Human Activity Recognition Using Inertial, Physiological and Environmental Sensors: A Comprehensive Survey

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The most common types of sensors used in activity recognition are accelerometers, mainly because of their small size and low cost.

“The most common types of sensors used in activity recognition are accelerometers, mainly due to their small size and low cost.”

May combine two claims.

Among all reviewed published papers, proposed HAR methods are based mostly on standalone devices, but the total number of smartphone- and smartwatch-based methods is higher than those based on standalone devices.

“Finally, as shown in Figure 8.(a), among all the reviewed published papers, the proposed HAR methods are based mostly on standalone devices. However, the total number of smartphone- and smartwatch-based methods are higher than those based on standalone devices.”

May combine two claims.

Among the 149 selected papers, 149 used accelerometers, 83 used gyroscopes in addition to accelerometers, and 27 used a magnetometer in addition to the accelerometer.

“As depicted in Figure 7.(a), 149 papers used accelerometers, 83 used gyroscopes in addition to accelerometers, and 27 used a magnetometer in addition to the accelerometer.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The average number of citations of the chosen articles was 46.

“The average number of citations was 46, and the distribution of the papers for each year is shown in Table II.”

May combine two claims.

All selected papers use at least one accelerometer or at least one accelerometer in combination with other sensors.

“Therefore, all the selected papers use at least one accelerometer or at least one accelerometer in combination with other sensors.”

May combine two claims.

No existing survey article presents a comprehensive meta-review of existing surveys, provides a comprehensive overview of different sensors, reports and compares performance metrics, and reports on dataset availability and popularity.

“To the best of our knowledge, no existing survey article is (1) presenting a comprehensive meta-review of the existing surveys, (2) providing a comprehensive overview of different sensors, (3) reporting and comparing performance metrics, and (4) reporting on dataset availability and popularity.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

In many cases, accelerometers are used together with other sensors including gyroscopes, magnetometers, compasses, pressure sensors, body temperature sensors, electromyography, oximetry sensors, and electrocardiographs.

“In many cases, accelerometers are used in conjunction with others sensors including gyroscopes, magnetometers, compasses, pressure sensors, body temperature sensors, electromyography, oximetry sensors, and electrocardiographs.”

May combine two claims.

Average activity recognition accuracy among the 53 DL-based papers and the 96 CML-based papers was 93% and 92.2%, respectively, showing almost the same recognition quality.

“Figure 1.(b) shows the average activity recognition accuracy, among the 53 DL-based papers and the 96 CML-based papers, that as visible (93% DL-based and 92.2% CML-based) present almost the same recognition quality.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Other detailed movement recognition involves specific movements of the arms, such as carrying or reaching an object, releasing it, frontal elevation, and other activities performed in relation to other objects.

“Other detailed movement recognition involves specific movements of the arms, such as carrying/reaching an object, releasing it, frontal elevation, and other activities that people can perform in relation to other objects [76]–[78].”

May combine two claims.

Among 149 selected published papers on HAR from 2015 to 2019, 53 were based on DL models and 96 were based on CML models.

“among a total of 149 selected published papers on HAR, 53 were based on DL models, and 96 were based on CML models.”

May combine two claims.

The keyword search produced a total of 249,110 records, from which 293 were selected based on the quality of the publication venue.

“Our keywords produced a total of 249110 records, among which we selected 293 based on the quality of the publication venue.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Other activities explored in HAR research include complex activities such as different phases of cooking, house cleaning, driving, smoking, swimming, or biking.

“However, other type of activities have been explored in the past few years, including complex activities, such as the different phases of cooking [59], house cleaning [4], [59]–[61], driving [62]–[65], smoking [66], swimming [67], or biking [6], [43], [64], [68].”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

The largest amount of data on daily life is collected via electronic devices such as smartphones, smartwatches, activity trackers, smart thermostats, and video cameras.

“However, as we discussed earlier, the largest amount of data on daily life is collected via electronic devices, such as smartphones, smartwatches, activity trackers, smart thermostats and video cameras.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Among all activities of daily life, the most popular activities in HAR research are walking, running, standing, sitting, walking upstairs, and walking downstairs.

“Among all ADL’s, the most popular activities in HAR research are walking, running, standing, sitting, walking upstairs and walking downstairs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The most commonly studied activities in these datasets are primary activities of daily life such as walking, running, sitting, standing, walking upstairs, walking downstairs, and sleeping.

“The most common studied activities are primary activities of daily life, such as walking, running, sitting, standing, walking upstairs, walking downstairs, and sleeping.”

May combine two claims.

CNN is the most popular DL model, referenced in 30 papers.

“The most popular model is the Convolutional Neural Network (CNN), which was referenced in 30 papers”

May combine two claims.

