literature review on mobile phone detector

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• My Account Login
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Review Article
• Open access
• Published: 18 October 2021

A systematic review of smartphone-based human activity recognition methods for health research

• Marcin Straczkiewicz   ORCID: orcid.org/0000-0002-8703-4451 1 ,
• Peter James 2 , 3 &
• Jukka-Pekka Onnela 1  

npj Digital Medicine volume  4 , Article number:  148 ( 2021 ) Cite this article

19k Accesses

80 Citations

23 Altmetric

Metrics details

• Predictive markers
• Public health
• Quality of life

Smartphones are now nearly ubiquitous; their numerous built-in sensors enable continuous measurement of activities of daily living, making them especially well-suited for health research. Researchers have proposed various human activity recognition (HAR) systems aimed at translating measurements from smartphones into various types of physical activity. In this review, we summarized the existing approaches to smartphone-based HAR. For this purpose, we systematically searched Scopus, PubMed, and Web of Science for peer-reviewed articles published up to December 2020 on the use of smartphones for HAR. We extracted information on smartphone body location, sensors, and physical activity types studied and the data transformation techniques and classification schemes used for activity recognition. Consequently, we identified 108 articles and described the various approaches used for data acquisition, data preprocessing, feature extraction, and activity classification, identifying the most common practices, and their alternatives. We conclude that smartphones are well-suited for HAR research in the health sciences. For population-level impact, future studies should focus on improving the quality of collected data, address missing data, incorporate more diverse participants and activities, relax requirements about phone placement, provide more complete documentation on study participants, and share the source code of the implemented methods and algorithms.

Similar content being viewed by others

Adults who microdose psychedelics report health related motivations and lower levels of anxiety and depression compared to non-microdosers

Principal component analysis

Effect of breathwork on stress and mental health: A meta-analysis of randomised-controlled trials

Introduction.

Progress in science has always been driven by data. More than 5 billion mobile devices were in use in 2020 1 , with multiple sensors (e.g., accelerometer and GPS) that can capture detailed, continuous, and objective measurements on various aspects of our lives, including physical activity. Such proliferation in worldwide smartphone adoption presents unprecedented opportunities for the collection of data to study human behavior and health. Along with sufficient storage, powerful processors, and wireless transmission, smartphones can collect a tremendous amount of data on large cohorts of individuals over extended time periods without additional hardware or instrumentation.

Smartphones are promising data collection instruments for objective and reproducible quantification of traditional and emerging risk factors for human populations. Behavioral risk factors, including but not limited to sedentary behavior, sleep, and physical activity, can all be monitored by smartphones in free-living environments, leveraging the personal or lived experiences of individuals. Importantly, unlike some wearable activity trackers 2 , smartphones are not a niche product but instead have become globally available, increasingly adopted by users of all ages both in advanced and emerging economies 3 , 4 . Their adoption in health research is further supported by encouraging findings made with other portable devices, primarily wearable accelerometers, which have demonstrated robust associations between physical activity and health outcomes, including obesity, diabetes, various cardiovascular diseases, mental health, and mortality 5 , 6 , 7 , 8 , 9 . However, there are some important limitations to using wearables for studying population health: (1) their ownership is much lower than that of smartphones 10 ; (2) most people stop using their wearables after 6 months of use 11 ; and (3) raw data are usually not available from wearable devices. The last point often forces investigators to rely on proprietary device metrics, which lowers the already low rate of reproducibility of biomedical research in general 12 and makes uncertainty quantification in the measurements nearly impossible.

Human activity recognition (HAR) is a process aimed at the classification of human actions in a given period of time based on discrete measurements (acceleration, rotation speed, geographical coordinates, etc.) made by personal digital devices. In recent years, this topic has been proliferating within the machine learning research community; at the time of writing, over 400 articles had been published on HAR methods using smartphones. This is a substantial increase from just a handful of articles published a few years earlier (Fig. 1 ). As data collection using smartphones becomes easier, analysis of the collected data is increasingly identified as the main bottleneck in health research 13 , 14 , 15 . To tackle the analytical challenges of HAR, researchers have proposed various algorithms that differ substantially in terms of the type of data they use, how they manipulate the collected data, and the statistical approaches used for inference and/or classification. Published studies use existing methods and propose new methods for the collection, processing, and classification of activities of daily living. Authors commonly discuss data filtering and feature selection techniques and compare the accuracy of various machine learning classifiers either on previously existing datasets or on datasets they have collected de novo for the purposes of the specific study. The results are typically summarized using classification accuracy within different groups of activities, such as ambulation, locomotion, and exercise.

Articles were published between January 2008 and December 2020, based on a search of PubMed, Scopus, and Web of Science databases (for details, see “Methods”).

To successfully incorporate developments in HAR into research in public health and medicine, there is a need to understand the approaches that have been developed and identify their potential limitations. Methods need to accommodate physiological (e.g., weight, height, age) and habitual (e.g., posture, gait, walking speed) differences of smartphone users, as well as differences in the built environment (e.g., buildings and green spaces) that provide the physical and social setting for human activities. Moreover, the data collection and statistical approaches typically used in HAR may be affected by location (where the user wears the phone on their body) and orientation of the device 16 , which complicates the transformation of collected data into meaningful and interpretable outputs.

In this paper, we systematically review the emerging literature on the use of smartphones for HAR for health research in free-living settings. Given that the main challenge in this field is shifting from data collection to data analysis, we focus our analysis on the approaches used for data acquisition, data preprocessing, feature extraction, and activity classification. We provide insight into the complexity and multidimensionality of HAR utilizing smartphones, the types of data collected, and the methods used to translate digital measurements into human activities. We discuss the generalizability and reproducibility of approaches, i.e., the features that are essential and applicable to large and diverse cohorts of study participants. Lastly, we identify challenges that need to be tackled to accelerate the wider utilization of smartphone-based HAR in public health studies.

Our systematic review was conducted by searching for articles published up to December 31, 2020, on PubMed, Scopus, and Web of Science databases. The databases were screened for titles, abstracts, and keywords containing phrases “activity” AND (“recognition” OR “estimation” OR “classification”) AND (“smartphone” OR “cell phone” OR “mobile phone”). The search was limited to full-length journal articles written in English. After removing duplicates, we read the titles and abstracts of the remaining publications. Studies that did not investigate HAR approaches were excluded from further screening. We then filtered out studies that employed auxiliary equipment, like wearable or ambient devices, and studies that required carrying multiple smartphones. Only studies that made use of commercially available consumer-grade smartphones (either personal or loaner) were read in full. We excluded studies that used the smartphone microphone or video camera for activity classification as they might record information about an individual’s surroundings, including information about unconsented individuals, and thus hinder the large-scale application of the approach due to privacy concerns. To focus on studies that mimicked free-living settings, we excluded studies that utilized devices strapped or glued to the body in a fixed position.

Our search resulted in 1901 hits for the specified search criteria (Fig. 2 ). After removal of articles that did not discuss HAR algorithms ( n  = 793), employed additional hardware ( n  = 150), or utilized microphones, cameras, or body-affixed smartphones ( n  = 149), there were 108 references included in this review.

The search was conducted in PubMed, Scopus, and Web of Science databases and included full-length peer-reviewed articles written in English. The search was carried out on January 2, 2021.

Most HAR approaches consist of four stages: data acquisition, data preprocessing, feature extraction, and activity classification (Fig. 3 ). Here, we provide an overview of these steps and briefly point to significant methodological differences among the reviewed studies for each step. Figure 4 summarizes specific aspects of each study. Of note, we decomposed data acquisition processes into sensor type, experimental environment, investigated activities, and smartphone location; we indicated which studies preprocessed collected measurements using signal correction methods, noise filtering techniques, and sensor orientation-invariant transformations; we marked investigations based on the types of signal features they extracted, as well as the feature selection approaches used; we indicated the adopted activity classification principles, utilized classifiers, and practices for accuracy reporting; and finally, we highlighted efforts supporting reproducibility and generalizability of the research. Before diving into these technical considerations, we first provide a brief description of study populations.

The map displays common aspects of HAR systems together with their operational definitions. The methodological differences between the reviewed studies are highlighted in Figure 4 .

The columns correspond to the 108 reviewed studies and the rows correspond to different technical aspects of each study. Cells marked with a cross (x) indicate that the given study used the given method, algorithm, or approach. Rows have been grouped to correspond to different stages of HAR, such as data processing, and color shading of rows indicates how frequently a particular aspect is present among the studies (darker shade corresponds to higher frequency).

Study populations

We use the term study population to refer to the group of individuals investigated in any given study. In the reviewed studies, data were usually collected from fewer than 30 individuals, although one larger study analyzed data from 440 healthy individuals 17 . Studies often included healthy adults in their 20s and 30s, with only a handful of studies involving older individuals. Most studies did not report the full distribution of ages, only the mean age or the age range of participants (Fig. 5 ). To get a sense of the distribution of participant ages, we attempted to reconstruct an overall approximate age distribution by assuming that the participants in each study are evenly distributed in age between the minimum and maximum ages, which may not be the case. A comparison of the reconstructed age distribution of study participants with nationwide age distributions clearly demonstrates that future HAR research in health settings needs to broaden the age spectrum of the participants. Less effort was devoted in the studies to investigating populations with different demographic and disease characteristics, such as elders 18 , 19 , 20 and individuals with Parkinson’s disease 21 .

Panel a displays age of the population corresponding to individual studies, typically described by its range (lines) or mean (dots). Panel b displays the reconstructed age distribution in the reviewed studies (see the text). Nationwide age distributions displayed in panel c of three countries offer a stark contrast with the reconstructed distribution of study participant ages.

Data acquisition

We use the term data acquisition to refer to a process of collecting and storing raw sub-second-level smartphone measurements for the purpose of HAR. The data are typically collected from individuals by an application that runs on the device and samples data from built-in smartphone sensors according to a predefined schedule. We carefully examined the selected literature for details on the investigated population, measurement environment, performed activities, and smartphone settings.

In the reviewed studies, data acquisition typically took place in a research facility and/or nearby outdoor surroundings. In such environments, study participants were asked to perform a series of activities along predefined routes and to interact with predefined objects. The duration and order of performed activities were usually determined by the study protocol and the participant was supervised by a research team member. A less common approach involved observation conducted in free-living environments, where individuals performed activities without specific instructions. Such studies were likely to provide more insight into diverse activity patterns due to individual habits and unpredictable real-life conditions. Compared to a single laboratory visit, studies conducted in free-living environments also allowed investigators to monitor behavioral patterns over many weeks 22 or months 23 .

Activity selection is one of the key aspects of HAR. The studies in our review tended to focus on a small set of activities, including sitting, standing, walking, running, and stair climbing. Less common activities involved various types of mobility, locomotion, fitness, and household routines, e.g., slow, normal, and brisk walking 24 , multiple transportation modes, such as by car, bus, tram, train, metro, and ferry 25 , sharp body-turns 26 , and household activities, like sweeping a floor or walking with a shopping bag 27 . More recent studies concentrated solely on walking recognition 28 , 29 . As shown in Fig. 4 , the various measured activities in the reviewed studies can be grouped into classes: “posture” refers to lying, sitting, standing, or any pair of these activities; “mobility” refers to walking, stair climbing, body-turns, riding an elevator or escalator, running, cycling, or any pair of these activities; “locomotion” refers to motorized activities; and “other” refers to various household and fitness activities or singular actions beyond the described groups.

The spectrum of investigated activities determines the choice of sensors used for data acquisition. At the time of writing, a standard smartphone is equipped with a number of built-in hardware sensors and protocols that can be used for activity monitoring, including an accelerometer, gyroscope, magnetometer, GPS, proximity sensor, and light sensor, as well as to collect information on ambient pressure, humidity, and temperature (Fig. 6 ). Accurate estimation of commonly available sensors over time is challenging given a large number of smartphone manufacturers and models, as well as the variation in their adoption in different countries. Based on global statistics on smartphone market shares 30 and specifications of flagship models 31 , it appears that accelerometer, gyroscope, magnetometer, GPS, and proximity and light sensors were fairly commonly available by 2010. Other smartphone sensors were introduced a couple of years later; for example, the barometer was included in Samsung Galaxy S III released in 2012, and thermometer and hygrometer were included in Samsung Galaxy S4 released in 2013.

Inertial sensors (accelerometer, gyroscope, and magnetometer) provide measurements with respect to the three orthogonal axes ( x , y , z ) of the body of the phone; the remaining sensors are orientation-invariant.

Our literature review revealed that the most commonly used sensors for HAR are the accelerometer, gyroscope, and magnetometer, which capture data about acceleration, angular velocity, and phone orientation, respectively, and provide temporally dense, high-resolution measurements for distinguishing among activity classes (Fig. 7 ). Inertial sensors were often used synchronously to provide more insight into the dynamic state of the device. Some studies showed that the use of a single sensor can yield similar accuracy of activity recognition as using multiple sensors in combination 32 . To alleviate the impact of sensor position, some researchers collected data using the built-in barometer and GPS sensors to monitor changes in altitude and geographic location 33 , 34 , 35 . Certain studies benefited from using the broader set of capabilities of smartphones; for example, some researchers additionally exploited the proximity sensor and light sensor to allow recognition of a measurement’s context, e.g., the distance between a smartphone and the individual’s body, and changes between in-pocket and out-of-pocket locations based on changes in illumination 36 , 37 . The selection of sensors was also affected by secondary research goals, such as simplicity of classification and minimization of battery drain. In these studies, data acquisition was carried out using a single sensor (e.g., accelerometer 22 ), a small group of sensors (e.g., accelerometer and GPS 38 ), or a purposely modified sampling frequency or sampling scheme (e.g., alternating between data collection and non-collection cycles) to reduce the volume of data collected and processed 39 . Supplementing GPS data with other sensor data was motivated by the limited indoor reception of GPS; satellite signals may be absorbed or attenuated by walls and ceilings 17 up to 60% of the time inside buildings and up to 70% of the time in underground trains 23 .

a A person is sitting by the desk with the smartphone placed in the front pants pocket; b a person is walking normally (~1.9 steps per second) with the smartphone placed in a jacket pocket; c a person is ascending stairs with the smartphone placed in the backpack; d a person is walking slowly (~1.4 steps per second) holding the smartphone in hand; e a person is jogging (~2.8 steps per second) with the smartphone placed in back short’s pocket.

Sampling frequency specifies how many observations are collected by a sensor within a 1-s time interval. The selection of sampling frequency is usually performed as a trade-off between measurement accuracy and battery drain. Sampling frequency in the reviewed studies typically ranged between 20 and 30 Hz for inertial sensors and 1 and 10 Hz for the barometer and GPS. The most significant variations were seen in studies where limited energy consumption was a priority (e.g., accelerometer sampled at 1 Hz 40 ) or if investigators used advanced signal processing methods, such as time-frequency decomposition methods, or activity templates that required higher sampling frequency (e.g., accelerometer sampled at 100 Hz 41 ). Some studies stated that inertial sensors sampled at 20 Hz provided enough information to distinguish between various types of transportation 42 , while 10 Hz sampling rate was sufficient to distinguish between various types of mobility 43 . One study demonstrated that reducing the sampling rate from 100 Hz to 12.5 Hz increased the duration of data collection by a factor of three on a single battery charge 44 .

A crucial parameter in the data acquisition process is the smartphone’s location on the body. This is important mainly because of the nonstationary nature of real-life conditions and the strong effect it has on the smartphone’s inertial sensors. The main challenge in HAR in free-living conditions is that data recorded by the accelerometer, gyroscope, and magnetometer sensors differ between the upper and lower body as the device is not affixed to any specific location or orientation 45 . Therefore, it is essential that studies collect data from as many body locations as possible to ensure the generalizability of results. In the reviewed literature, study participants were often instructed to carry the device in a pants pocket (either front or back), although a number of studies also considered other placements, such as jacket pocket 46 , bag or backpack 47 , 48 , and holding the smartphone in the hand 49 or in a cupholder 50 .

To establish the ground truth for physical activity in HAR studies, data were usually annotated manually by trained research personnel or by the study participants themselves 51 , 52 . However, we also noted several approaches that automated this process both in controlled and free-living conditions, e.g., through a designated smartphone application 22 or built-in step counter combined paired with GPS data 53 ., used a built-in step counter and GPS data to produce “weak” labels. The annotation was also done using the built-in microphone 54 , video camera 18 , 20 , or an additional body-worn sensor 29 .

Finally, the data acquisition process in the reviewed studies was carried out on purposely designed applications that captured data. In studies with online activity classification, the collected data did not leave the device, but instead, the entire HAR pipeline was implemented on the smartphone; in contrast, studies using offline classification transmitted data to an external (remote) server for processing using a cellular, Wi-Fi, Bluetooth, or wired connection.

Data preprocessing

We use the term data preprocessing to refer to a collection of procedures aimed at repairing, cleaning, and transforming measurements recorded for HAR. The need for such step is threefold: (1) measurement systems embedded in smartphones are often less stable than research-grade data acquisition units, and the data might therefore be sampled unevenly or there might be missingness or sudden spikes that are unrelated to an individual’s actual behavior; (2) the spatial orientation (how the phone is situated in a person’s pocket, say) of the device influences tri-axial measurements of inertial sensors, thus potentially degrading the performance of the HAR system; and (3) despite careful planning and execution of the data acquisition stage, data quality may be compromised due to other unpredictable factors, e.g., lack of compliance by the study participants, unequal duration of activities in the measurement (i.e., dataset imbalance), or technological issues.

In our literature review, the first group of obstacles was typically addressed using signal processing techniques (in Fig. 4 , see “standardization”). For instance, to alleviate the mismatch between requested and effective sampling frequency, researchers proposed the use of linear interpolation 55 or spline interpolation 56 (Fig. 8 ). Such procedures were imposed on a range of affected sensors, typically the accelerometer, gyroscope, magnetometer, and barometer. Further time-domain preprocessing considered data trimming, carried out to remove unwanted data components. For this purpose, the beginning and end of each activity bout, a short period of activity of a specified kind, were clipped as nonrepresentative for the given activity 46 . During this stage, the researchers also dealt with dataset imbalance, which occurs when there are different numbers of observations for different activity classes in the training dataset. Such a situation makes the classifier susceptible to overfitting in favor of the larger class; in the reviewed studies, this issue was resolved using up-sampling or down-sampling of data 17 , 57 , 58 , 59 . In addition, the measurements were processed for high-frequency noise cancellation (i.e., “denoising”). The literature review identified several methods suitable for this task, including the use of low-pass finite impulse response filters (with a cutoff frequency typically equal to 10 Hz for inertial sensors and 0.1 Hz for barometers) 60 , 61 , which remove the portion of the signal that is unlikely to result from the activities of interest; weighted moving average 55 ; moving median 45 , 62 ; and singular-value decomposition 63 . GPS data were sometimes de-noised based on the maximum allowed positional accuracy 64 .

Standardization includes relabeling ( a ), when labels are reassigned to better match transitions between activities; trimming ( b ), when part of the signal is removed to balance the dataset for system training; interpolation ( c ), when missing data are filled in based on adjacent observations; and denoising ( d ), when the signal is filtered from redundant components. The transformation includes normalization ( e ), when the signal is normalized to unidimensional vector magnitude; rotation ( f ), when the signal is rotated to a different coordinate system; and separation ( g ), when the signal is separated into linear and gravitational components. Raw accelerometer data are shown in gray, and preprocessed data are shown using different colors.

Another element of data preprocessing considers device orientation (in Fig. 4 , see “transformation”). Smartphone measurements are sensitive to device orientation, which may be due to clothing, body shape, and movement during dynamic activities 57 . One of the popular solutions reported in the literature was to transform the three-dimensional signal into a univariate vector magnitude that is invariant to rotations and more robust to translations. This procedure was often applied to accelerometer, gyroscope, and magnetometer data. Accelerometer data were also subjected to digital filtering by separating the signal into linear (related to body motions) and gravitational (related to device spatial orientation) acceleration 65 . This separation was typically performed using a high-pass Butterworth filter of low order (e.g., order 3) with a cutoff frequency below 1 Hz. Other approaches transformed tri-axial into bi-axial measurement with horizontal and vertical axes 49 , or projected the data from the device coordinate system into a fixed coordinate system (e.g., the coordinate system of a smartphone that lies flat on the ground) using a rotation matrix (Euler angle-based 66 or quaternion 47 , 67 ).

Feature extraction

We use the term feature extraction to refer to a process of selecting and computing meaningful summaries of smartphone data for the goal of activity classification. A typical extraction scheme includes data visualization, data segmentation, feature selection, and feature calculation. A careful feature extraction step allows investigators not only to understand the physical nature of activities and their manifestation in digital measurements, but also, and more importantly, to help uncover hidden structures and patterns in the data. The identified differences are later quantified through various statistical measures to distinguish between activities. In an alternative approach, the process of feature extraction is automated using deep learning, which handles feature selection using simple signal processing units, called neurons, that have been arranged in a network structure that is multiple layers deep 59 , 68 , 69 , 70 . As with many applications of deep learning, the results may not be easily interpretable.

The conventional approach to feature extraction begins with data exploration. For this purpose, researchers in our reviewed studies employed various graphical data exploration techniques like scatter plots, lag plots, autocorrelation plots, histograms, and power spectra 71 . The choice of tools was often dictated by the study objectives and methods. For example, research on inertial sensors typically presented raw three-dimensional data from accelerometers, gyroscopes, and magnetometers plotted for the corresponding activities of standing, walking, and stair climbing 50 , 72 , 73 . Acceleration data were often inspected in the frequency domain, particularly to observe periodic motions of walking, running, and cycling 45 , and the impact of the external environment, like natural vibration frequencies of a bus or a subway 74 . Locomotion and mobility were investigated using estimates of speed derived from GPS. In such settings, investigators calculated the average speed of the device and associated it with either the group of motorized (car, bus, train, etc.) or non-motorized (walking, cycling, etc.) modes of transportation.

In the next step, measurements are divided into smaller fragments (also, segments or epochs) and signal features are calculated for each fragment (Fig. 9 ). In the reviewed studies, this segmentation was typically conducted using a windowing technique that allows consecutive windows to overlap. The window size usually had a fixed length that varied from 1 to 5 s, while the overlap of consecutive windows was often set to 50%. Several studies that investigated the optimal window size supported this common finding: short windows (1–2 s) were sufficient for recognizing posture and mobility, whereas somewhat longer windows (4–5 s) had better classification performance 75 , 76 , 77 . Even longer windows (10 s or more) were recommended for recognizing locomotion modes or for HAR systems employing frequency-domain features calculated with the Fourier transform (resolution of the resulting frequency spectrum is inversely proportional to window length) 42 . In principle, this calibration aims to closely match the window size with the duration of a single instance of the activity (e.g., one step). Similar motivation led researchers to seek more adaptive segmentation methods. One idea was to segment data based on specific time-domain events, like zero-cross points (when the signal changes value from positive to negative or vice versa), peak points (local maxima), or valley points (local minima), which represent the start and endpoints of a particular activity bout 55 , 57 . This allowed for segments to have different lengths corresponding to a single fundamental period of the activity in question. Such an approach was typically used to recognize quasiperiodic activities like walking, running, and stair climbing 63 .

An analyzed measurement ( a ) is segmented into smaller fragments using a sliding window ( b ). Depending on the approach, each segment may then be used to compute time-domain ( c ) or frequency-domain features ( d ), but also it may serve as the activity template ( e ), or as input for deep learning networks that compute hidden (“deep”) features ( f ). The selected feature extraction approach determines the activity classifier: time- and frequency-domain features are paired with machine learning classifiers ( g ) and activity templates are investigated using distance metrics ( h ), while deep features are computed within embedded layers of convolutional neural networks ( i ).

The literature described a large variety of signal features used for HAR, which can be divided into several categories based on the initial signal processing procedure. This enables one to distinguish between activity templates (i.e., raw signal), deep features (i.e., hidden features calculated within layers of deep neural networks), time-domain features (i.e., statistical measures of time-series data), and frequency-domain features (i.e., statistical measures of frequency representation of time-series data). The most popular features in the reviewed papers were calculated from time-domain signals as descriptive statistics, such as local mean, variance, minimum and maximum, interquartile range, signal energy (defined as the area under the squared magnitude of the considered continuous signal), and higher-order statistics. Other time-domain features included mean absolute deviation, mean (or zero) crossing rate, regression coefficients, and autocorrelation. Some studies described novel and customized time-domain features, like histograms of gradients 78 , and the number of local maxima and minima, their amplitude, and the temporal distance between them 39 . Time-domain features were typically calculated over each axis of the three-dimensional measurement or orientation-invariant vector magnitude. Studies that used GPS also calculated average speed 64 , 79 , 80 , while studies that used the barometer analyzed the pressure derivative 81 .

Signals transformed to the frequency domain were less exploited in the literature. A commonly performed signal decomposition used the fast Fourier transform (FFT) 82 , 83 , an algorithm that converts a temporal sequence of samples to a sequence of frequencies present in that sample. The essential advantage of frequency-domain features over time-domain features is their ability to identify and isolate certain periodic components of performed activities. This enabled researchers to estimate (kinetic) energy within particular frequency bands associated with human activities, like gait and running 51 , as well as with different modes of locomotion 74 . Other frequency-domain features included spectral entropy and parameters of the dominant peak, e.g., its frequency and amplitude.

Activity templates function essentially as blueprints for different types of physical activity. In the HAR systems, we reviewed, these templates were compared to patterns of observed raw measurements using various distance metrics 38 , 84 , such as the Euclidean or Manhattan distance. Given the heterogeneous nature of human activities, activity templates were often enhanced using techniques similar to dynamic time warping 29 , 57 , which measures the similarity of two temporal sequences that may vary in speed. As an alternative to raw measurements, some studies used signal symbolic approximation, which translates a segmented time-series signal into sequences of symbols based on a predefined mapping rule (e.g., amplitude between −1 and −0.5 g represents symbol “a”, amplitude between −0.5 and 0 g represents symbol “b”, and so on) 85 , 86 , 87 .

More recent studies utilized deep features. In these approaches, smartphone data were either fed to deep neural networks as raw univariate or multivariate time series 35 , 48 , 60 or preprocessed into handcrafted time- and frequency-domain feature vectors 82 , 83 . Within the network layers, the input data were then transformed (e.g., using convolution) to produce two-dimensional activation maps that revealed hidden spatial relations between axes and sensors specific to a given activity. To improve the resolution of input data, one study proposed to split the integer and decimal values of accelerometer measurements 41 .

In the reviewed articles, the number of extracted features typically varied from a few to a dozen. However, some studies purposely calculated too many features (sometimes hundreds) and let the analytical method perform variable selection, i.e., identify those features that were most informative for HAR 88 . Support vector machines 81 , 89 , gain ratio 43 , recursive feature elimination 38 , correlation-based feature selection 51 , and principal component analysis 90 were among the popular feature selection/dimension reduction methods used.

Activity classification

We use the term activity classification to refer to a process of associating extracted features with particular activity classes based on the adopted classification principle. The classification is typically performed by a supervised learning algorithm that has been trained to recognize patterns between features and labeled physical activities in the training dataset. The fitted model is then validated on separate observations, using a validation dataset, usually data obtained from the same group of study participants. The comparison between predictions made by the model and the known true labels allows one to assess the accuracy of the approach. This section summarizes the methods used in classification and validation, and also provides some insights into reporting on HAR performance.

The choice of classifier aims to identify a method that has the highest classification accuracy for the collected datasets and for the given data processing environment (e.g., online vs. offline). The reviewed literature included a broad range of classifiers, from simple decision trees 18 , k-nearest neighbors 65 , support vector machines 91 , 92 , 93 , logistic regression 21 , naïve Bayes 94 , and fuzzy logic 64 to ensemble classifiers such as random forest 76 , XGBoost 95 , AdaBoost 45 , 96 , bagging 24 , and deep neural networks 48 , 60 , 82 , 97 , 98 , 99 . Simple classifiers were frequently compared to find the best solution in the given measurement scenario 43 , 53 , 100 , 101 , 102 . A similar type of analysis was implemented for ensemble classifiers 79 . Incremental learning techniques were proposed to adapt the classification model to new data streams and unseen activities 103 , 104 , 105 . Other semi-supervised approaches were proposed to utilize unlabeled data to improve the personalization of HAR systems 106 and data annotation 53 , 70 . To increase the effectiveness of HAR, some studies used a hierarchical approach, where the classification was performed in separate stages and each stage could use a different classifier. The multi-stage technique was used for gradual decomposition of activities (coarse-grained first, then fine-grained) 22 , 37 , 52 , 60 and to handle the predicament of changing sensor location (body location first, then activity) 91 . Multi-instance multi-label approaches were adapted for the classification of complex activities (i.e., activities that consist of several basic activities) 62 , 107 as well as for recognition of basic activities paired with different sensor locations 108 .

Classification accuracy could also be improved by using post-processing, which relies on modifying the initially assigned label using the rules of logic and probability. The correction was typically performed based on activity duration 74 , activity sequence 25 , and activity transition probability and classification confidence 80 , 109 .

The selected method is typically cross-validated, which splits the collected dataset into two or more parts—training and testing—and only uses the part of the data for testing that was not used for training. The literature mentions a few cross-validation procedures, with k -fold and leave-one-out cross-validation being the most common 110 . Popular train-test proportions were 90–10, 70–30, and 60–40. A validation is especially valuable if it is performed using studies with different demographics and smartphone use habits. Such an approach allows one to understand the generalizability of the HAR system to real-life conditions and populations. We found a few studies that followed this validation approach 18 , 21 , 71 .

Activity classification is the last stage of HAR. In our review, we found that analysis results were typically reported in terms of classification accuracy using various standard metrics like precision, recall, and F-score. Overall, the investigated studies reported very high classification accuracies, typically above 95%. Several comparisons revealed that ensemble classifiers tended to outperform individual or single classifiers 27 , 77 , and deep-learning classifiers tended to outperform both individual and ensemble classifiers 48 . More nuanced summaries used the confusion matrix, which allows one to examine which activities are more likely to be classified incorrectly. This approach was particularly useful for visualizing classification differences between similar activities, such as normal and fast walking or bus and train riding. Additional statistics were usually provided in the context of HAR systems designed to operate on the device. In this case, activity classification needed to be balanced among acceptable classifier performance, processing time, and battery drain 44 . The desired performance optimum was obtained by making use of dataset remodeling (e.g., by replacing the oldest observations with the newest ones), low-cost classification algorithms, limited preprocessing, and conscientious feature selection 45 , 86 . Computation time was sometimes reported for complex methods, such as deep neural networks 20 , 82 , 111 and extreme learning machine 112 , as well as for symbolic representation 85 , 86 and in comparative analyses 46 . A comprehensive comparison of results was difficult or impossible, as discussed below.

Over the past decade, many studies have investigated HAR using smartphones. The reviewed literature provides detailed descriptions of essential aspects of data acquisition, data preprocessing, feature extraction, and activity classification. Studies were conducted with one or more objectives, e.g., to limit technological imperfections (e.g., no GPS signal reception indoors), to minimize computational requirements (e.g., for online processing of data directly on the device), and to maximize classification accuracy (all studies). Our review summarizes the most frequently used methods and offers available alternatives.

As expected, no single activity recognition procedure was found to work in all settings, which underlines the importance of designing methods and algorithms that address specific research questions in health while keeping the specifics of the study cohort in mind (e.g., age distribution, the extent of device use, and nature of disability). While datasets were usually collected in laboratory settings, there was little evidence that algorithms trained using data collected in these controlled settings could be generalized to free-living conditions 113 , 114 . In free-living settings, duration, frequency, and specific ways of performing any activity are subject to context and individual ability, and these degrees of freedom need to be considered in the development of HAR systems. Validation of these data in free-living settings is essential, as the true value of HAR systems for public health will come through transportable and scalable applications in large, long-term observational studies or real-world interventions.

Some studies were conducted with a small number of able-bodied volunteers. This makes the process of data handling and classification easier but also limits the generalizability of the approach to more diverse populations. The latter point was well demonstrated in two of the investigated studies. In the first study, the authors observed that the performance of a classifier trained on a young cohort significantly decreases if validated on an older cohort 18 . Similar conclusions can be drawn from the second study, where the observations on healthy individuals did not replicate in individuals with Parkinson’s disease 21 . These facts highlight the role of algorithmic fairness (or fairness of machine learning), the notion that the performance of an algorithm should not depend on variables considered sensitive, such as race, ethnicity, sexual orientation, age, and disability. A highly visible example of this was the decision of some large companies, including IBM, to stop providing facial recognition technology to police departments for mass surveillance 115 , and the European Commission has considered a ban on the use of facial recognition in public spaces 116 . These decisions followed findings demonstrating the poor performance of facial recognition algorithms when applied to individuals with dark-skin tones.

The majority of the studies we reviewed utilized stationary smartphones at a single-body position (i.e., a specific pants pocket), sometimes even with a fixed orientation. However, such scenarios are rarely observed in real-life settings, and these types of studies should be considered more as proofs of concept. Indeed, as demonstrated in several studies, inertial sensor data might not share similar features across body locations 49 , 117 , and smartphone orientation introduces additional artifacts to each axis of measurement which make any distribution-based features (e.g., mean, range, skewness) difficult to use without appropriate data preprocessing. Many studies provided only incomplete descriptions of the experimental setup and study protocol and provided few details on demographics, environmental context, and the details of the performed activities. Such information should be reported as fully and accurately as possible.

Only a few studies considered classification in a context that involves activities outside the set of activities the system was trained on; for example, if the system was trained to recognize walking and running, these were the only two activities that the system was later tested on. However, real-life activities are not limited to a prescribed set of behaviors, i.e., we do not just sit still, stand still, walk, and climb stairs. These classifiers, when applied to free-living conditions, will naturally miss the activities they were not trained on but will also likely overestimate those activities they were trained on. An improved scheme could assume that the observed activities are a sample from a broader spectrum of possible behaviors, including periods when the smartphone is not on a person, or assess the uncertainty associated with the classification of each type of activity 84 . This could also provide for an adaptive approach that would enable observation/interventions suited to a broad range of activities relevant for health, including decreasing sedentary behavior, increasing active transport (i.e., walking, bicycling, or public transit), and improving circadian patterns/sleep.

The use of personal digital devices, in particular smartphones, makes it possible to follow large numbers of individuals over long periods of time, but invariably investigators need to consider approaches to missing sensor data, which is a common problem. The importance of this problem is illustrated in a recent paper that introduced a resampling approach to imputing missing smartphone GPS data; the authors found that relative to linear interpolation—the naïve approach to missing spatial data—imputation resulted in a tenfold reduction in the error averaged across all daily mobility features 118 . On the flip side of missing data is the need to propagate uncertainty, in a statistically principled way, from the gaps in the raw data to the inferences that investigators wish to draw from the data. It is a common observation that different people use their phones differently, and some may barely use their phones at all; the net result is not that the data collected from these individuals are not useful, but instead the data are less informative about the behavior of this individual than they ideally might be. Dealing with missing data and accounting for the resulting uncertainty is important because it means that one does not have to exclude participants from a study because their data fail meet some arbitrary threshold of completeness; instead, everyone counts, and every bit of data from each individual counts.

The collection of behavioral data using smartphones understandably raises concerns about privacy; however, investigators in health research are well-positioned to understand and address these concerns given that health data are generally considered personal and private in nature. Consequently, there are established practices and common regulations on human subjects’ research, where informed consent of the individual to participate is one of the key foundations of any ethically conducted study. Federated learning is a machine learning technique that can be used to train an algorithm across decentralized devices, here smartphones, using only local data (data from the individual) and without the need to exchange data with other devices. This approach appears at first to provide a powerful solution to the privacy problem: the personal data never leave the person’s phone and only the outputs of the learning process, generally parameter estimates, are shared with others. This is where the tension between privacy and the need for reproducible research arises, however. The reason for data collection is to produce generalizable knowledge, but according to an often-cited study, 65% of medical studies were inconsistent when retested and only 6% were completely reproducible 12 . In the studies reviewed here, only 4 out of 108 made the source code or the methods used in the study publicly available. For a given scientific question, studies that are not replicable require the collection of more private and personal data; this highlights the importance of reproducibility of studies, especially in health, where there are both financial and ethical considerations when conducting research. If federated learning provides no possibility to confirm data analyses, to re-analyze data using different methods, or to pool data across studies, it by itself cannot be the solution to the privacy problem. Nevertheless, the technique may act as inspiration for developing privacy-preserving methods that also enable future replication of studies. One possibility is to use publicly available datasets (Table 1 ). If sharing of source code were more common, HAR methods could be tested on these publicly available datasets, perhaps in a similar way as datasets of handwritten digits are used to test classification methods in machine learning research. Although some efforts have been made in this area 42 , 119 , 120 , 121 , the recommended course of action assumes collecting and analyzing data from a large spectrum of sensors on diverse and understudied populations and validating classifiers against widely accepted gold standards.

When accurate, reproducible, and transportable methods coalesce to recognize a range of relevant activity patterns, smartphone-based HAR approaches will provide a fundamental tool for public health researchers and practitioners alike. We hope that this paper has provided to the reader some insights into how smartphones may be used to quantify human behavior in health research and the complexities that are involved in the collection and analysis of such data in this challenging but important field.

Data availability

Aggregated data analyzed in this study are available from the corresponding author upon request.

Code availability

Scripts used to process the aggregated data are available from the corresponding author upon request.

Association, G. The mobile economy 2020. https://www.gsma.com/mobileeconomy/wp-content/uploads/2020/03/GSMA_MobileEconomy2020_Global.pdf (2020).

Mercer, K. et al. Acceptance of commercially available wearable activity trackers among adults aged over 50 and with chronic illness: a mixed-methods evaluation. JMIR mHealth uHealth 4 , e7 (2016).