Among the 149 analyzed HAR models, 87 were tested on a single dataset, and 62 were tested on more than one dataset.

“Among the 149 analyzed HAR models, 87 models were tested on a single dataset, with the remaining 62 tested on more than one dataset.”

May combine two claims.

SVM is the most commonly used classical ML model, used in 35 papers, achieving an average accuracy of 92.3% over an average of 12 activities.

“the most commonly used model was the Support Vector Machine (SVM) model [4], [6], [58], [60], [69], [78], [79], [82], [85], [92], [95], [109], [118], [127], [131], [132], [136], [138], [142], [145], [155], [168], [169], [171], [173], [182], [184], [186], [187], [189], [190], [209]–[212] which was used in 35 papers, achieving an average accuracy of 92.3% over an average of 12 activities.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The open-source HAR datasets listed in Table VI predominantly use accelerometer, gyroscope, and magnetometer sensors.

“Such datasets are largely based on accelerometer, gyroscope, and magnetometer sensor data.”

May combine two claims.

28 HAR methodologies were tested on two datasets, 21 on three datasets, fewer than 10 on 4-6 datasets, and only one methodology was tested on 14 datasets.

“we found that 28 HAR methodologies were tested on two datasets, 21 HAR methodologies on three datasets, less than 10 HAR methodologies on 4–6 datasets, and only one methodology [219] was tested on a total of 14 datasets.”

Probably not a substantial claim: may not pass S3 (load-bearing).

A variety of sensors and devices are used for data collection.

“a variety of sensors and devices are used for data collection.”

May combine two claims. The quote only partly states this claim.

RF is the fourth most used classical ML model, used in 15 papers, obtaining an average accuracy of 93.3% over an average of 10 activities.

“and the Random Forest (RF) [6], [57], [69], [72], [78]–[80], [82], [92], [93], [95], [96], [175], [185], [212], which was used in 15 papers, obtaining an average accuracy of 93.3% over an average of 10 activities of daily life.”

May combine two claims.

The most used normalization steps are z-normalization, min-max, and linear interpolation, preceded by filtering steps such as outlier detection, Butterworth, median, high-pass, or statistical filters.

“Among all such techniques the most used ones are: z-normalization [75], [120], min-max [70], [127], and linear interpolation [102], [111] are the most used normalization steps, preceded by a filtering step based on the application of outlier detection [70], [117], [163], Butterworth [82], [101], [117], [123], [127], [128], [152], [155], [174], [189], median [74], [101], [117], [127], [132], [147],”

Probably not a substantial claim: may not pass S3 (load-bearing).

CML models are more popular than DL models.

“CML models still enjoy great popularity compared to those based on the relatively more recent and more advanced models such as the DL models.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

CML-based approaches are still widely used and generate outstanding results without the computational costs.

“CML-based approaches are still widely used, and they generate outstanding results without the computational costs.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Many articles used different classification models, not just one, for better performance.

“many articles made use of different classification models and not just one model for achieving better performance”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

The most widely used, if not indispensable, sensor is the accelerometer, which can be used with other sensors such as the gyroscope or magnetometer.

“the most widely used used, if not indispensable, sensor is the accelerometer, which can be used in conjunction with other sensors such as the gyroscope or the magnetometer.”

May combine two claims.

Other DL-based articles use time-domain, frequency-domain, or both time- and frequency-domain features.

“other DL-based articles do make use of time-domain (Tab. VIII, row 12) features, frequency domain (Tab. VIII, row 11) and both time and frequency domain (Tab. VIII, rows 13 and 14) features.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Some articles tested their approaches using different models.

“some of the articles have tested their approaches using different models.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Activity recognition algorithms based on ML and DL are becoming central in HAR.

“Activity recognition algorithms based on ML and DL are becoming central in HAR.”

May combine two claims.

Most proposed HAR models are only tested on a limited number of activities and captured in a single controlled environment.

“most proposed HAR models are only tested on a limited number of activities and captured in a single controlled environment.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The number of activities in these datasets ranges from two to thirty-three.

“the number of activities in these datasets ranges from two [202] to thirty-three [205] (Table VI).”

May combine two claims.

kNN is the second most used classical ML model, used in 23 papers, achieving an average accuracy of 93.7% over an average of 12 activities.

“The second most used model is the classical k-Nearest Neighbor (kNN) model [4], [6], [42], [60], [61], [69], [78], [79], [92], [95], [96], [113], [118], [122], [127], [136], [142], [145], [162], [164], [169], [173], [186], which was used in 23 papers, achieving an average accuracy of 93.7% over an average of 12 activities of daily life.”

May combine two claims.

LSTM is the second most used DL model, used in 17 papers.