Article   PubMed   PubMed Central   Google Scholar  

Anderson, M. & Perrin, A. Tech adoption climbs among older adults. http://www.pewinternet.org/wp-content/uploads/sites/9/2017/05/PI_2017.05.17_Older-Americans-Tech_FINAL.pdf (2017).

Taylor, K. & Silver, L. Smartphone ownership is growing rapidly around the world, but not always equally. http://www.pewresearch.org/global/wp-content/uploads/sites/2/2019/02/Pew-Research-Center_Global-Technology-Use-2018_2019-02-05.pdf (2019).

Cooper, A. R., Page, A., Fox, K. R. & Misson, J. Physical activity patterns in normal, overweight and obese individuals using minute-by-minute accelerometry. Eur. J. Clin. Nutr. 54 , 887–894 (2000).

Article   CAS   PubMed   Google Scholar  

Ekelund, U., Brage, S., Griffin, S. J. & Wareham, N. J. Objectively measured moderate- and vigorous-intensity physical activity but not sedentary time predicts insulin resistance in high-risk individuals. Diabetes Care 32 , 1081–1086 (2009).

Legge, A., Blanchard, C. & Hanly, J. G. Physical activity, sedentary behaviour and their associations with cardiovascular risk in systemic lupus erythematosus. Rheumatology https://doi.org/10.1093/rheumatology/kez429 (2019).

Loprinzi, P. D., Franz, C. & Hager, K. K. Accelerometer-assessed physical activity and depression among U.S. adults with diabetes. Ment. Health Phys. Act. 6 , 79–82 (2013).

Article   Google Scholar  

Smirnova, E. et al. The predictive performance of objective measures of physical activity derived from accelerometry data for 5-year all-cause mortality in older adults: National Health and Nutritional Examination Survey 2003–2006. J. Gerontol. Ser. A https://doi.org/10.1093/gerona/glz193 (2019).

Wigginton, C. Global Mobile Consumer Trends, 2nd edition. Deloitte, https://www2.deloitte.com/content/dam/Deloitte/us/Documents/technology-media-telecommunications/us-global-mobile-consumer-survey-second-edition.pdf (2017).

Coorevits, L. & Coenen, T. The rise and fall of wearable fitness trackers. Acad. Manag . 2016 , https://doi.org/10.5465/ambpp.2016.17305abstract (2016).

Prinz, F., Schlange, T. & Asadullah, K. Believe it or not: how much can we rely on published data on potential drug targets? Nat. Rev. Drug Discov. 10 , 712 (2011).

Kubota, K. J., Chen, J. A. & Little, M. A. Machine learning for large-scale wearable sensor data in Parkinson’s disease: concepts, promises, pitfalls, and futures. Mov. Disord. 31 , 1314–1326 (2016).

Article   PubMed   Google Scholar  

Iniesta, R., Stahl, D. & McGuffin, P. Machine learning, statistical learning and the future of biological research in psychiatry. Psychol. Med. 46 , 2455–2465 (2016).

Article   CAS   PubMed   PubMed Central   Google Scholar  

Kuehn, B. M. FDA’s foray into big data still maturing. J. Am. Med. Assoc. 315 , 1934–1936 (2016).

Straczkiewicz, M., Glynn, N. W. & Harezlak, J. On placement, location and orientation of wrist-worn tri-axial accelerometers during free-living measurements. Sensors 19 , 2095 (2019).

Esmaeili Kelishomi, A., Garmabaki, A. H. S., Bahaghighat, M. & Dong, J. Mobile user indoor-outdoor detection through physical daily activities. Sensors 19 , 511 (2019).

Article   PubMed Central   Google Scholar  

Del Rosario, M. B. et al. A comparison of activity classification in younger and older cohorts using a smartphone. Physiol. Meas. 35 , 2269–2286 (2014).

Del Rosario, M. B., Lovell, N. H. & Redmond, S. J. Learning the orientation of a loosely-fixed wearable IMU relative to the body improves the recognition rate of human postures and activities. Sensors 19 , 2845 (2019).

Nan, Y. et al. Deep learning for activity recognition in older people using a pocket-worn smartphone. Sensors 20 , 7195 (2020).

Albert, M. V., Toledo, S., Shapiro, M. & Kording, K. Using mobile phones for activity recognition in Parkinson’s patients. Front. Neurol. 3 , 158 (2012).

Liang, Y., Zhou, X., Yu, Z. & Guo, B. Energy-efficient motion related activity recognition on mobile devices for pervasive healthcare. Mob. Netw. Appl. 19 , 303–317 (2014).

Gjoreski, H. et al. The university of Sussex-Huawei locomotion and transportation dataset for multimodal analytics with mobile devices. IEEE Access 6 , 42592–42604 (2018).

Wu, W., Dasgupta, S., Ramirez, E. E., Peterson, C. & Norman, G. J. Classification accuracies of physical activities using smartphone motion sensors. J. Med. Internet Res. 14 , e130 (2012).

Guvensan, M. A., Dusun, B., Can, B. & Turkmen, H. I. A novel segment-based approach for improving classification performance of transport mode detection. Sensors 18 , 87 (2018).

Pei, L. et al. Human behavior cognition using smartphone sensors. Sensors 13 , 1402–1424 (2013).

Della Mea, V., Quattrin, O. & Parpinel, M. A feasibility study on smartphone accelerometer-based recognition of household activities and influence of smartphone position. Inform. Heal. Soc. Care 42 , 321–334 (2017).

Klein, I. Smartphone location recognition: a deep learning-based approach. Sensors 20 , 214 (2020).

Casado, F. E. et al. Walking recognition in mobile devices. Sensors 20 , 1189 (2020).

O’Dea, S. Global smartphone market share worldwide by vendor 2009–2020. https://www.statista.com/statistics/271496/global-market-share-held-by-smartphone-vendors-since-4th-quarter-2009/ (2021).

GSMArena. https://www.gsmarena.com/ (2021). Accessed 24 March 2021.

Shoaib, M., Bosch, S., Durmaz Incel, O., Scholten, H. & Havinga, P. J. M. Fusion of smartphone motion sensors for physical activity recognition. Sensors 14 , 10146–10176 (2014).

Vanini, S., Faraci, F., Ferrari, A. & Giordano, S. Using barometric pressure data to recognize vertical displacement activities on smartphones. Comput. Commun. 87 , 37–48 (2016).

Wan, N. & Lin, G. Classifying human activity patterns from smartphone collected GPS data: a fuzzy classification and aggregation approach. Trans. GIS 20 , 869–886 (2016).

Gu, Y., Li, D., Kamiya, Y. & Kamijo, S. Integration of positioning and activity context information for lifelog in urban city area. Navigation 67 , 163–179 (2020).

Miao, F., He, Y., Liu, J., Li, Y. & Ayoola, I. Identifying typical physical activity on smartphone with varying positions and orientations. Biomed. Eng. Online 14 , 32 (2015).

Lee, Y.-S. & Cho, S.-B. Layered hidden Markov models to recognize activity with built-in sensors on Android smartphone. Pattern Anal. Appl. 19 , 1181–1193 (2016).

Martin, B. D., Addona, V., Wolfson, J., Adomavicius, G. & Fan, Y. Methods for real-time prediction of the mode of travel using smartphone-based GPS and accelerometer data. Sensors 17 , 2058 (2017).

Oshin, T. O., Poslad, S. & Zhang, Z. Energy-efficient real-time human mobility state classification using smartphones. IEEE Trans. Comput. 64 , 1680–1693 (2015).

Google Scholar  

Shin, D. et al. Urban sensing: Using smartphones for transportation mode classification. Comput. Environ. Urban Syst. 53 , 76–86 (2015).

Hur, T. et al. Iss2Image: a novel signal-encoding technique for CNN-based human activity recognition. Sensors 18 , 3910 (2018).

Gjoreski, M. et al. Classical and deep learning methods for recognizing human activities and modes of transportation with smartphone sensors. Inf. Fusion 62 , 47–62 (2020).

Wannenburg, J. & Malekian, R. Physical activity recognition from smartphone accelerometer data for user context awareness sensing. IEEE Trans. Syst. Man, Cybern. Syst. 47 , 3143–3149 (2017).

Yurur, O., Labrador, M. & Moreno, W. Adaptive and energy efficient context representation framework in mobile sensing. IEEE Trans. Mob. Comput. 13 , 1681–1693 (2014).

Li, P., Wang, Y., Tian, Y., Zhou, T.-S. & Li, J.-S. An automatic user-adapted physical activity classification method using smartphones. IEEE Trans. Biomed. Eng. 64 , 706–714 (2017).

PubMed   Google Scholar  

Awan, M. A., Guangbin, Z., Kim, C.-G. & Kim, S.-D. Human activity recognition in WSN: a comparative study. Int. J. Networked Distrib. Comput. 2 , 221–230 (2014).

Chen, Z., Zhu, Q., Soh, Y. C. & Zhang, L. Robust human activity recognition using smartphone sensors via CT-PCA and online SVM. IEEE Trans. Ind. Inform. 13 , 3070–3080 (2017).

Zhu, R. et al. Efficient human activity recognition solving the confusing activities via deep ensemble learning. IEEE Access 7 , 75490–75499 (2019).

Yang, R. & Wang, B. PACP: a position-independent activity recognition method using smartphone sensors. Inf 7 , 72 (2016).

Gani, M. O. et al. A light weight smartphone based human activity recognition system with high accuracy. J. Netw. Comput. Appl. 141 , 59–72 (2019).

Reddy, S. et al. Using mobile phones to determine transportation modes. ACM Trans. Sens. Networks 6 , 1–27 (2010).

Guidoux, R. et al. A smartphone-driven methodology for estimating physical activities and energy expenditure in free living conditions. J. Biomed. Inform. 52 , 271–278 (2014).

Cruciani, F. et al. Automatic annotation for human activity recognition in free living using a smartphone. Sensors 18 , 2203 (2018).

Micucci, D., Mobilio, M. & Napoletano, P. UniMiB SHAR: A dataset for human activity recognition using acceleration data from smartphones. Appl. Sci. 7 , 1101 (2017).

Derawi, M. & Bours, P. Gait and activity recognition using commercial phones. Comput. Secur. 39 , 137–144 (2013).

Gu, F., Khoshelham, K., Valaee, S., Shang, J. & Zhang, R. Locomotion activity recognition using stacked denoising autoencoders. IEEE Internet Things J. 5 , 2085–2093 (2018).

Chen, Y. & Shen, C. Performance analysis of smartphone-sensor behavior for human activity recognition. IEEE Access 5 , 3095–3110 (2017).

Javed, A. R. et al. Analyzing the effectiveness and contribution of each axis of tri-axial accelerometer sensor for accurate activity recognition. Sensors 20 , 2216 (2020).

Mukherjee, D., Mondal, R., Singh, P. K., Sarkar, R. & Bhattacharjee, D. EnsemConvNet: a deep learning approach for human activity recognition using smartphone sensors for healthcare applications. Multimed. Tools Appl. 79 , 31663–31690 (2020).

Avilés-Cruz, C., Ferreyra-Ramírez, A., Zúñiga-López, A. & Villegas-Cortéz, J. Coarse-fine convolutional deep-learning strategy for human activity recognition. Sensors 19 , 1556 (2019).

Guiry, J. J., van de Ven, P. & Nelson, J. Multi-sensor fusion for enhanced contextual awareness of everyday activities with ubiquitous devices. Sensors 14 , 5687–5701 (2014).

Saha, J., Chowdhury, C., Ghosh, D. & Bandyopadhyay, S. A detailed human activity transition recognition framework for grossly labeled data from smartphone accelerometer. Multimed. Tools Appl . https://doi.org/10.1007/s11042-020-10046-w (2020).

Ignatov, A. D. & Strijov, V. V. Human activity recognition using quasiperiodic time series collected from a single tri-axial accelerometer. Multimed. Tools Appl. 75 , 7257–7270 (2016).

Das, R. D. & Winter, S. Detecting urban transport modes using a hybrid knowledge driven framework from GPS trajectory. ISPRS Int. J. Geo-Information 5 , 207 (2016).

Arif, M., Bilal, M., Kattan, A. & Ahamed, S. I. Better physical activity classification using smartphone acceleration sensor. J. Med. Syst. 38 , 95 (2014).

Heng, X., Wang, Z. & Wang, J. Human activity recognition based on transformed accelerometer data from a mobile phone. Int. J. Commun. Syst. 29 , 1981–1991 (2016).

Gao, Z., Liu, D., Huang, K. & Huang, Y. Context-aware human activity and smartphone position-mining with motion sensors. Remote Sensing 11 , 2531 (2019).

Kang, J., Kim, J., Lee, S. & Sohn, M. Transition activity recognition using fuzzy logic and overlapped sliding window-based convolutional neural networks. J. Supercomput. 76 , 8003–8020 (2020).

Shojaedini, S. V. & Beirami, M. J. Mobile sensor based human activity recognition: distinguishing of challenging activities by applying long short-term memory deep learning modified by residual network concept. Biomed. Eng. Lett. 10 , 419–430 (2020).

Mairittha, N., Mairittha, T. & Inoue, S. On-device deep personalization for robust activity data collection. Sensors 21 , 41 (2021).

Khan, A. M., Siddiqi, M. H. & Lee, S.-W. Exploratory data analysis of acceleration signals to select light-weight and accurate features for real-time activity recognition on smartphones. Sensors 13 , 13099–13122 (2013).

Ebner, M., Fetzer, T., Bullmann, M., Deinzer, F. & Grzegorzek, M. Recognition of typical locomotion activities based on the sensor data of a smartphone in pocket or hand. Sensors 20 , 6559 (2020).

Voicu, R.-A., Dobre, C., Bajenaru, L. & Ciobanu, R.-I. Human physical activity recognition using smartphone sensors. Sensors 19 , 458 (2019).

Hur, T., Bang, J., Kim, D., Banos, O. & Lee, S. Smartphone location-independent physical activity recognition based on transportation natural vibration analysis. Sensors 17 , 931 (2017).

Bashir, S. A., Doolan, D. C. & Petrovski, A. The effect of window length on accuracy of smartphone-based activity recognition. IAENG Int. J. Comput. Sci. 43 , 126–136 (2016).

Lu, D.-N., Nguyen, D.-N., Nguyen, T.-H. & Nguyen, H.-N. Vehicle mode and driving activity detection based on analyzing sensor data of smartphones. Sensors 18 , 1036 (2018).

Wang, G. et al. Impact of sliding window length in indoor human motion modes and pose pattern recognition based on smartphone sensors. Sensors 18 , 1965 (2018).

Jain, A. & Kanhangad, V. Human activity classification in smartphones using accelerometer and gyroscope sensors. IEEE Sens. J. 18 , 1169–1177 (2018).

Bedogni, L., Di Felice, M. & Bononi, L. Context-aware Android applications through transportation mode detection techniques. Wirel. Commun. Mob. Comput. 16 , 2523–2541 (2016).

Ferreira, P., Zavgorodnii, C. & Veiga, L. edgeTrans—edge transport mode detection. Pervasive Mob. Comput. 69 , 101268 (2020).

Gu, F., Kealy, A., Khoshelham, K. & Shang, J. User-independent motion state recognition using smartphone sensors. Sensors 15 , 30636–30652 (2015).

Li, X., Wang, Y., Zhang, B. & Ma, J. PSDRNN: an efficient and effective HAR scheme based on feature extraction and deep learning. IEEE Trans. Ind. Inform. 16 , 6703–6713 (2020).

Zhao, B., Li, S., Gao, Y., Li, C. & Li, W. A framework of combining short-term spatial/frequency feature extraction and long-term IndRNN for activity recognition. Sensors 20 , 6984 (2020).

Huang, E. J. & Onnela, J.-P. Augmented movelet method for activity classification using smartphone gyroscope and accelerometer data. Sensors 20 , 3706 (2020).

Montero Quispe, K. G., Sousa Lima, W., Macêdo Batista, D. & Souto, E. MBOSS: a symbolic representation of human activity recognition using mobile sensors. Sensors 18 , 4354 (2018).

Sousa Lima, W., de Souza Bragança, H. L., Montero Quispe, K. G. & Pereira Souto, E. J. Human activity recognition based on symbolic representation algorithms for inertial sensors. Sensors 18 , 4045 (2018).

Bragança, H., Colonna, J. G., Lima, W. S. & Souto, E. A smartphone lightweight method for human activity recognition based on information theory. Sensors 20 , 1856 (2020).

Saeedi, S. & El-Sheimy, N. Activity recognition using fusion of low-cost sensors on a smartphone for mobile navigation application. Micromachines 6 , 1100–1134 (2015).

Bilal, M., Shaikh, F. K., Arif, M. & Wyne, M. F. A revised framework of machine learning application for optimal activity recognition. Clust. Comput. 22 , 7257–7273 (2019).

Shi, D., Wang, R., Wu, Y., Mo, X. & Wei, J. A novel orientation- and location-independent activity recognition method. Pers. Ubiquitous Comput. 21 , 427–441 (2017).

Antos, S. A., Albert, M. V. & Kording, K. P. Hand, belt, pocket or bag: practical activity tracking with mobile phones. J. Neurosci. Methods 231 , 22–30 (2014).

Shi, J., Zuo, D. & Zhang, Z. Transition activity recognition system based on standard deviation trend analysis. Sensors 20 , 3117 (2020).

Garcia-Gonzalez, D., Rivero, D., Fernandez-Blanco, E. & Luaces, M. R. A public domain dataset for real-life human activity recognition using smartphone sensors. Sensors 20 , 2200 (2020).

Saeedi, S., Moussa, A. & El-Sheimy, N. Context-aware personal navigation using embedded sensor fusion in smartphones. Sensors 14 , 5742–5767 (2014).

Zhang, W., Zhao, X. & Li, Z. A comprehensive study of smartphone-based indoor activity recognition via Xgboost. IEEE Access 7 , 80027–80042 (2019).

Ferrari, A., Micucci, D., Mobilio, M. & Napoletano, P. On the personalization of classification models for human activity recognition. IEEE Access 8 , 32066–32079 (2020).

Zhou, B., Yang, J. & Li, Q. Smartphone-based activity recognition for indoor localization using a convolutional neural network. Sensors 19 , 621 (2019).

Pires, I. M. et al. Pattern recognition techniques for the identification of activities of daily living using a mobile device accelerometer. Electronics 9 , 509 https://www.mdpi.com/2079-9292/9/3/509#cite (2020).

Alo, U. R., Nweke, H. F., Teh, Y. W. & Murtaza, G. Smartphone motion sensor-based complex human activity identification using deep stacked autoencoder algorithm for enhanced smart healthcare system. Sensors 20 , 6300 https://www.mdpi.com/2079-9292/9/3/509#cite (2020).

Otebolaku, A. M. & Andrade, M. T. User context recognition using smartphone sensors and classification models. J. Netw. Comput. Appl. 66 , 33–51 (2016).

Zhuo, S. et al. Real-time smartphone activity classification using inertial sensors—recognition of scrolling, typing, and watching videos while sitting or walking. Sensors 20 , 655 (2020).