“The second most used model was the Long Short-Term Memory (LSTM) model, which was used in 17 papers”

May combine two claims.

NN is the fifth most used classical ML model, used in 14 papers, obtaining an average accuracy of 93.5% over an average of 8 activities.

“The fifth most used model is the Neural Networks (NN) [4], [78], [92], [98], [114], [136], [142], [145], [148], [157], [173], [183]–[185], which was used in 14 papers, obtaining an average accuracy of 93.5% over an average of 8 activities of daily life.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Many HAR methodologies use overlapping segments between consecutive windows, which provides the model with knowledge of previous context.

“Besides, many methodologies maintain an overlapping part between two consecutive segments. This part provides the model with knowledge of the previous context.”

May combine two claims.

Most reviewed CML-based articles use both time- and frequency-domain features and apply noise removal preprocessing.

“most of the reviewed articles (Tab. VIII, row 7) make use of both time and frequency domain features, and the raw data was initially pre-processed with noise removal techniques.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Most sensors in these datasets are embedded in smartphones and smartwatches.

“Most of such sensors are embedded into smartphones and smartwatches”

May combine two claims.

Some methodologies do not use any feature extraction technique, and in some cases noise is not considered.

“other methodologies (Tab. VIII, rows 8 and 9) do not make use of any features extraction technique, and in some cases, the presence of noise is not considered”

May combine two claims.

RNN models were used in 8 papers, over an average of 14 activities, obtaining an average accuracy of 95%.

“Recurrent Neural Network (RNN) were used in [8], [56], [89], [112], [129], [213], [216], [217], over an average number of 14 obtaining an average accuracy of 95%.”

May combine two claims.

The remaining DL papers (Other) were based on models such as Autoencoders, Inception Neural Networks, or other frameworks, totaling 7 papers with an average accuracy of 91.1% and an average of 17 activities.

“the rest of the papers (indicated by Other in Figure 9) where based on models such as Autoencoders [71], [123], Inception Neural Networks (INN), or the other frameworks [105] for a total of 7 papers with an average accuracy of 91.1% and an average number of 17 activities of daily life.”

May combine two claims.

DT is the third most used classical ML model, used in 19 papers, obtaining an average accuracy of 94.2% over an average of 8 activities.

“The third and fourth most used model are the Decision Tree (DT) model [6], [78], [85], [94], [95], [113], [136], [142], [145], [159], [165], [173], [177], [178], [184], [185], [193], [208], [210], [212], which was used in 19 papers, obtaining an average accuracy of 94.2% over an average of 8 activities of daily life”

May combine two claims.

Advanced activity tracking devices can recognize sleeping and the neurological stages of sleep, cycling through nREM stages 1 to 4 and REM.

“Advanced devices can recognize sleeping and the neurological stages of sleep (i.e., cycling through nREM (stages 1–4) and REM) [86]”

• cites [86] sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

The large inter-subject and intra-subject variability of human movement entails considerable difficulty in developing a methodology that generalizes among all subjects.

“Such variability entails a considerable difficulty in developing a methodology that manages to generalize among all subjects.”

• AssumptionRests on an assumption: Human movement patterns are very hard to recognize because of large inter-subject and intra-subject variability.

May combine two claims.

While DL methods produce high accuracy on large activity datasets, in many HAR applications Classic Machine Learning models might be better suited because of small dataset size, lower input dimensionality, and availability of expert knowledge.

“While DL methods produce high accuracy results on large activity datasets, in many HAR applications Classic Machine Learning (CML) models might be better suited due to the small size of the dataset, lower dimensionality of the input data, and availability of expert knowledge in formulating the problem [11].”

• cites [11] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Gated Recurrent Neural Networks have shown advantages in modeling sequential dependencies in long-term time-series.

“Gated RNNs have shown advantages in modeling sequential dependencies in long-term time-series [44].”

• cites [44] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Deep Learning techniques can automatically generate optimal features from raw input data without human intervention, making it possible to identify unknown patterns.

“Since DL is based on the idea of the data representation, such techniques can automatically generate optimal features, starting from the raw input data, without any human intervention, making it possible to identify the unknown patterns that otherwise would remain hidden or unknown [44].”

• cites [44] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Human Activity Recognition has become a popular research topic in the last decade because of its importance in health care, interactive gaming, sports, and general monitoring systems.

“Human Activity Recognition (HAR) has become a popular topic in the last decade due to its importance in many areas, including health care, interactive gaming, sports, and monitoring systems for general purposes [1].”

• cites [1] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Because of the limitations of Deep Learning, Classic Machine Learning methods are still preferred in some areas, especially when the training dataset is small or fast training is required.