Asim, Y., Azam, M. A., Ehatisham-ul-Haq, M., Naeem, U. & Khalid, A. Context-aware human activity recognition (CAHAR) in-the-wild using smartphone accelerometer. IEEE Sens. J. 20 , 4361–4371 (2020).

Zhao, Z., Chen, Z., Chen, Y., Wang, S. & Wang, H. A class incremental extreme learning machine for activity recognition. Cogn. Comput. 6 , 423–431 (2014).

Abdallah, Z. S., Gaber, M. M., Srinivasan, B. & Krishnaswamy, S. Adaptive mobile activity recognition system with evolving data streams. Neurocomputing 150 , 304–317 (2015).

Guo, H., Chen, L., Chen, G. & Lv, M. Smartphone-based activity recognition independent of device orientation and placement. Int. J. Commun. Syst. 29 , 2403–2415 (2016).

Cruciani, F. et al. Personalizing activity recognition with a clustering based semi-population approach. IEEE ACCESS 8 , 207794–207804 (2020).

Saha, J., Ghosh, D., Chowdhury, C. & Bandyopadhyay, S. Smart handheld based human activity recognition using multiple instance multiple label learning. Wirel. Pers. Commun . https://doi.org/10.1007/s11277-020-07903-0 (2020).

Mohamed, R., Zainudin, M. N. S., Sulaiman, M. N., Perumal, T. & Mustapha, N. Multi-label classification for physical activity recognition from various accelerometer sensor positions. J. Inf. Commun. Technol. 17 , 209–231 (2018).

Wang, C., Xu, Y., Liang, H., Huang, W. & Zhang, L. WOODY: a post-process method for smartphone-based activity recognition. IEEE Access 6 , 49611–49625 (2018).

Garcia-Ceja, E. & Brena, R. F. An improved three-stage classifier for activity recognition. Int. J. Pattern Recognit. Artif. Intell . 32 , 1860003 (2018).

Ravi, D., Wong, C., Lo, B. & Yang, G.-Z. A deep learning approach to on-node sensor data analytics for mobile or wearable devices. IEEE J. Biomed. Heal. Inform. 21 , 56–64 (2017).

Chen, Z., Jiang, C. & Xie, L. A novel ensemble ELM for human activity recognition using smartphone sensors. IEEE Trans. Ind. Inform. 15 , 2691–2699 (2019).

van Hees, V. T., Golubic, R., Ekelund, U. & Brage, S. Impact of study design on development and evaluation of an activity-type classifier. J. Appl. Physiol. 114 , 1042–1051 (2013).

Sasaki, J. et al. Performance of activity classification algorithms in free-living older adults. Med. Sci. Sports Exerc. 48 , 941–950 (2016).

Allyn, B. IBM abandons facial recognition products, condemns racially biased surveillance. https://www.npr.org/2020/06/09/873298837/ibm-abandons-facial-recognition-products-condemns-racially-biased-surveillance (2020).

Chee, F. Y. EU mulls five-year ban on facial recognition tech in public areas. https://www.reuters.com/article/uk-eu-ai/eu-mulls-five-year-ban-on-facial-recognition-tech-in-public-areas-idINKBN1ZF2QN (2020).

Saha, J., Chowdhury, C., Chowdhury, I. R., Biswas, S. & Aslam, N. An ensemble of condition based classifiers for device independent detailed human activity recognition using smartphones. Information 9 , 94 (2018).

Barnett, I. & Onnela, J.-P. Inferring mobility measures from GPS traces with missing data. Biostatistics 21 , e98–e112 (2020).

Wang, L. et al. Enabling reproducible research in sensor-based transportation mode recognition with the Sussex-Huawei dataset. IEEE ACCESS 7 , 10870–10891 (2019).

Lee, M. H., Kim, J., Jee, S. H. & Yoo, S. K. Integrated solution for physical activity monitoring based on mobile phone and PC. Healthc. Inform. Res. 17 , 76–86 (2011).

Fahim, M., Fatima, I., Lee, S. & Park, Y.-T. EFM: evolutionary fuzzy model for dynamic activities recognition using a smartphone accelerometer. Appl. Intell. 39 , 475–488 (2013).

Yurur, O., Liu, C. H. & Moreno, W. Light-weight online unsupervised posture detection by smartphone accelerometer. IEEE Internet Things J. 2 , 329–339 (2015).

Awan, M. A., Guangbin, Z., Kim, H.-C. & Kim, S.-D. Subject-independent human activity recognition using Smartphone accelerometer with cloud support. Int. J. Ad Hoc Ubiquitous Comput. 20 , 172–185 (2015).

Chen, Z., Wu, J., Castiglione, A. & Wu, W. Human continuous activity recognition based on energy-efficient schemes considering cloud security technology. Secur. Commun. Netw. 9 , 3585–3601 (2016).

Guo, J. et al. Smartphone-based patients’ activity recognition by using a self-learning scheme for medical monitoring. J. Med. Syst. 40 , 140 (2016).

Walse, K. H., Dharaskar, R. V. & Thakare, V. M. A study of human activity recognition using AdaBoost classifiers on WISDM dataset. IIOAB J. 7 , 68–76 (2016).

Lee, K. & Kwan, M.-P. Physical activity classification in free-living conditions using smartphone accelerometer data and exploration of predicted results. Comput. Environ. Urban Syst. 67 , 124–131 (2018).

Ahmad, N. et al. SARM: salah activities recognition model based on smartphone. Electronics 8 , 881 (2019).

Usman Sarwar, M. et al. Recognizing physical activities having complex interclass variations using semantic data of smartphone. Softw. Pract. Exp . 51 , 532–549 (2020).

Kwapisz, J. R., Weiss, G. M. & Moore, S. A. Activity recognition using cell phone accelerometers. in Proceedings of the Fourth International Workshop on Knowledge Discovery from Sensor Data 10–18 https://doi.org/10.1145/1964897.1964918 (Association for Computing Machinery, 2010).

Sharma, A., Singh, S. K., Udmale, S. S., Singh, A. K. & Singh, R. Early transportation mode detection using smartphone sensing data. IEEE Sens. J . 1 , https://doi.org/10.1109/JSEN.2020.3009312 (2020).

Chen, Z. et al. Smartphone sensor-based human activity recognition using feature fusion and maximum full a posteriori. IEEE Trans. Instrum. Meas. 69 , 3992–4001 (2020).

Vavoulas, G. Chatzaki, C. Malliotakis, T. Pediaditis, M. & Tsiknakis, M. The MobiAct Dataset: recognition of activities of daily living using smartphones. In Proceedings of the International Conference on Information and Communication Technologies for Ageing Well and e-Health (eds. Röcker, C., Ziefle, M., O’Donoghue, J, Maciaszek, L. & Molloy W.) Vol. 1: ICT4AWE, (ICT4AGEINGWELL 2016) 143–151, https://www.scitepress.org/ProceedingsDetails.aspx?ID=VhZYzluZTNE=&t=1 (SciTePress, 2016).

Shoaib, M., Bosch, S., Incel, O. D., Scholten, H. & Havinga, P. J. M. Complex human activity recognition using smartphone and wrist-worn motion sensors. Sensors 16 , 426 (2016).

Lockhart, J. W. et al. Design considerations for the WISDM smart phone-based sensor mining architecture. in Proceedings of the Fifth International Workshop on Knowledge Discovery from Sensor Data. 25–33 https://doi.org/10.1145/2003653.2003656 (Association for Computing Machinery, 2011).

Vaizman, Y., Ellis, K. & Lanckriet, G. Recognizing detailed human context in the wild from smartphones and smartwatches. IEEE Pervasive Comput. 16 , 62–74 (2017).

Sztyler, T. & Stuckenschmidt, H. On-body localization of wearable devices: an investigation of position-aware activity recognition. in 2016 IEEE International Conference on Pervasive Computing and Communications (PerCom) 1–9 https://ieeexplore.ieee.org/document/7456521 (IEEE, 2016).

Malekzadeh, M., Clegg, R. G., Cavallaro, A. & Haddadi, H. Mobile sensor data anonymization. in Proceedings of the International Conference on Internet of Things Design and Implementation. 49–58 https://doi.org/10.1145/3302505.3310068 (ACM, 2019).

Carpineti, C., Lomonaco, V., Bedogni, L., Felice, M. D. & Bononi, L. Custom dual transportation mode detection by smartphone devices exploiting sensor diversity. in 2018 IEEE International Conference on Pervasive Computing and Communications Workshops (PerCom Workshops) 367–372 https://ieeexplore.ieee.org/document/8480119 (IEEE, 2018).

Ichino, H., Kaji, K., Sakurada, K., Hiroi, K. & Kawaguchi, N. HASC-PAC2016: large scale human pedestrian activity corpus and its baseline recognition. in Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing: Adjunct. 705–714 https://doi.org/10.1145/2968219.2968277 (Association for Computing Machinery, 2016).

Download references

Acknowledgements

Drs. Straczkiewicz and Onnela are supported by NHLBI award U01HL145386 and NIMH award R37MH119194. Dr. Onnela is also supported by the NIMH award U01MH116928. Dr. James is supported by NCI award R00CA201542 and NHLBI award R01HL150119.

Author information

Authors and affiliations.

Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, MA, 02115, USA

Marcin Straczkiewicz & Jukka-Pekka Onnela

Department of Population Medicine, Harvard Medical School and Harvard Pilgrim Health Care Institute, Boston, MA, 02215, USA

Peter James

Department of Environmental Health, Harvard T.H. Chan School of Public Health, Boston, MA, 02115, USA

You can also search for this author in PubMed   Google Scholar

Contributions

M.S. conducted the review, prepared figures, and wrote the initial draft. P.J. and J.P.O. revised the manuscript. J.P.O. supervised the project. All authors reviewed the manuscript.

Corresponding author

Correspondence to Marcin Straczkiewicz .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Straczkiewicz, M., James, P. & Onnela, JP. A systematic review of smartphone-based human activity recognition methods for health research. npj Digit. Med. 4 , 148 (2021). https://doi.org/10.1038/s41746-021-00514-4

Download citation

Received : 24 March 2021

Accepted : 13 September 2021

Published : 18 October 2021

DOI : https://doi.org/10.1038/s41746-021-00514-4

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

This article is cited by

Unlocking the potential of smartphone and ambient sensors for adl detection.

• Marija Stojchevska
• Mathias De Brouwer
• Femke Ongenae

Scientific Reports (2024)

Self-supervised learning for human activity recognition using 700,000 person-days of wearable data

• Aiden Doherty

npj Digital Medicine (2024)

Activity recognition based on smartphone sensor data using shallow and deep learning techniques: A Comparative Study

• Asif Iqbal Middya
• Sarvajit Kumar
• Sarbani Roy

Multimedia Tools and Applications (2024)

A “one-size-fits-most” walking recognition method for smartphones, smartwatches, and wearable accelerometers

• Marcin Straczkiewicz
• Emily J. Huang
• Jukka-Pekka Onnela

npj Digital Medicine (2023)

Data filtering and deep learning for enhanced human activity recognition from UWB radars

• Julien Maitre
• Kévin Bouchard
• Sébastien Gaboury

Journal of Ambient Intelligence and Humanized Computing (2023)

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• Sensors (Basel)

Smart Roads for Autonomous Accident Detection and Warnings

Abdul mateen.

1 Information and Communication Engineering Department, Yeungnam University, Gyeongsan 38541, Korea; kp.ude.bsitsauuf@neetamludba (A.M.); rk.ca.uny@girtahknayaran (N.K.)

2 Department of Computer Science, Federal Urdu University of Arts, Science & Technology, Islamabad 45570, Pakistan; moc.liamg@3891finahdihazdammahum

Muhammad Zahid Hanif

Narayan khatri, sihyung lee.

3 School of Computer Science and Engineering, Kyungpook National University, Daegu 41566, Korea; rk.ca.unk@eelgnuyhis

Seung Yeob Nam

Associated data.

Not applicable.

An increasing number of vehicles on the roads increases the risk of accidents. In bad weather (e.g., heavy rainfall, strong winds, storms, and fog), this risk almost doubles due to bad visibility as well as road conditions. If an accident happens, especially in bad weather, it is important to inform approaching vehicles about it. Otherwise, there might be another accident, i.e., a multiple-vehicle collision (MVC). If the Emergency Operations Center (EOC) is not informed in a timely fashion about the incident, fatalities might increase because they do not receive immediate first aid. Detecting humans or animals would undoubtedly provide us with a better answer for reducing human fatalities in traffic accidents. In this research, an accident alert light and sound (AALS) system is proposed for auto accident detection and alerts with all types of vehicles. No changes are required in non-equipped vehicles (nEVs) and EVs because the system is installed on the roadside. The idea behind this research is to make smart roads (SRs) instead of equipping each vehicle with a separate system. Wireless communication is needed only when an accident is detected. This study is based on different sensors that are used to build SRs to detect accidents. Pre-saved locations are used to reduce the time needed to find the accident’s location without the help of a global positioning system (GPS). Additionally, the proposed framework for the AALS also reduces the risk of MVCs.

1. Introduction

Traveling is one of the basic needs of every person who lives in cities or villages. There are several ways to travel from one place to another by air, water, rail, and road in various types of vehicle, e.g., cars, motorbikes, buses, and trucks. Roads are the foremost source of linking between cities and villages. Due to the ease in traveling by road, vehicles have become the main way people travel. The chances of vehicular accidents (Vas) have increased with the growing number of vehicles on the roads. During a journey, one does not know what will happen on the next road, particularly during bad weather conditions (BWC). In such a situation, driving can be difficult due to bad visibility, which can lead to an accident. It was also noticed that in BWC, multiple vehicle collisions (MVCs) can occur owing to delays in receiving information about an incident. According to one study by the Islamabad police, there were 9582 accidents from 2016 to 2017 all over Pakistan, involving 11,317 vehicles, leading to 5047 fatalities and 12,696 persons injured [ 1 ].

Vehicles can be divided into two main groups: equipped vehicles (Evs) and non-equipped vehicles (nEVs). Evs have sensing capabilities to avoid or detect accidents. Evs include vehicles equipped with a smartphone-based application or a microcontroller with different sensors. It uses GSM, LTE, or 5G to send messages; a GPS for finding locations; and GPRS, LTE, or 5G for internet connectivity. All old vehicles with no capability for sensing an accident are nEVs. Thus, the benefits from accident detection and alerts are not provided in nEVs. A question arises about why we rely on vehicle sensors or smartphone-based systems. The GSM signal is weak in many distant areas, and communication links might be unstable in those areas. On the other hand, a GPS requires 10 to 15 min to fix a location for the first time, which leads to late broadcasts. The main goal of this research is to:

• 1. introduce a new framework of smart road based on multiple sensors to save the lives of people injured in an accident, and protect people and vehicles against MVCs;
• 2. detect Vas autonomously without using vehicular sensors;
• 3. alert drivers of approaching vehicles about an accident, even without vehicular communications;
• 4. inform an Emergency Operations Center (EOC) about an accident and its location without needing a GPS.

According to one dictionary, the word accident is defined as “an uncertain event which may lead to injury, loss of life, or property damage” [ 2 ]. An accident is also defined as “an event in which one or more vehicles hit in such a way that may lead to loss of life or injury or vehicle damage, resulting in a traffic blockage on the road”. The risks for loss of life, injuries, and other damage may increase if an incident is not reported to an EOC in a timely fashion. Lives can be saved by sending timely information about an accident through an automated mechanism. Moreover, quick automobile accident detection and an alert system are required to protect approaching vehicles against an MVC. Several methods have been implemented in advanced vehicles (Avs) for avoiding an accident. An accident threat is detected through sensors installed in vehicles or by using smartphone sensors. Previous researchers have used accelerometers, smoke detectors, infrared (IR) obstacle sensors, proximity sensors, and biosensors to detect an accident. A microcontroller board collects sensor information. A GPS receiver is used for location tracking and monitoring speeds, while GSM, LTE, and 5G technologies send information about an accident. Event information can include the location and time, videos, and images of the event for the EOC. There are two methods to detect accidents autonomously; we can equip all vehicles, or we can develop smart roads that can sense any type of accident and convey messages wirelessly to an EOC. To protect vehicles around the accident area, a new framework is proposed, i.e., an accident alert light and sound (AALS) system. In this framework, drivers of oncoming vehicles are informed by blinking lights and a siren after detection of an accident. This blinking light is so bright that it can be seen easily in BWC. Wireless communications on these smart roads (SRs) is provided with the HC-12 serial communication module of the Arduino Uno board. These SRs act as live roads because of their sensing power.

The communication protocols can be GSM or 5G technology in vehicle-to-everything (V2X) systems [ 3 ] to send messages about an accident. V2X can be realized from integration of cellular 5G and new radio (NR) access technology (i.e., 5G NR V2X). Mobile-phone-based applications use built-in sensor data to detect accidents. One of the major drawbacks of using mobile phone sensors for accident detection is false positives, i.e., incorrectly detecting an accident. The sensors used in mobile phones and Avs need calibration, which is difficult to achieve or may not be possible. When a car jerks due to an accident, mobile technology might turn off, or a battery might be displaced, preventing information about the accident from reaching an EOC in a timely manner.

In this research, an AALS system is proposed to detect automobile accidents and send alerts to other vehicles. No changes are required to nEVs and Evs, because the whole system is installed on the roadside, leading to the new concept of smart roads. A smart road has the capability to detect an incident on the road and send this information to the EOC. In case of an accident, the AALS system uses the event-driven wireless sensor network (EDWSN) protocol [ 4 ] for communication between nodes. Otherwise, all the nodes are isolated from each other. Each node works independently to detect an accident. When an accident happens on part of an SR, the AALS system detects it through its sensing capabilities, and generates light and sound alerts. The purpose of these alerts is to inform drivers of approaching vehicles about the accident so they can take precautionary measures, e.g., reduce speed or stop to avoid the accident. In this way, the risk of MVCs is reduced, leading to fewer fatalities and less vehicle damage. With the help of wireless communications, an accident event alert message is sent to a nearby node in a chain process using the EDWSN protocol until it reaches the EOC. The communication takes less than a minute or two to reach the EOC from the node detecting the accident. The message received by the EOC contains all the important information about the location of the node detecting the accident with the complete address, direction of the road, and a fire alert (if fire is also detected). The EOC sends a rescue team with an ambulance and fire brigade (whatever is required) to the accident’s location to help save lives. The AALS system cannot reset itself; it is reset by the EOC rescue team after operations are performed and the road is clear. To reset the AALS system, a hard-reset button is pressed on the node detecting the accident. Then, the reset message is generated and sent to other nodes in a chain process until it reaches the EOC. After forwarding the message, each node resets itself and returns to the normal working state.

Main contributions of this research are:

• 1. A brief survey on the state of the art related to pre-accident as well as post-accident models, frameworks, and techniques;
• 2. Identification and reporting of limitations in previous studies related to accident detection;
• 3. The concept of a smart road with an event-sensing capability, plus implementation and testing through various experiments;
• 4. Demonstration of a new and modern way to quickly detect accidents and communicate with nearby vehicles and EOCs.

The remainder of the paper is arranged as follows. Section 2 describes previous work related to the detection of accidents and MVCs on roads. It discusses previous studies and highlights their limitations. Section 3 discusses the problem background, and provides insight into the proposed AALS and its main components. Section 4 outlines details on the hardware and experimental setup for the AALS. Section 5 presents the experiments conducted and their outcomes (with graphs). Finally, Section 6 concludes the paper and suggests future research directions.

2. Related Works

A large amount of research is being carried out in the domain of accident avoidance and accident alarms by a large number of researchers and practitioners. In order to avoid accidents, many approaches are utilized to enhance safety. For ease of reference, the literature on accident detection and avoidance is separated into three approaches: stand-alone, cooperative, and hybrid. Stand-alone approaches use sensors, such as radar and light detection and ranging (LiDAR), for accident avoidance and detection, whereas cooperative approaches rely on V2X technology, and hybrid approaches combine the two methods.

2.1. Stand-Alone Approaches

A system based on a programmable integrated circuit (PIC) microcontroller and an ultrasonic sensor installed in vehicles was proposed by Govindarajulu and Ezhumalai [ 5 ]. The ultrasonic sensor’s role is to detect any obstruction, speed bump, steep curve, or humans on the road. The sensor’s input is processed by the microcontroller, which acts as a computing device. If an abnormality is discovered, vehicle drivers are notified so they can take appropriate action, such as slowing down or changing lanes. The time delay is controlled by the microprocessor, and ultrasonic sensors measure the distance by using an echo produced by an impediment to a high-frequency sound pulse. A voice message is created if the distance between the vehicle and the obstacle is less than a preset threshold.

Bahgat et al. [ 6 ] proposed model-based visual sensing for avoidance of vehicular accidents. Image processing is implemented for detection of a nearby vehicle to maintain a safe distance by alerting the driver. In their work, a backend server monitors the behavior of various advanced vehicles through sensors and mobile communications. All the calculations and estimations are performed on the server. When the server finds any chance of a collision, it provides alerts to drivers. The proposed system is complicated and involves high computational costs.

Rapp et al. proposed a technique that utilizes a short-range radar for self-driving vehicles by using normal distribution transform (NDT) [ 7 ]. The radar is used to predict lane changes with simultaneous localization and mapping (SLAM) and landmark extraction. A semi-Markov process is adopted to model the varying conditions and environments in a 2D representation. The main contribution of their work is a grid map representation where every cell’s dynamics are modeled as a dynamic process. However, their proposed idea can solve the problem of localization only in a limited area.

Barjenbruch et al. [ 8 ] estimated the velocity and yaw rate (angular velocity during rotation) with a spatial-data radar sensor. They optimized the discrepancy in position from radar detection and Doppler velocity (to calculate the expected Doppler shift for existing pose hypothesis to each mapped landmark) for w.r.t. moving vehicles. The implementation of this approach in real systems is difficult due to high computation costs. Moreover, even small errors in methods that calculate relative motion between consecutive sensor readings will accumulate when summed over a longer time [ 9 ].

In order to drive safely, Hyung et al. [ 10 ] used cruise control technology in self-driving automobiles. In the cruise control system, the vehicle can detect and maintain a safe distance from the car ahead of it. Brakes and acceleration are controlled by using the proportional-integral-derivative (PID) approach, which calculates three basic coefficients to produce optimal responses. The PID approach also adjusts the accelerator and brake pedals to maintain a safe distance between vehicles. Distance errors caused by a preceding vehicle can be determined from distance information obtained by a laser scanner installed on the front of a vehicle. Acceleration control in each successive vehicle provides velocity control by transmitting acceleration position sensor (APS) signals, which can be artificially generated by the speed control unit to the electronic control unit (ECU). Their work was tested using simulators (Carsim and Simulink), and proved highly useful.

An automated deep learning (DL)-based system [ 11 ] was developed for detecting accidents from video data. The system uses visual components in temporal order to represent traffic collisions. As a result, the model architecture is composed of a visual-features-extraction phase followed by transient pattern identification. Convolution and recurrent layers are used in the training phase to learn visual and temporal features. In public traffic accident datasets, accuracy of 98% was attained in detection of accidents, demonstrating a strong capacity for detection independent of the road structure. The solution is limited to automobile crashes, not motorbikes, bicycles, and pedestrians. Furthermore, the model makes mistakes when determining accident segments under poor illumination (e.g., at night), at low resolutions, and when there are occlusions.

A system suggested by Kim et al. [ 12 ] includes a LiDAR sensor to measure speed, and a prediction algorithm based on the stop sight distance (SSD) formula to calculate braking distance. Red light running (RLR) is predicted by a sensor that monitors the speed of the vehicle approaching an intersection. The results demonstrated that RLR can be predicted from the front and side of a moving vehicle using LiDAR and the SSD algorithm. The vehicle’s weight is fixed in this method; however, in practice, the braking distance varies depending on the vehicle’s weight.

Xie et al. [ 13 ] suggested a method for recognizing and tracking obstacles in 3D LiDAR point clouds. The point clouds generated by 3D LiDAR after road segmentation are first rasterized, and then additional relevant cells are added. A clustering technique is then used to group the obstacle cells in the next stage. Static obstacles are discovered using a multi-frame fusion approach by evaluating the clusters. Finally, moving obstacles are tracked using an upgraded dynamic tracking point model and a regular Kalman filter. The proposed method was tested in a variety of demanding settings, but has yet to deal with scenarios where barriers are extremely near one another.

A systematic approach to multi-LiDAR data processing [ 14 ] was introduced, consisting of calibration, filtering, clustering, and classification. Within the multi-LiDAR architecture, the accuracy of obstacle detection is enhanced by using noise filtering and clustering. The applied filtering method is based on sampling point occupancy rates (Ors). An adaptive search (AS) approach is used to improve density-based spatial clustering of applications with noise (DBSCAN). Furthermore, when AS-DBSCAN is combined with the proposed OR-based filtering, more robust and precise obstacle detection is achieved. The agglomerative hierarchical clustering technique with complete linkage is applied. The same procedure is then employed with a single linkage. If the two targets occupy neighboring segments or share the same segment, and their longitudinal separation is less than 0.8 m, the clustering algorithm utilized might aggregate detection of vehicles into a cluster, resulting in a merged/unresolved report. An inside perception test and an on-road test were performed on a fully instrumented, autonomous hybrid electric automobile. Experiments proved that the proposed algorithms are reliable and practical.

Using LiDAR, a mechanism for estimating the contour of approaching vehicles in the pre-crash phase was explored [ 15 ]. The data are initially combined, and an environmental model containing the identified items and their properties, such as position and shape, is computed. The mechanism extracts the main parameters of the vehicles from a LiDAR 3D point cloud through a convex hull algorithm. The relevant vehicle contour and longitudinal axis are then calculated. On both static and dynamic measures, tests on vehicles revealed good accuracy and durability of the estimation. Despite the fact that a convex hull algorithm examines a set of points from the reflection point clusters’ outermost region, it defines its bounds by connecting two neighboring points with straight lines. The accuracy of contour prediction may be affected by the fact that a vehicle’s contour is often a higher degree curve. Furthermore, due to noise-induced change in reflection points, evaluating individual points for analysis can introduce mistakes [ 16 ].

2.2. Cooperative Approaches

Jerry et al. [ 17 ] presented a vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communication system that leverages short-range communication technologies. The on-board unit (OBU) and the roadside unit (RSU) are the two primary components of this system. The OBD includes a processor, vehicle sensors, a touch screen, speakers, radio link modules, Bluetooth, and a GPS module. The OBU is mounted on vehicles and connected with the system and the in-vehicle network. The primary function of the OBU is to offer vehicle-to-vehicle communication and to gather diagnostic messages. The RSU is a stationary unit installed on roadways and in other conspicuous locations, such as gas stations and toll plazas, to facilitate V2I communication. The RSU receives and sends information to the OBU for ITS and internet services. Messages about collisions or speed limits are shown on a display screen.

A survey by Eskandarian et al. [ 18 ] discussed the latest techniques and algorithms in connected and automated vehicles (CAVs) for the identification, perception, planning, and control of vehicles. They addressed remarkable issues related to decision making on path tracking and control of vehicle communications in a cooperative environment. They mentioned how, in a CAV, stability and loss of signals is the major problem in wireless communications. Finally, issues and challenges unsolved in CAVs were outlined for researchers and vendors.

A framework for safe lane changing in connected and automated vehicles was proposed by Zheng et al. [ 19 ]. Their lane-changing approach was based on V2V and V2I communications for better movement and traffic operations. The study suggests that changing lanes is risky when the vehicle’s declaration is more than the threshold value of a leading vehicle. Total time, surrogate safety measures, and traffic waves are among the characteristics used to make lane-changing decisions. A heuristic technique is used to identify the lane-changing zones before simulation. A cooperative strategy is implemented near highway off-ramps via vehicle-based control of the aforementioned applications. The main limitation of this work is that it did not consider the environment of human-driven vehicles and the diversity in calculating lane-changing behaviors in different and unexplored places.

A cooperative neighboring vehicle positioning system (CNVPS) [ 20 ] was introduced by Nam et al. that leverages GPS technology and other sensors to improve localization. The GPS is used to provide locations owing to its availability 24 h a day, 7 days a week, and in any weather. The technology uses a local map to identify and interact with nearby cars (using V2V communication) in order to obtain an accurate relative position. It uses maximum likelihood estimation to obtain the correct position of a nearby vehicle. The suggested that CNVPS refreshes positions faster; however, it has a high deployment cost and can only be used for already predicted behavior. To overcome GPS refinement concerns, a DL-based cooperative vehicle localization algorithm, called a graph convolution network-CNVPS (GCN-CNVPS), was proposed [ 21 ]. To facilitate V2V communication, the technique uses GPS technology to obtain the coordinates, relative distances, angles to neighboring vehicles, and received signal strength indication (RSSI) information from a basic safety message (BSM). Three DL-based approaches that require less hardware were presented to address the challenge of localization. Finally, comparison results showed that the GCN-CNVPS effectively utilizes temporal and spatial correlations with lower overhead.

Through fuzzy logic and the Dempster–Shafer Theory, a predictive multi-criteria technique [ 22 ] was employed for multi-criteria decision-making in Avs following other vehicles and for lane changes. Dempster’s rule is used to identify risk values for hypotheses and trajectories, while fuzzy logic is used to handle uncertain situations and nonlinear data. All calculations are based on parameters such as accessibility, traffic rules, passenger preferences (for comfort), safety, and energy consumption. The fuzzy rules are based on data learning in real time and on traffic rules. The impact of sensors and their relative relevance to criteria and categories, ways to relax if it is safer, and driver preferences are considered for a more flexible decision. This method is scalable and not restricted to the circumstances presented. Using a collision avoidance algorithm (CAA) and a collision avoidance strategy, the authors suggested an enhanced collision avoidance (eCA) technique [ 23 ]. The eCA technique is responsible for identifying vehicles’ trajectories in order to avoid crashes. The next vehicles’ trajectories are anticipated in particular by projecting the positions of previous vehicles on curved or straight segments through status indicated by blinking lights. The CAS, on the other hand, regulates the speed of the vehicle in order to avoid accidents. A simulation was run on SUMO and NS-3, demonstrating the effectiveness of the strategy by avoiding nearly every collision under all of the tested circumstances.

Ijjina et al. proposed an object-tracking-based algorithm approach to detect accidents (termed Mask-RCNN) [ 24 ]. The chance of an accident is computed using the speed and trajectory of the vehicle. On CCTV surveillance footage, it has a high detection rate and a low false alarm rate. The model was tested in a variety of situations (including rain, storms, and low light), and proved effective. They also noted a low false alarm rate of 0.53%.

An accident management system was proposed in [ 25 ] that makes use of cellular technology in public transportation. This method enables communication across various components, including those in ambulances, RSUs, and servers. Furthermore, in this system, an optimal route-planning algorithm (ORPA) is proposed to optimize aggregate spatial utilization of road networks while lowering the travel cost to operate a vehicle. The ORPA was evaluated through simulations, and findings were compared with other current algorithms. In congested areas, the proposed method can also be used to offer fast routes for ambulances. All vehicles, including ambulances, are required to have a route indicator installed, as well as the ability to use remote correspondence. The ORPA outperformed in terms of average speed and travel duration, according to the evaluation data. The proposed system only works for predicted patterns and can fail due to unpredicted behavior of traffic.

With the advancement of 5G technology, V2X, which emphasizes safe driving and traffic management for slightly advanced cars, is rapidly progressing. To ensure the sum rate and compliance with reliability constraints, a unicast communication paradigm [ 26 ] was suggested. In a highway scenario, the authors chose signal power and buffer size as the key restrictions for V2X communication. The focus of the study was device-to-device (D2D) resource distribution. Their proposed resource allocation algorithm was applied to resource sharing between user equipment and cellular user equipment. When compared to other existing methods, simulation results revealed that the proposed algorithm had very promising performance.

Another study compared DSRC and LTE in infrastructure mode (LTE-I) and LTE-D2D in terms of average end-to-end delay and packet delivery ratio (PDR) under various communication conditions [ 27 ], which were achieved by varying the communication perimeter, message generation frequency, and road traffic intensity.

In the context of V2V, Karoui et al. [ 28 ] measured the signal-to-interference-plus-noise ratio and reference signal received power in two scenarios to benchmark ITS-G5 and LTE-V2X (mode 3), in terms of the end-to-end latency and radio frequency conditions. The study looked at the influence of substantial data traffic on an ITS alert service, and the influence of handover on ITS safety services in the context of V2V.

Along with MEC, blockchain can be used to protect and manage 5G V2X. Other research revealed the current condition of V2X, as well as its evolution based on cellular 5G and non-cellular 802.11bd [ 29 ], exploring how blockchain can be integrated with 5G-based MEC vehicular networks. With future research initiatives, the concerns and limitations in existing edge computing and 5G V2X are addressed.

2.3. Hybrid Approaches

Adithya et al. [ 30 ] created a method for vehicular communications using ZigBee wireless technology in order to prevent accidents. A GPS modem is utilized in this system to determine an accident’s exact location. Blynk applications on smartphones are used to save data to the cloud. When a car’s distance from a neighboring vehicle is lower than a certain threshold, the vehicle uses ZigBee to send messages to the adjacent vehicle, alerting the driver to take suitable action.

A collision avoidance system was developed [ 31 ] based on ZigBee where GPS technology is used to find a vehicle’s current location. In this system, the brakes of the vehicle are controlled using the automatic braking system (ABS) in the recent vehicles. The proposed model is used in Avs with a protective system already installed. The research they carried out was limited to line-of-sight visualization by a camera and IR sensors.

Many researchers and practitioners have adopted cameras, radar, and LiDAR in vehicles with a GPS for safe or autonomous driving. For the study in [ 32 ], the localization problem was solved by using radar data (pulse-based short-range radar) of the route in relation to a fresh traversal. The approach requires only the latest traversal history for the route, rather than storing and processing large amounts of data. The proposed approach was tested and validated to localize a truck on a road by using a five-minute dataset. Results after the experiments showed root mean square (RMS) errors of 7.3 cm laterally and 37.7 cm longitudinally, with the worst case being 27.8 cm and 115.1 cm, respectively [ 33 ]. Another problem with this approach is unexpected objects in dynamic surroundings.

In recent years, optical camera communication (OCC) [ 34 ] has been employed for communication between vehicles and RSUs. For a variety of activities, including traffic sign recognition and determining the distance between two vehicles, OCC employs a light-emitting diode (LED), an image sensor, and infrared (IR) rays. In a vehicle ad hoc network (VANET), the highest data rate reached to date with OCC is 55 Mbps. That paper included a survey of the literature on OCC, receiver and transmitter topologies, as well as some open research problems. Lee [ 35 ] investigated a highway accident detection system using CCTV and a Calogero–Moser system (an integrable system used at a crossroads). The effort attempted to overcome the difficulty in object recognition in shade and nighttime environments. In order to develop an accurate detection system, the flow of a vehicle trace was discovered to be similar to a Wigner distribution in terms of level spacing. Later on, Lee developed an advanced road traffic analytics processing system [ 36 ] that can process and analyze all the data in near real time. The suggested framework was evaluated using data collected from 41 CCTV cameras along the Icheon-to-Gangneung highway in Korea. The prediction method was implemented in a neighborhood between CCTV cameras.

Wei et al. [ 37 ] suggested an approach in which LiDAR and a camera are utilized to detect and prevent a vehicle from entering a restricted space. The beacons were normal orange traffic cones with a vertical luminous pole. They reduced false-positive LiDAR detection by projecting beacons in camera footage using DL, and by validating the detection using neural-network-learned projection from the camera into the LiDAR space. The utility of the suggested approach was proved via data acquired at Mississippi State University’s Center for CAVS.

Al-Mayouf et al. [ 38 ] proposed a system for accident management using cellular-technology-based VANETs. This technique ensures real-time communication between various vehicles, including crashed vehicles and ambulances, and roadside communicating units that send all updates to a server. An algorithm is used for optimal routing of an ambulance to its target, i.e., the location of the injury. In a vehicle, two types of sensors are installed: one is a bio-medical sensor that helps monitor the heart rate of the driver, and the other is a vehicular sensor to collect information about acceleration, temperature, tire pressure, etc. The microcontroller obtains inputs from these sensors and manipulates them to find any abnormalities. If abnormal values are found, it obtains the current location through a GPS module. After obtaining location information, it sends the information to nearby RSUs, which forward it to a central server to guide the ambulance to the target location. The drawback of this technique is that if the cellular signals are not present or weak, the required communication may fail, and the delay may be life-threatening.

Sarmin et al. [ 39 ] used GSM, a GPS, and an accelerometer-based system for detecting accidents. It provides alerts by sending an SMS to an emergency number when any accident is detected. The Arduino microcontroller is used for processing. The GPS and the accelerometer are attached to input pins, and GSM is used for the I/O port. The processor is responsible for all calculations performed locally, and then an alert is sent to an emergency number using the GSM cellular network. However, the proposed solution has no ability to inform other vehicles about the incident.

Fernandez et al. [ 40 ] developed a system to detect accidents using GPS information. In this technique, the receiver extracts the complete GPRMC sentence by demodulating RF signals from GPS satellites. The sentence holds all necessary information about the speed, time, and location of the vehicle. The last two values for each attribute are stored in memory and manipulated by the local processor for detection of any abnormalities. If the processor finds an accident, it raises a HIGH flag that results in autonomous initialization of emergency procedures. Before sending the message, the processor waits for manual cancellation by the driver if no help is required. After five seconds, an SMS alert is generated by the processor with the destination of the emergency.

Ali et al. [ 41 ] proposed a smartphone-based accident detection and notification system. The internal sensors of a smartphone are used to detect accidents, and the camera is used for recording images and video. If an accident happens, the video and/or images are sent to an emergency number. One major drawback of this approach is that every smartphone is prone to false positives.

Fernandes et al. [ 42 ] built an application for Android phones with an algorithm for autonomous detection of accidents. Mobile phones are connected to a single-board computer (SBC) through a USB serial cable. The accident detection algorithm takes values from the internal sensors of smartphones and vehicles. Whenever an accident occurs, the application detects it and broadcasts a message to all nearby vehicles. The nearby vehicles are informed by the color on the phone screen: green means the road is clear, and red means there is an accident nearby. Every smartphone, however, is prone to false positives.

Dias et al. proposed and developed a system that utilizes GSM, a GPS, the IoT, and GPRS technology with different vehicular sensors and microcontrollers [ 43 ]. An Arduino microcontroller receives input from vehicular sensors and sends a message via GSM modules using cellular technology if it obtains an abnormal value. Information about an incident’s location is taken from the GPS. The proposed system is expensive due to the multiple technologies and sensors.

Masini et al. [ 44 ] investigates a possible approach toward 5G VAN. The vehicular networks can lease available spectrum from existing cellular networks through base stations (BS) or access points using dynamic spectrum access (AP). They compared LTE performance from current legacy solutions to projected 5G trends. Shrestha et al. [ 45 ] investigated an evolution of V2X communication based on 802.11bd and 5G NR, and explain that multi-access edge computing (MEC) can bring cloud closer to vehicular nodes. They also introduce a combination of blockchain and 5G-based MEC vehicular networks as a possible solution for the security, privacy protection, and content caching issues in the existing 5G V2X environment.

The study [ 46 ] looks at how well vehicle visible light networks (VVLNs) perform in terms of increasing message delivery rates via full-duplex transmission (FD). According to their tests in which they compared their results with half-duplex (HD), it was discovered that using FD in urban environments increased message delivery ratio by 10%.

By putting sensors on the road, the VANET is integrated with the low-cost wireless sensor network (WSN) to buffer and convey information concerning unsafe situations [ 47 ]. The designed system’s prototype has been implemented and tested in the field. The findings show a variety of design tradeoffs and show that right parameters can provide sufficient safety and energy efficiency. In contrast to our work, this study started with the goal of improving driving safety by providing information about road conditions. They gathered and processed sensor data in order to extract information useful for safe driving and transmit it to vehicles that require it.

A novel biologically inspired networking architecture [ 48 ] was presented in order to take advantage of the first available DSRC-equipped vehicles as temporary RSUs. Cars acting as temporary RSUs can halt for short periods of time and operate as a communication bridge for other vehicles in the network. Designing local rules and the algorithms that apply them is the basis for the suggested approach. According to results, the proposed method significantly improves message reachability and connection. In order to reduce RSUs, they deployed vehicles as RSUs to improve message reachability and network connectivity.

A reinforcement learning (RL) algorithm [ 49 ] is utilized for partially observable intelligent transportation systems (ITS) based on dedicated short-range communications (DSRC). This system’s performance is evaluated for various car flows, detection rates, and road network topologies. This method can effectively reduce the average waiting time of vehicles at a junction even with a poor detection rate. However, their work focuses on reducing traffic congestion by controlling traffic lights, instead of reducing or responding to traffic accidents efficiently.

A deep reinforcement learning-based resource allocation system [ 50 ] was developed that includes remote radio head grouping and vehicle clustering. It balances system energy efficiency with service quality and dependability. In terms of performance, noise ratio, feasible data rate, and system energy efficiency, the new method is compared to three current algorithms using simulations. In comparison to our work, this study aims to assign various communication resources in order to improve system capacity while consuming the least amount of energy possible from periodic message traffic overheads.

2.4. Analysis of Previous Approaches

All the previous accident-related techniques are based on some sort of continuous monitoring in the vehicle of its surroundings through various sensors with the help of a microcontroller-based processing unit. Calibration of these devices from time to time is necessary for proper function, which becomes costly. Communication between vehicles is carried out by wireless technology. Although a GPS offers easy and accessible localization, the precision of the GPS still has room for further improvement in providing accuracy. To be more specific, a GPS suffers influences from several factors (e.g., receiver noise; a multipath effect), such that the received GPS coordinates have large errors in the actual coordinates of the vehicle, thereby posing a threat to the safety of Avs or the precision of ITS applications [ 21 ]. Another problem with GPS technology is that not all driving surfaces have satellite visibility [ 32 ]. The received GPS data can be influenced in urban areas by building occlusions, making the data less accurate [ 13 ]. On the other hand, post-accident techniques use a GPS to detect and find the location of the accident, with GSM and 5G technology for messaging to emergency service centers [ 51 , 52 ]. These techniques require internet connection. Sometimes, owners do not want to make changes in their vehicles, such as installing vehicular sensors. Another issue is the calibration of these sensors from time to time to avoid false readings. Sensor and communication-link checks for proper function are not always easy. Most of the previous techniques for accident detection apply to Evs only. Our work is also different from other studies [ 44 , 45 , 46 , 47 , 48 , 49 , 50 ] where the major focus was to provide information about road conditions to drivers, reduce traffic congestion by controlling traffic lights, and manage communication resources to improve system capacity. A summary on the analysis of previous approaches is provided in Appendix A .

3. Proposed Accident Alert Light and Sound System

In BWC, MVCs can happen where a number of approaching vehicles can lead to another accident. In this case, damage and the number of injured people, and/or fatalities might increase. The common cause for this type of accident is poor visibility whereby drivers cannot see the accident until they come upon it (approximately 10 to 15 m away). At that distance, braking will not stop the vehicles in time and, as a result, they become part of the accident. In modern vehicles, some preventive and protective systems are installed for the safety of the driver and passengers. However, in Third World countries, for example, more than 80% of vehicles have no factory-installed protective systems [ 53 ]. Until now, no system can handle such situations for all vehicles in an old or late model. Based on the literature review, almost all the work carried out so far is either used for modern vehicles or to make older vehicles into Evs. Evs have some system installed in either a mobile phone application or a microcontroller-based system. All vehicles that have any type of preventive measure and alert system are considered equipped vehicles. Evs act as a source generating accident alert messages. However, if a vehicle is damaged badly or burned in an accident, the system may fail, and no information may be sent to an EOC. In a mobile-phone-based system, the mobile phone may be damaged in an accident. It is not easy to convert all old vehicles into Evs due to financial costs. Another limitation is the need for timely calibration of the sensors in smartphone-based and vehicle-based systems, which is difficult to do. Mobile phones are often prone to false positives, and GSM signals may drop in some areas. If a GPS is used to find locations, major drawbacks are inaccurate and depend on satellite signals. It is difficult to use pre-saved locations while vehicles are moving. It is also notable that no research work has been carried out for accident detection by nEVs.

There is a need for a system that works from outside the vehicle for detection of an accident and generating alerts for approaching vehicles in order to avoid MVCs. A new concept of making smart roads is introduced here. SRs have different types of sensors and actuators installed for autonomous detection of accidents. SRs have nodes (a completely independent system) that are nearly 50 m apart. Because these nodes are fixed on the roadside, it is possible to use their pre-saved locations. A node that detects an accident sends its pre-saved location to an EOC for the quickest rescue operation, saving lives and reducing damage. The main part of the SR is an alert system, called an AALS, which makes drivers of approaching vehicles aware of an accident.

The AALS is a system designed for nEVs and EVs to avoid MVCs. Most of the previous researchers have tried to build a system in which vehicles are able to sense and communicate with each other through devices. In short, they try to make all vehicles equipped. However, equipping all vehicles with the same type of prevention and communication mechanism is difficult to achieve. Modern vehicles may have a factory-built protection system installed. In the system proposed here, a golden yellow blinking light is used with a siren to alert vehicles to an accident. According to one study of different colored lights, red and golden yellow light are visible from a long distance and in bad weather [ 54 ]. Red lights are already used for traffic signals; thus, using golden yellow light is a good option. Another parallel way of alerting drivers is with a loud siren, so drivers can hear it and take precautionary measures to avoid being involved in an MVC, especially in BWC. These alerts are generated through an auto mechanism. The proposed system will be installed on both sides of the road. A complete system based on applied studies was built with hardware (including microcontroller boards), different sensing units, and with alert functions. A node can be powered from existing light poles on the roads or highways; otherwise, solar panels can be deployed.

The basic functions of the AALS system are accident detection from outside vehicles (i.e., from the roadside) and alerting oncoming vehicles about an accident by blinking a light and making a sound. The AALS system communicates with the next node through EDWSN, a wireless communication protocol that uses two HC-12 modules with low power consumption. It functions only when an accident is detected by the AALS or when the reset button is pressed. When an accident happens, certain events occur. For example, when brakes are applied at high speed, a squealing sound is produced; when a vehicle hits another vehicle, it produces a loud sound that can be heard from a distance. If glass breaks, it produces an audible sound; if a vehicle is burning, the temperature of the environment rises, and smoke is produced; when a vehicle suddenly stops in the middle of the road, it becomes an obstacle for other vehicles. By considering all these factors, a foolproof system is available for accident detection. The steps of the proposed algorithm for a node in the system are shown in Figure 1 .

Proposed algorithm for an AALS node.

A Way to Smart Roads (SRs)

SRs are roads that have some sort of sensing power given with the help of different types of transducers (devices that convert one form of energy to another) as well as control devices with communication capabilities. Several nodes are installed on the sides of the road to enable sensing at those points. By increasing the number of nodes, sensing power can improve. These nodes hold all the necessary sensing devices and a microcontroller board with a wireless communications system. The distance between nodes is directly proportional to the transmission power of the sensing devices. SRs do not discriminate among vehicle types, and can detect accidents involving Evs as well as nEVs. Similarly, SRs generate light and sound alerts that can be seen and heard by drivers and passengers of approaching vehicles.

A smart road has sensing capabilities for accident and/or event detection. A two-way road is shown in Figure 2 , each side with AALS nodes which are installed at a distance of 50 m apart. Labels A, B, C, etc., denote the sensing nodes, and each node communicates with neighboring nodes. If a sensor of node D detects an accident, it sends a message to its immediate downstream node, i.e., node C, which will send the received message to node B, and this procedure is repeated until the message finally reaches the EOC. Similarly, on the other side of the road, if an accident is detected by node Y, it sends a message to node X, which relays the message to its previous immediate node, i.e., node W. This procedure is repeated until the message reaches the EOC. A message cannot be received by any node other than its immediately adjacent node. For example, if node D senses an accident, the message is sent to C, B, A, and the EOC. Different channels are used to perform this type of communication. Each channel is in the 400 kHz bandwidth range. To avoid overlapping signals, the immediately adjacent node should be a part of five channels. The transmitting channel and baud rate of one node are similar to the receiving channel and baud rate of its previous adjacent node. For example, if channel 20 is selected for the transmission of a message for node C, then the same channel (i.e., 20) should be selected for node B and other previous adjacent nodes.

Illustration of a smart road.

A functional block diagram of a node in the proposed AALS system is shown in Figure 3 .

Block diagram of a node in the proposed AALS system.

Approximately 4000 nodes are required to cover a 100-km two-lane road. Each node has the ability to sense and recognize an accident from loud sounds or obstacle detection by using its different sensors. The Arduino Uno microcontroller takes input from the IR, microphone, smoke sensors, and the HC-12 receiver. A smoke sensor is used to detect fire. When a sound is generated due to an accident between vehicles or by a vehicle hitting an object, the microphone picks up the sound and sends it to the microcontroller, which compares the level with a preset threshold. It declares an accident if the level is higher than the threshold. However, if the sound is not greater than the threshold, the microcontroller discards it. In the meantime, the IR sensor works in parallel to detect obstacles on the road. The microcontroller compares the IR value to check whether it is lower than the set threshold and if the elapsed time is more than five seconds (the time taken by a long vehicle to pass in front of the IR sensor at 20 km/h). When an obstacle remains for more than five seconds, the microcontroller recognizes it as an accident. Meanwhile, if the smoke sensor detects any smoke in the atmosphere, the value of the fire bit changes to HIGH. The controller executes three actions if an accident is detected. First, it blinks the light continuously until the system is reset; secondly, it sounds the siren to alert drivers of approaching vehicles so they can take measures to avoid an MVC; and third, it sends a message through the HC-12 transmitter to the immediately adjacent node. A message contains three types of information: the pre-saved location of the detecting node, the direction of traffic on the road, and fire information (If recorded). Communication between nodes is one way, i.e., back toward oncoming vehicles to make their drivers aware of the incident ahead. Every node does the same job in case of accident detection until the message reaches the EOC. When the EOC receives an accident message, it sends an ambulance to the incident location and the fire brigade in case of fire. A rescue operation will be performed to save the lives by giving treatment and hospitalizing the injured peoples if necessary. After the road is cleared by the rescue team, the first communication node will be reset via the hardware push button. Four functions activate when the reset button is pressed: first is to switch off the blinking light; second is to stop the siren; third is to send the RESET message to its adjacent node; and fourth is to reset the controller to its initial state. The RESET message is also sent backwards and will be relayed until received by the EOC. This message indicates to the EOC that the road is now clear, and the rescue operation completed successfully.

4. Hardware and Experimental Setup

Developing and testing the entire system on hardware is an interesting part of the proposed work. The experimental setup, microcontroller boards, and sensors are selected, considering the cost and availability. We selected an Arduino Uno R3 board and a PIC 4550 microcontroller [ 55 , 56 , 57 ] with different sensors and actuators. The Arduino website [ 57 ] is an open-source helping website that provides a basic code for different sensor modules that work with Arduino boards. The proposed AALS system is built to minimize computing power needed to detect accidents, and does not need artificial intelligence or a machine learning technique. With the AALS system, accidents are detected by monitoring input from different sensors and comparing them to threshold values. After the successful detection of an accident, the alerts are generated to protect oncoming vehicles from MVCs. The help request message is generated and sent wirelessly to the immediately downstream node. All nodes are interconnected wirelessly in such a way that each node listens to the next node and sends a message downstream to the node. Messages are sent only in case of an accident or after a RESET event, where the RESET button was pressed. To build the whole system, the following sensor modules and components are used:

• 1. Arduino Uno R3;
• 2. An IR sensor module
• 3. A microphone sensor module;
• 4. A smoke detection module;
• 5. A GPS 8M module;
• 6. A HC-12 wireless communication module;
• 7. Breadboards and jumper wires;
• 8. A relay module;
• 9. A golden yellow light and siren;
• 10. 16X2 LCD with an I2C module.

4.1. Arduino Uno R3

An Arduino Uno R3 board is microcontroller-based open-source prototype hardware [ 57 , 58 ] that can be programmed through software called the Arduino IDE. The programming language for the Arduino IDE is C++. Arduino boards have several digital and analog pins, where the digital pin is configurable for input or output by using the “pinMODE” command. Digital pins have only two states: HIGH (1) and LOW (0). However, analog pins have a value that varies from 0 V to 5 V. Most of the sensors, e.g., temperature, sound, and smoke, have analog output, although some provide digital output based on some threshold value set by an installed potentiometer. There are various kinds of Arduino boards available: the Arduino Uno R3, Mega 2560, Due, Ethernet, LilyPad Arduino 328, and their variants. In this research, the Arduino Uno R3 is selected due to its functionality and cost. It works on a 5V DC supply from a USB port. When not connected to a USB port, it can operate on 7 V to 12 V DC from an external adapter. The pins of this board are shown in Figure 4 .

Arduino Uno R3 pin descriptions [ 55 ].

4.2. IR Sensor Module

This module consists of two parts: the IR transmitter and the IR receiver. The IR transmitter is based on an infrared LED that emits light at an infrared frequency, while the IR receiver is based on a photodiode that can sense reflected infrared light and convert it into an electrical signal [ 59 ]. Figure 5 a shows the IR sensor module with its three pins. Two pins are used for powering up the module (+5 V and GND), while the third pin is used for output. The output is HIGH when light is not reflected back and vice versa. The range of sensitivity can be adjusted through a variable. This module is normally used for obstacle detection; however, it can be used for counting objects. such as people, vehicles, etc.

( a ) IR sensor module [ 55 ]; ( b ) microphone sensor module [ 60 ]; ( c ) smoke sensor module [ 61 ].

4.3. Microphone Sensor Module

A microphone is a device that converts sound energy waves into electrical signals. There are two main types of microphone: dynamic and condenser microphones. The microphone sensor module for this system is the condenser type. There is a built-in transistor to amplify the electrical signal generated by the sound wave energy. The strength of the electrical signal is proportional to the loudness of the sound waves, and the frequency is equal to the frequency of the sound waves. There are four pins in this module; two are used for powering up, i.e., +5 V and 0 V, and the other two are output pins. One pin is for analog output, and the other is for digital output. The digital output pin goes to HIGH when the loudness of a sound is greater than the threshold set with a variable; otherwise, it remains at LOW. The voltage of the analog pin varies from 0 to 5 V, depending upon loudness (the energy carried by the sound waves). The analog voltage is measured in numbers from 0 to 1023. Therefore, 0–5 V is divided into 1024 equal-length intervals. The microphone module is shown in Figure 5 b.

4.4. Smoke Detection Module

The smoke detector is used to detect smoke or fire by sensing combustible gases present in the atmosphere that ionize into a substance present between the two conducting plates. When ionization occurs, varying voltages are produced, indicating harmful gases in the air. The smoke sensor is pictured in Figure 5 c. There are four pins in this module: two are used for powering up and the other two are output pins (analog and digital).

4.5. GPS 8M Module

A GPS receiver receives radio frequency (RF) signals at a very high frequency, as shown in Figure 6 a. A GPS receives static data broadcast from different geo positioning satellites. These data contain information about the location, time, altitude, etc., in the form of National Marine Electronics Association (NMEA) sentences. These sentences are difficult to understand, and are therefore parsed by the website ( www.freenmea.com ) (accessed on 21 April 2021). To obtain the latitude and longitude, the parse button is pressed after copying a sentence that starts with $GNGLL to the website. These coordinates are transferred to the Google website to determine the physical location.

( a ) GPS 8M module [ 62 ]; ( b ) HC-12 module [ 63 , 64 , 65 ]; ( c ) breadboard and jumper wires [ 56 ].

4.6. HC-12 Wireless Communication Module

The HC-12 is an RF transmitter and receiver used for wireless communications. These modules replace wired communication with a half-duplex wireless serial port. Half duplex means that one module transmits data while another receives it. The HC-12 module is used because of its long range up to 1 km and because its operating frequency ranges from 433 to 437 MHz. Figure 6 b shows the pair of HC-12 modules in the proposed system. The frequency range is divided into 100 channels with each having a 400 kHz bandwidth. By selecting different channels, they can communicate with different devices. Modules with the same baud rate and on the same channel can communicate with each other within their communication ranges [ 62 , 63 , 64 , 65 ].

4.7. Breadboard and Jumper Wires

A breadboard is a device for temporarily making electronic circuits without a permanent connection from soldering. Jumper wires are used to connect different components on the breadboard to control devices, such as Arduino boards. The breadboard consists of five interconnected holes vertically separated from other vertically interconnected holes. The horizontally interconnected holes for the + supply connection are separated into other horizontally interconnected “–” supply connection holes line. A power connection is present on both horizontal sides of a breadboard. There are two types of jumper wire; one is male-to-male and the other is male-to-female. Male-to-male jumper wires connect the breadboard and the Arduino Uno board, whereas male-to-female jumper wires connect different sensors directly to the Arduino Uno board without a breadboard. A breadboard and male-to-male jumper wires are shown in Figure 6 c.

4.8. Relay Module

A relay module is used with the output of a controlling device that requires more power to operate than its normal output. An Arduino Uno board has a maximum 40 mA output current, which can light up an LED, but cannot control the motor, light bulb, or siren, which operate on high voltage and current rates. Another use of a relay module is to isolate and protect the controller from the output load. A relay module may have one, two, four, eight, and twelve relays for controlling one, two, four, eight, and twelve devices, respectively. Here, two relay modules are used to control two output devices (i.e., the light and siren). Figure 7 a shows the two-channel relay modules for the Arduino Uno board.

( a ) Two-channel relay module with pinouts [ 66 ]; ( b ) golden yellow light bulb and siren.

4.9. Golden Yellow Light and Siren

It is difficult to see white light in the daytime or when fog or smoke is in the air. Using a white light does not help people see things in bad weather (e.g., heavy rainfall, storms, or fog). Red lights and golden yellow lights can be seen in such situations from very far away [ 47 ]. Red lights are already used in traffic signals. However, so far, golden yellow light is not used in traffic management for any type of alert. Therefore, it is a good choice for an alert signal when it blinks. The bulb that produces golden yellow light to accompany the siren is shown in Figure 7 b. The siren is used for alerts by generating a specific loud sound that can be heard from a distance. It is used where a generic alert to the public is required. In the same way, the siren is used in this proposed system to alert drivers of approaching vehicles about an incident ahead.

The last module is the LCD 16×2 which is used to display information at a node during the installation.

4.10. Wiring Diagram of an AALS System Node

The pinout connections shown in Figure 8 are used in this research. The HC-12 module has four pins: the VCC pin is connected to +5 V, the GND pin grounds the module, the RX pin connects to pin 5, and the TX pin connects to pin 2 of the Arduino Uno board through a breadboard. LCD 16X2 with the I2C module also has four pins where the VCC pin is connected to +5 V, the GND pin grounds the module, the SDA pin connects to analog pin A4, and the SCL pin connects to analog pin A5 through the breadboard. The IR obstacle module has three pins: the VCC pin connects to +5 V, the GND pin grounds the module, and last output pin is connected to analog input pin A0 through the breadboard. The microphone sensor module has four pins, with the VCC pin connected to +5 V, the GND pin grounding the sensor, digital output not connected to any pin, and the analog output pin connected to A1 of the Arduino Uno board through the breadboard. The smoke sensor module has four pins: the VCC pin is connected to +5 V, the GND pin grounds the module, digital output is not used, and the analog output pin connects to analog input A2 of the Arduino Uno board through the breadboard. The siren and golden yellow light are connected to the Arduino Uno board through a relay module that isolates the output of the Arduino from the high-powered output devices. The two-channel relay module has four input pins: the first is VCC (+5 V) and the second is GND (0 V), with the third and fourth pins being connected to pin 12 for the siren and digital pin 13 for the light, respectively. The relay module controls output devices that operate at 5 V; however, it can control ON or OFF voltages from 5 to 30 V from a DC supply and form 110 to 220 V from an AC supply. It connects a 12 V battery to the siren and light.

Wiring diagram for the AALS system node.

In the AALS system, two HC-12 modules are used: one for transmitting messages and the other for receiving messages. The pin connections are such that for the HC-12 transmitter, the Tx pin is not connected, while the Rx pin is not connected for the HC-12 receiver module. This setting is made for one-way communication. If we use only one HC-12 module, then it sends and receives messages on the same channel; hence, the repeated message of the other node is received back in a loop. If a message is received in a loop, it results in a microcontroller being busy, and communication takes place all the time, which is not acceptable. Two HC-12 modules are used to minimize the risk of loopback communications and to ensure only event-driven communications.

The number of sensors deployed on a node is determined by the coverage area and distance. Because the IR sensor has a range of five meters, a node will require 10 IR obstacle sensors. If we replace the IR obstacle sensor with a TF mini IR time-of-flight distance sensor, the number of sensors required drops to four. The smoke sensor can detect smoke at a radius of 7.5 to 10 m (a diameter of 15 to 20 m); hence, three to four smoke sensors are mounted on a node [ 67 ]. Microphone sensors have an eight-meter coverage range; hence, six to seven microphone sensors are installed on each node. However, some directional microphones perform at up to 40 m; thus, two directional microphones are used in the proposed system. For a 50-m distance, one siren and one golden yellow light are required.

5. Experiments

Different experiments are performed to find the threshold values for the different sensor modules which will be used for the detection of accident. These values have great importance in the successful detection of false-proof accident detection systems.

5.1. Finding Threshold for the Microphone Sensor Module

The microphone sensor module is attached to the Arduino Uno board, as shown in the wiring diagram from uploading a sketch (a program) to the Arduino board through the Arduino IDE. The sketch contains the basic pin configuration of the Arduino board and the serial print command for showing the incoming values on the serial monitor or serial plotter. The serial monitor is an output screen that shows data on the monitor and the serial plotter shows the output in the form of a graph which can be easily understood. Figure 9 shows the microphone output of the serial plotter.

Microphone output on the serial plotter.

When no sound is produced externally, the microphone senses only background noise. In Figure 9 , background noise is shown before and after the accident’s sound. When a sound is played through a loudspeaker, the amplitude increases. The normal background sound ranges from 518 to 532 ADC (calculated by taking the ratio of the analog signal’s voltage to the reference voltage: ADC = (11.003 * dB) − 83.2073 [ 68 ]). However, when an accident sound is produced, the range suddenly changes from 480 to 555 ADC. If the range exceeds 532 ADC or is less than 518 ADC, there is a chance of an accident. To make it more precise, another experiment was completed by playing the sound of a passing vehicle into a microphone through a loudspeaker. Figure 10 shows the waveform for the sound of the passing vehicle on the serial plotter. The range of the output waveform varies from 512 to 534 ADC. Figure 10 shows that it varies from 519 to 531 ADC against the background sound. It was noted that there are very small differences in the amplitude of the waveform produced when only background noise is presented or from sound due to vehicles passing by, as shown in Figure 11 . In one case, the amplitude of the waveform reached 540 ADC when an accident sound is produced. Various accident sounds are produced, and a similar effect provided loudness that is nearly the same. An accident can be detected from the loudness of sound.

A passing vehicle’s waveform.

Background noise waveform.

5.2. Finding a Threshold for the IR Sensor

The IR sensor module is simpler than the microphone sensor because it produces analog output nearly equal to 5 V; in other words, on the Arduino serial monitor, the value is greater than 1000 when there is no obstacle present. However, the serial monitor showed a reading of less than 20 when an obstacle came in front of it. An obstacle in the module’s light passageway is easy to detect. However, the problem is that it detects every obstacle that passes in front of it, but we have to detect only vehicles that stop. To handle this situation, software is used so that, if an obstacle remains in place, the value generated is greater than the value generated when a vehicle simply passes in front of the IR sensor. Almost all long vehicles pass a point on a free road in less than five seconds. So, we set a time limit of 5000 ms, which can be changed according to requirements. If an object is detected by the IR sensor for more than five seconds (the threshold), it is deemed an accident.

5.3. Finding the Smoke Sensor Threshold

The smoke sensor has a normal range of about 150 to 220 when there is no smoke, as seen in Figure 12 .

Smoke detector’s waveform when there is no smoke.

Fire or smoke are found in the 230 to 270 sensor range, as seen in Figure 13 . After performing several experiments, it is noticed that a value greater than 240 (the threshold) indicates fire.

Smoke detector’s waveform when there is smoke.

5.4. Experiments for Finding Locations

The GPS 8M module [ 69 ] is used for recording location information and saving it permanently. Because it takes longer to establish the first fix, once the location is determined that information is saved to the Arduino Uno board [ 70 ]. The temporary connection of the GPS 8M module with the Arduino Uno board is shown in Figure 14 .

Wiring diagram of GPS 8M module connected to Arduino Uno board [ 69 ].

A sketch is uploaded to the Arduino Uno board (as shown in Figure 15 ) to obtain the location information on the serial monitor.

Sketch Uploaded to Arduino Uno Board.

When the GPS module is connected to the board, it sends the received data to a serial port. On the serial monitor, the data are viewed in the form of NMEA sentences. When it starts, the data received are not complete, as shown in the text box.

It takes (10 to 15 mins) to the module after a cold restart to obtain and lock in the data. The data are shown in another text box.

$GNVTG,,,,,,,,,N*2E

$GNGGA,,,,,,0,00,99.99,,,,,,*56

$GNGSA,A,1,,,,,,,,,,,,,99.99,99.99,99.99*2E

$GPGSV,1,1,00*79

$GLGSV,1,1,00*65

$GNGLL,,,,,,V,N*7A

$GNRMC,,V,,,,,,,,,,N*4D

From the numbers in the information shown on the serial monitor (below), we require only the highlighted information. It is copied and pasted to the freenmea.com website and parsed for the location. To obtain the exact latitude and longitude, the location is searched for by using Google Maps. The procedure or function is shown below.

$GNVTG,,T,,M,0.327,N,0.606,K,A*3B

$GNGGA,070234.00,3336.52621,N,07306.48790,E,1,09,1.02,514.3,M,−40.6,M,,*61

$GNGSA,A,3,24,15,17,28,19,02,,,,,,,1.72,1.02,1.39*14

$GNGSA,A,3,79,73,81,,,,,,,,,,1.72,1.02,1.39*13

$GPGSV,3,1,12,01,18,042,18,02,09,198,23,03,03,084,28,06,36,175,20*71

$GPGSV,3,2,12,15,12,271,25,17,68,012,25,19,71,278,27,21,00,034,*71

$GPGSV,3,3,12,22,01,058,,24,22,315,38,28,59,048,28,30,35,156,*7B

$GLGSV,2,1,07,73,14,322,34,78,01,131,,79,47,111,25,80,60,352,19*60

$GLGSV,2,2,07,81,28,030,26,82,81,025,,83,46,208,*51

$GNGLL,3336.52621,N,07306.48790,E,070234.00,A,A*72

$GNRMC,070235.00,A,3336.52625,N,07306.48784,E,0.382,,280421,,,A*6F

We stored the latitude, longitude, and address permanently (in code) to the Arduino Uno as a pre-saved location. After obtaining and saving the information, the GPS module was uninstalled and was no longer part of the system. The pre-saved location is used when an accident occurs. After the successful detection of an accident, the message on the location is sent to the immediately adjacent node.

The original NMEA sentence is:

The following output can be obtained from the execution of the above line:

33°36′31.33″N, 73°6′29.57″E

The following latitude and longitude values are shown when above coordinates are searched on the Google Maps:

33.608703, 73.108214

Finally, the following physical address is shown against the above latitude and longitude values:

“Street Number 7, Dhoke Raja Muhammad Khan, Rawalpindi, Islamabad Capital Territory 46000, Pakistan”

For the complete code for one node of the AALS, see Appendix B .

5.5. Validation of AALS

Each module is tested for proper working order and was found to operate as expected. All modules of the AALS system were integrated to build a complete working system. It works as a single unit where at least two nodes must be present at a distance of 50 m apart. Therefore, two nodes were built in a lab for testing and validation. By default, each node was in the listening position with all its sensing capabilities. Each node is able to detect accidents through the mic, IR sensor, and smoke sensor modules. Whenever an accident occurs in the area of a node, only that node responds and sends messages to another node. In this experimental setup, if a node has only one HC-12 communication module (i.e., to both receive and send a message), then it might suffer from the message loopback problem due to the echo of the same message. To overcome this problem, two HC-12 modules were used, tuned to different channels: one module for transforming and another for receiving. Each pair of nodes communicated with each other in an unidirectional manner. The transmitting module channel was the same as the receiving channel of the other module. The final AALS system node has two HC-12 wireless communication modules: one for receiving and the other for transmitting. After the experiments were completed, we made the following observations.

When there are 20 nodes along a 1 km stretch of road:

Minimum transmission time = 80 ms = 0.08 s;

Maximum transmission time = 1600 ms = 1.6 s.

When there are 2000 nodes on a 100 km stretch of road:

Minimum transmission time = 8000 ms = 8 s;

Maximum transmission time = 160,000 ms = 160 s.

The HC-12′s signal transmission time from one node to another node ranged between 4 ms and 80 ms.

• 1. When there was an accident, the time to detect it with sound was 100 ms.
• 2. The time to detect an accident using IR was 100 ms + Threshold (e.g., 100 + 5000 = 5100 ms).
• 3. The time required to detect smoke was 30 to 60 s.

The salient features of the AALS system node are as follows.

• 1. It can sense the sound produced by an accident/crash.
• 2. It can sense smoke from a fire.
• 3. It can sense an obstacle on the road for a period longer than the set threshold, e.g., 10 s.
• 4. When an accident was detected by sound or obstacle detection, the alert comprising light and sound was generated on the node.
• 5. A message with the location, traffic direction, and fire detection information was sent to the immediately adjacent node.
• 6. When a message about an accident was received by a node, it retransmitted the message to its adjacent node. This process then continued from node to node until the message was received by the EOC.
• 7. Each node that entered the retransmission phase also generated an alert using light and sound to warn oncoming vehicles.
• 8. All oncoming vehicles’ drivers and passengers were able to see the blinking lights and can hear the siren. An oncoming vehicle’s driver ensured their safety by taking precautionary measures.
• 9. When a rescue team reached the accident location and completed their tasks, the node was reset. All nodes were reset when they received the RESET message from the node that started the communication.

6. Conclusions and Future Work

The proposed AALS system provides a solution to the problems of actual location detection of an accident in order to protect other drivers and the authorities. It provides an alert, warning drivers of nEVs and EVs about an accident ahead. It helps the ambulance and fire brigade reach the destination easily by providing the exact location of the incident. Each node is a small but complete system that can sense its surroundings, and, in the case of an accident, generates alerts comprising a blinking light and a siren. The alerts will help drivers of oncoming vehicles to protect themselves from a mishap. Each node is responsible for communicating in one direction opposite the flow of traffic. By making smart roads using this AALS system, drivers are informed in time about any dangerous situation on the road. Moreover, in the case of an accident, the rescue team can reach the site without delay through an automatic process. Once the system is established, it works for a long time with limited maintenance. For better results, the EOCs should be at equidistant locations about 40–50 km apart at maximum. Each EOC should be equipped with an ambulance and fire brigade and should be responsible for sending a rescue team to the incident location. The EOC should also be aware of any ambulance and/or fire brigade already moving that might be near the location.

Current work is limited as the proposed system was installed and tested in a lab by simulating various accident scenarios. All results were found to be accurate after experiments. However, differences between the lab environment and the actual road environment may cause test results to differ from those expected in the operational environment. We will investigate the detection accuracy of the proposed system in a more practical situation on the road in our future work. An enhanced version of the AALS can be developed that will detect an accident from a longer distance (more than 100 km). Automatic fire extinguishers can be used to extinguish fires without the need to send a fire brigade. Other sensors can be used to provide information about the road conditions to drivers of nEVs and EVs through different display mechanisms. The same procedure can be applied to an individual lane to show its status (e.g., traffic density, single-lane incidents).

Appendix A. Analysis of Previous Approaches

Table A1 provides a summary of the previous research with their adopted techniques and limitations.

State-of-the-art accident avoidance approaches.

Appendix B. Code of the AALS System for One Node

//This code is built for “Autonomous Accident Detection & Alert for Non-equipped Vehicles”

// This code is the property of Muhammad Zahid Hanif, MSCS FUUAST ISB.

// Library added for code simplification

#include < SoftwareSerial .h> // For serial communication though digital pins

#include <LCD_I2C.h>

LCD_I2C lcd(0x27); // Default address

SoftwareSerial HC12(2, 5); // HC-12 TX Pin, HC-12 RX Pin

// Variable Declarations

// Pin mapping

const int lightPin = 13; // used for alert by blinking light

const int sirenPin = 12; // used for alert by siren sound

const int resetPin = 11; // used for reset with pushbutton

const int softresetPin = 10; // software-based reset

const int irSensor = A0; // used for IR sensor input

const int micSensor = A1; // used for mic sensor input

const int smokeSensor = A2; // used for smoke sensor input

//PRE- Location Save

const String GPSLatitude = “32.783280”;

const String GPSLongitude = “72.703564”;

const String mylocation = “Lat:32.783280, Long:72.703564; Kallar Kahar, Chakwal, Punjab, Pakistan, N2S”;

// Sensor threshold settings

const int irThreshold = 100;// less than this threshold detects obstacle

const int micThreshold = 540; // greater than this threshold detects accident

const int smokeThreshold = 240; // greater than this value indicates fire

const int timeElapsedThreshold = 5000; // 5 s is for a long vehicle passing by

// Variables for storage input data

int irSensorValue = 0;

int micSensorValue = 0;

int smokeSensorValue = 0;

int resetPinValue = 0;

int timeElapse = 0;

int accidentSound = 0;

int fireDetected = 0;

void setup () {

// run once:

Serial .begin(9600); // serial baud rate is set to 9600

HC12.begin(9600); // Serial port to HC12

// gps.begin(9600); // Serial port to GPS

lcd.begin(); // If using more I2C devices using the wire library use lcd.begin(false)

lcd.backlight();

pinMode (lightPin, OUTPUT); // Declare lightPin as output pin

pinMode (sirenPin, OUTPUT); // Declare sirenPin as output pin

pinMode (resetPin, INPUT); // Declare sirenPin as input pin

pinMode (irSensor, INPUT); // Declare irSensor as input pin

pinMode (micSensor, INPUT); // Declare micSensor as input pin

pinMode (smokeSensor, INPUT);// Declare smokeSensor as input pin

pinMode (softresetPin, OUTPUT); // Connect this pin to reset of Arduino Uno board

digitalWrite (softresetPin, HIGH); // Make it high because reset is low

void loop() {

// runs repeatedly:

// get values from input pins

irSensorValue = analogRead (irSensor);

micSensorValue = analogRead (micSensor);

smokeSensorValue = analogRead (smokeSensor);

resetPinValue = digitalRead (resetPin);

// Check if hardreset button is pushed

if (resetPinValue == 1) {

HC12.write(“Reset”);

Serial .println(“HardReset”);

digitalWrite (softresetPin, LOW);

// serial print the values of sensors

Serial .print(“IR Sensor has value = “);

Serial .println(irSensorValue);

Serial .print(“mic Sensor has value = “);

Serial .println(micSensorValue);

Serial .print(“smoke Sensor has value = “);

Serial .println(smokeSensorValue);

// comparision of sensor values

if (irSensorValue < irThreshold) {

timeElapse += 100;

Serial .println(timeElapse);

if ((timeElapse <= timeElapsedThreshold) && (irSensorValue > irThreshold)) {

timeElapse = 0;

if (micSensorValue > micThreshold) {

accidentSound = 1;

Serial .println(“Accident sound detected”);

if (smokeSensorValue > smokeThreshold) {

fireDetected = 1;

Serial .println(“Fire detected”);

// Send message if accident detected

if ((accidentSound == 1) || (timeElapse >= timeElapsedThreshhold)) {

HC12.write(“Lat:32.783280,Long:72.703564;Kallar Kahar,Chakwal, Punjab, Pakistan, N2S”);

Serial .println(“Accident Detected”);

HC12.write(“Accident Detected”);

tone(sirenPin, 1000, 70);

digitalWrite(lightPin, HIGH);

if (fireDetected == 1) {

HC12.write(“Help: Fire brigade “);

Serial .println(“Help: Fire brigade “);

// lcd display

lcd.print (“Lati: “);

lcd.print (GPSLatitude); // print smoke detector value

lcd.setCursor(0, 1); // curser position

lcd.print(“long: “);

lcd.print(GPSLongitude);// print IR sensor value

// if msg received by HC12 module

// HC12 Serial communication using pin 2 and pin 5

String MsgReceive = ““;

byte incomingByte;

while (HC12.available() > 0) {

incomingByte = HC12.read();

//HC12.write(incomingByte);

MsgReceive += char(incomingByte); // incoming msg is stored in MsgReceive variable

// changing

timeElapse = 5500;

// Serial.println(“zahid while”); //////

// Serial.write(HC12.read()); // Send data to serial monitor //////

Serial .println(MsgReceive);

if (MsgReceive == ““) {

//Serial.println(“No Message Receive”); ;

else if (MsgReceive == “Reset”)

digitalWrite(lightPin, LOW);

digitalWrite(sirenPin, LOW);

irSensorValue = 0;

micSensorValue = 0;

smokeSensorValue = 0;

resetPinValue = 0;

accidentSound = 0;

fireDetected = 0;

else if (MsgReceive != ““)

{ tone(sirenPin, 1000, 70);

delay(100); // Wait for 50 milliseconds

lcd.clear();

//Serial.println(timeElapse);

Author Contributions

Conceptualization, A.M.; methodology, M.Z.H.; software, M.Z.H. and N.K.; validation, A.M., S.L. and S.Y.N.; investigation, S.L.; resources, A.M.; writing—original draft preparation, A.M.; writing—review and editing, S.Y.N.; visualization, N.K.; supervision, A.M.; project administration, S.Y.N.; funding acquisition, S.Y.N. All authors have read and agreed to the published version of the manuscript.

This research was supported in part by the National Research Foundation of Korea (NRF), with a grant funded by the Korean government (MSIT) (2020R1A2C1010366). This research was supported in part by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (2021R1A6A1A03039493).

Institutional Review Board Statement

Informed consent statement, data availability statement, conflicts of interest.

The authors declare no conflict of interest.

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Mobile phone data in transportation research: methods for benchmarking against other data sources

• Published: 02 January 2021
• Volume 48 , pages 2883–2905, ( 2021 )

Cite this article

• Andreas Dypvik Landmark 1 ,
• Petter Arnesen 2 ,
• Carl-Johan Södersten   ORCID: orcid.org/0000-0003-1586-8462 2 &
• Odd André Hjelkrem 2  

1407 Accesses

10 Citations

1 Altmetric

Explore all metrics

The ubiquity of personal cellular phones in society has led to a surging interest in using Big Data generated by mobile phones in transport research. Studies have suggested that the vast amount of data could be used to estimate origin–destination (OD) matrices, thereby potentially replacing traditional data sources such as travel surveys. However, constructing OD matrices from mobile phone data (MPD) entails multiple challenges, and the lack of ground truth hampers the evaluation and validation of the estimated matrices. Furthermore, national laws may prohibit the distribution of MPD for research purposes, compelling researchers to work with pre-compiled OD matrices with no insight into the methods used. In this paper, we analyse a set of such pre-compiled OD matrices from the greater Oslo area and perform validation procedures against several sources to assess the quality and robustness of the OD matrices as well as their usefulness in transportation planning applications. We find that while the OD matrices correlate well with other sources at a low resolution, the reliability decreases when a finer level of detail is chosen, particularly when comparing shorter trips between neighbouring areas. Our results suggest that coarseness of data and privacy concerns restrict the usefulness of MPD in transport research in the case where OD matrices are pre-compiled by the operator.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Price includes VAT (Russian Federation)

Instant access to the full article PDF.

Rent this article via DeepDyve

Institutional subscriptions

Similar content being viewed by others

Assessing the Value of New Big Data Sources for Transportation Planning: Benton Harbor, Michigan Case Study

Emerging Big Data Sources for Public Transport Planning: A Systematic Review on Current State of Art and Future Research Directions

Analysis of Spatial–Temporal Characteristics Based on Mobile Phone Data

Availability of data and material.

Data and material used have not been made available due to privacy requirements.

Code availability

Code used has not been made available.

Aguiléra, V., Allio, S., Benezech, V., Combes, F., Milion, C.: Using cell phone data to measure quality of service and passenger flows of Paris transit system. Transp. Res. Part C Emerg. Technol. 43 , 198–211 (2014)

Article   Google Scholar  

Ahas, R., Aasa, A., Silm, S., Tiru, M.: Daily rhythms of suburban commuters’ movements in the Tallinn metropolitan area: case study with mobile positioning data. Transp. Res. Part C Emerg. Technol. 18 , 45–54 (2010)

Aksehirli, E., Li, Y.: Predicting MRT trips in Singapore by creating a mobility behavior model based on GSM data. In: 2018 IEEE International Conference on Data Mining Workshops (ICDMW), pp. 632–639. IEEE (2018)

Alexander, L., Jiang, S., Murga, M., González, M.C.: Origin–destination trips by purpose and time of day inferred from mobile phone data. Transp. Res. Part C Emerg. Technol. 58 , 240–250 (2015)

Asgari, F.: Inferring User Multimodal Trajectories from Cellular Network Metadata in Metropolitan Areas. Institut National des Télécommunications, Évry (2016)

Google Scholar  

Bachir, D., Khodabandelou, G., Gauthier, V., El Yacoubi, M., Puchinger, J.: Inferring dynamic origin–destination flows by transport mode using mobile phone data. Transp. Res. Part C Emerg. Technol. 101 , 254–275 (2019)

Bassolas, A., Ramasco, J.J., Herranz, R., Cantú-Ros, O.G.: Mobile phone records to feed activity-based travel demand models: MATSim for studying a cordon toll policy in Barcelona. Transp. Res. Part A Policy Pract. 121 , 56–74 (2019)

Becker, R., Cáceres, R., Hanson, K., Isaacman, S., Loh, J.M., Martonosi, M., Rowland, J., Urbanek, S., Varshavsky, A., Volinsky, C.: Human mobility characterization from cellular network data. Commun. ACM 56 , 74–82 (2013)

Calabrese, F., Di Lorenzo, G., Liu, L., Ratti, C.: Estimating Origin–Destination Flows Using Opportunistically Collected Mobile Phone Location Data from One Million Users in Boston Metropolitan Area (2011a)

Calabrese, F., Di Lorenzo, G., Liu, L., Ratti, C.: Estimating origin–destination flows using mobile phone location data. IEEE Pervasive Comput. 10 , 36–44 (2011b)

Calabrese, F., Diao, M., Di Lorenzo, G., Ferreira Jr., J., Ratti, C.: Understanding individual mobility patterns from urban sensing data: a mobile phone trace example. Transp. Res. Part C Emerg. Technol. 26 , 301–313 (2013)

Chatzikokolakis, K., Elsalamouny, E., Palamidessi, C., Pazii, A.: Methods for Location Privacy: A comparative overview. In: Foundations and Trends® in Privacy and Security, vol. 1, no. 4, pp. 199–257. Now publishers inc. https://doi.org/10.1561/3300000017 (2017)

Chen, C., Ma, J., Susilo, Y., Liu, Y., Wang, M.: The promises of big data and small data for travel behavior (aka human mobility) analysis. Transp. Res. Part C Emerg. Technol. 68 , 285–299 (2016)

Danafar, S., Piorkowski, M., Krysczcuk, K.: Bayesian framework for mobility pattern discovery using mobile network events. In: 2017 25th European Signal Processing Conference (EUSIPCO), pp. 1070–1074. IEEE (2017)

Di Lorenzo, G., Sbodio, M., Calabrese, F., Berlingerio, M., Pinelli, F., Nair, R.: Allaboard: visual exploration of cellphone mobility data to optimise public transport. IEEE Trans. Visual Comput. Graph. 22 , 1036–1050 (2015)

Doyle, J., Hung, P., Kelly, D., Mcloone, S.F., Farrell, R.: Utilising Mobile Phone Billing Records for Travel Mode Discovery (2011)

Drageide, V.: Towards Privacy Management of Information Systems. The University of Bergen, Bergen (2009)

Forrest, T.L., Pearson, D.F.: Comparison of trip determination methods in household travel surveys enhanced by a global positioning system. Transp. Res. Rec. 1917 , 63–71 (2005)

García, P., Herranz, R., Javier, J.: Big data analytics for a passenger-centric air traffic management system. Presented at the 6th SESAR Innovation Days, Delft, Netherlands (2016)

García-Albertos, P., Picornell, M., Salas-Olmedo, M.H., Gutiérrez, J.: Exploring the potential of mobile phone records and online route planners for dynamic accessibility analysis. Transp. Res. Part A Policy Pract. 125 , 294–307 (2019)

Gundlegård, D., Rydergren, C., Breyer, N., Rajna, B.: Travel demand estimation and network assignment based on cellular network data. Comput. Commun. 95 , 29–42 (2016)

Hjorthol, R., Engebretsen, Ø., Uteng, T.P.: Den nasjonale Reisevaneundersøkelsen 2013/14: Nøkkelrapport. Transportøkonomisk institutt, Oslo (2014)

Holleczek, T., Yin, S., Jin, Y., Antonatos, S., Goh, H.L., Low, S., Shi-Nash, A.: Traffic measurement and route recommendation system for mass rapid transit (MRT). In: Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1859–1868. ACM (2015)

Horn, C., Kern, R.: Deriving public transportation timetables with large-scale cell phone data. Procedia Comput. Sci. 52 , 67–74 (2015)

Huang, Z., Ling, X., Wang, P., Zhang, F., Mao, Y., Lin, T., Wang, F.-Y.: Modeling real-time human mobility based on mobile phone and transportation data fusion. Transp. Res. Part C Emerg. Technol. 96 , 251–269 (2018)

Huang, H., Cheng, Y., Weibel, R.: Transport mode detection based on mobile phone network data: a systematic review. Transp. Res. Part C Emerg. Technol. 101 , 297–312 (2019)

Iqbal, M.S., Choudhury, C.F., Wang, P., González, M.C.: Development of origin–destination matrices using mobile phone call data. Transp. Res. Part C Emerg. Technol. 40 , 63–74 (2014)

Isaacman, S., Becker, R., Cáceres, R., Kobourov, S., Martonosi, M., Rowland, J., Varshavsky, A.: Ranges of human mobility in Los Angeles and New York. In 2011 IEEE International Conference on Pervasive Computing and Communications Workshops (PERCOM Workshops), pp. 88–93. IEEE (2011)

Kalatian, A., Shafahi, Y.: Travel mode detection exploiting cellular network data. In: MATEC Web of Conferences, pp. 03008. EDP Sciences (2016)

Larijani, A.N., Olteanu-Raimond, A.-M., Perret, J., Brédif, M., Ziemlicki, C.: Investigating the mobile phone data to estimate the origin destination flow and analysis; case study: Paris region. Transp. Res. Procedia 6 , 64–78 (2015)

Li, G., Chen, C.-J., Peng, W.-C., YI, C.-W.: Estimating crowd flow and crowd density from cellular data for mass rapid transit. In: Proceedings of the 6th International Workshop on Urban Computing (in Conjunction with ACM KDD 2017) (2017)

Machanavajjhala, A., Kifer, D., Abowd, J., Gehrke, J., Vilhuber, L.: Privacy: theory meets practice on the map. In: 2008 IEEE 24th International Conference on Data Engineering, pp. 277–286. IEEE (2008)

Mamei, M., Bicocchi, N., Lippi, M., Mariani, S., Zambonelli, F.: Evaluating origin–destination matrices obtained from CDR data. Sensors (Basel) 19 , 4470 (2019)

Montero, L., Ros-Roca, X., Herranz, R., Barceló, J.: Fusing mobile phone data with other data sources to generate input OD matrices for transport models. Transp. Res. Procedia 37 , 417–424 (2019)

Ni, L., Wang, X.C., Chen, X.M.: A spatial econometric model for travel flow analysis and real-world applications with massive mobile phone data. Transp. Res. Part C Emerg. Technol. 86 , 510–526 (2018)

Phithakkitnukoon, S., Sukhvibul, T., Demissie, M., Smoreda, Z., Natwichai, J., Bento, C.: Inferring social influence in transport mode choice using mobile phone data. EPJ Data Sci. 6 , 11 (2017)

Poonawala, H., Kolar, V., Blandin, S., Wynter, L., Sahu, S.: Singapore in motion: Insights on public transport service level through farecard and mobile data analytics. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and data mining, pp. 589–598. ACM (2016)

Qu, Y., Gong, H., Wang, P.: Transportation mode split with mobile phone data. In: 2015 IEEE 18th International Conference on Intelligent Transportation Systems, pp. 285–289. IEEE (2015)

Schjalm, A.: Kvalitetsundersøkelsen for Folke- og boligtelling 1990. In: Norway, S. (ed.) Oslo—Kongsvinger: Statistics Norway (1996)

Schlaich, J., Otterstätter, T., Friedrich, M.: Generating trajectories from mobile phone data. In: Proceedings of the 89th Annual Meeting Compendium of Papers. Transportation Research Board of the National Academies (2010)

Smoreda, Z., Olteanu-Raimond, A.-M., Couronné, T.: Spatiotemporal data from mobile phones for personal mobility assessment. Transp. Surv. Methods Best Pract. Decis. Mak. 41 , 745–767 (2013)

Sørensen, A.Ø., Bjelland, J., Bull-Berg, H., Landmark, A.D., Akhtar, M.M., Olsson, N.O.: Use of mobile phone data for analysis of number of train travellers. J. Rail Transp. Plan. Manag. 8 , 123–144 (2018)

Statistics Norway: Classification of Statistical Tract and Basic Statistical Unit (2019). Available: https://www.ssb.no/en/klass/klassifikasjoner/1 . Accessed 29 Nov 2019

Stopher, P.R., Greaves, S.P.: Household travel surveys: Where are we going? Transp. Res. Part A Policy Pract. 41 , 367–381 (2007)

Stopher, P., Fitzgerald, C., Xu, M.: Assessing the accuracy of the Sydney Household Travel Survey with GPS. Transportation 34 , 723–741 (2007)

Toole, J.L., Colak, S., Sturt, B., Alexander, L.P., Evsukoff, A., González, M.C.: The path most traveled: travel demand estimation using big data resources. Transp. Res. Part C Emerg. Technol. 58 , 162–177 (2015)

Vazifeh, M.M., Zhang, H., Santi, P., Ratti, C.: Optimizing the deployment of electric vehicle charging stations using pervasive mobility data. Transp. Res. Part A Policy Pract. 121 , 75–91 (2019)

Wang, F., Chen, C.: On data processing required to derive mobility patterns from passively-generated mobile phone data. Transp. Res. Part C Emerg. Technol. 87 , 58–74 (2018)

Wang, H., Calabrese, F., Di Lorenzo, G., Ratti, C.: Transportation mode inference from anonymized and aggregated mobile phone call detail records. In: 13th International IEEE Conference on Intelligent Transportation Systems, pp. 318–323. IEEE (2010)

Wang, Z., He, S.Y., Leung, Y.: Applying mobile phone data to travel behaviour research: a literature review. Travel Behav. Soc. 11 , 141–155 (2018)

Wolf, J., Loechl, M., Thompson, M., Arce, C.: Trip rate analysis in GPS-enhanced personal travel surveys. In: Stopher, P.R., Jones, P. (eds.) Transport Survey Quality and Innovation. Emerald Group Publishing Limited, Bingley (2003)

World Bank: Mobile Cellular Subscriptions (per 100 People) (2019). Available: https://data.worldbank.org/indicator/it.cel.sets.p2 . Accessed 29 Nov 2019

Wu, W., Cheu, E.Y., Feng, Y., Le, D.N., Yap, G.E., Li, X.: Studying intercity travels and traffic using cellular network data. In: Mobile Phone Data for Development: Net Mob 2013 (2013)

Wu, C., Thai, J., Yadlowsky, S., Pozdnoukhov, A., Bayen, A.: Cellpath: fusion of cellular and traffic sensor data for route flow estimation via convex optimization. Transp. Res. Part C Emerg. Technol. 59 , 111–128 (2015)

Wu, L., Yang, B., Jing, P.: Travel mode detection based on GPS raw data collected by smartphones: a systematic review of the existing methodologies. Information 7 , 67 (2016)

Xu, C., Ji, M., Chen, W., Zhang, Z.: Identifying travel mode from GPS trajectories through fuzzy pattern recognition. In: 2010 Seventh International Conference on Fuzzy Systems and Knowledge Discovery, pp. 889–893. IEEE (2010)

Yamada, Y., Uchiyama, A., Hiromori, A., Yamaguchi, H., Higashino, T.: Travel estimation using control signal records in cellular networks and geographical information. In: 2016 9th IFIP Wireless and Mobile Networking Conference (WMNC), pp. 138–144. IEEE (2016)

Zheng, Y., Chen, Y., Li, Q., Xie, X., Ma, W.-Y.: Understanding transportation modes based on GPS data for web applications. ACM Trans. Web (TWEB) 4 , 1 (2010)

Download references

No funding to declare.

Author information

Authors and affiliations.

Department of Technology Management, SINTEF, Trondheim, Norway

Andreas Dypvik Landmark

Department of Mobility and Economics, SINTEF, Trondheim, Norway

Petter Arnesen, Carl-Johan Södersten & Odd André Hjelkrem

You can also search for this author in PubMed   Google Scholar

Contributions

Journal submission guidelines do not indicate that these are mandatory.

Corresponding author

Correspondence to Carl-Johan Södersten .

Ethics declarations

Conflict of interest.

The authors declare no conflict of interest.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Reprints and permissions

About this article

Dypvik Landmark, A., Arnesen, P., Södersten, CJ. et al. Mobile phone data in transportation research: methods for benchmarking against other data sources. Transportation 48 , 2883–2905 (2021). https://doi.org/10.1007/s11116-020-10151-7

Download citation

Published : 02 January 2021

Issue Date : October 2021

DOI : https://doi.org/10.1007/s11116-020-10151-7

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Mobile phone data
• Origin–destination (OD) estimation
• Travel surveys
• Find a journal
• Publish with us
• Track your research

Mobile phone detection and alcohol detection with engine locking for the prevention of car accidents

Ieee account.

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.