“Because of such limitations, in some areas still CML methods are preferred, especially when the training dataset is quite small, or when fast training is a requirement.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The most common sensor fusion methods are typically based on Bayesian estimation, Kalman Filters, and Particle Filtering.

“The most common sensor fusion methods are typically based on Bayesian estimation, Kalman Filters, and Particle Filtering techniques [58].”

• cites [58] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Deep Learning models present limitations including being black-box models, difficult interpretation, requiring large datasets for training, and high computational cost.

“However, as already mentioned, DL models also present some limitation [45]: Black-box models, interpretation is not easy and inherent,Require large datasets for training,High computational cost.”

• cites [45] Marcus, 2018, Deep learning: A critical appraisalNot checked yet

  Not checked in the free run.

May combine two claims.

HAR research has seen an explosion in Deep Learning methods, resulting in an increase in recognition accuracy.

“In terms of algorithmic implementation, HAR research has seen an explosion in Deep Learning (DL) methods, resulting in an increase in recognition accuracy [5], [7].”

• cites [5] sourceNot checked yet

  Not checked in the free run.

• cites [7] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Deep Learning algorithms have become popular in many domains because of their superior performance.

“On the other side, in recent years, DL algorithms have become popular in many domains, due to their superior performance [4].”

• cites [4] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Collecting video data for HAR presents privacy and computational issues.

“Cameras have been widely used in the HAR applications, however collecting video data presents many issues regarding privacy and computational requirements [10].”

• cites [10] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

HAR algorithms face challenges including the complexity and variety of daily activities, intra-subject and inter-subject variability, the trade-off between performance and privacy, computational efficiency in embedded and portable devices, and difficulty of data annotation.

“Despite myriad applications, HAR algorithms still face many challenges, including 1) complexity and variety of daily activities, 2) intra-subject and inter-subject variability for the same activity, 3) the trade-off between performance and privacy, 4) computational efficiency in embedded and portable devices, and 5) difficulty of data annotation [3].”

• cites [3] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The review's retrieval process followed the PRISMA template for systematic reviews.

“Figure 4 shows our retrieval process based on PRISMA template for systematic reviews [37].”

• cites [37] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

DL models have a unique ability to recognize more complex activities while maintaining high accuracy.

“DL models have a unique ability to recognize more complex activities, while maintaining high accuracy.”

May combine two claims.

The interpretability issue is related to tests performed using commercial tools, lack of open source code, and authors not publicly providing their source code.

“Such an issue is related to tests performed using commercial tools, lack of open source code, and authors who do not publicly provide their source code.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

DL models do not require a data preprocessing stage.

“they do not require a data preprocessing stage.”

• AssumptionRests on an assumption: DL networks perform automatic feature extraction without human intervention, unlike traditional machine-learning algorithms.

May combine two claims.

HAR researchers still prefer classic ML models, mainly because they require a smaller amount of data and less computational power than DL models.

“HAR researchers still prefer classic ML models, mainly because they require a smaller amount of data and less computational power than DL models.”

• AssumptionRests on an assumption: CML models require a smaller amount of training data and lower computational requirements than DL models.

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Sensor-based HAR has many advantages compared to vision-based HAR methodologies, which pose privacy concerns and are constrained by computational requirements.

“sensor-based HAR have many advantages compared to vision-based HAR methodologies, which pose privacy concerns and are constrained by computational requirements.”

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

This situation shows the challenge of identifying a methodology superior to the others.

“This situation shows the challenge of identifying a methodology superior to the others.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Because HAR imposes constraints such as reduced latency, memory constraint, and computational constraints, the common supervised classifiers except SVM are appropriate for low-resource environments given their low computational and memory requirements.

“Furthermore, since HAR imposes specific constraints, such as reduced latency, memory constraint, and computational constraints, these classifiers, except for SVM, are appropriate for low-resource environments given their low computational and memory requirements.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Inertial sensors are widely used in clinical and healthcare applications.

“Such inertial sensors are widely used in clinical and healthcare applications. [194]”

• cites [194] sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

HAR systems have become a growing research area in the past decade, achieving impressive progress.

“HAR systems have become a growing research area in the past decade, achieving impressive progress.”

May combine two claims.

Preprocessing raw data eliminates noise and prepares it for recognition models.

“The presence of noise is handled by preprocessing the raw data to eliminate this interference and prepare the data for being feed to the recognition models [35].”

• cites [35] sourceNot checked yet

  Not checked in the free run.

May combine two claims.

CML models require a smaller amount of training data and lower computational requirements than DL models.

“CML models require a smaller amount of training data, as well as lower computational requirements.”

• AssumptionRests on an assumption: CML models require a smaller amount of training data and lower computational requirements than DL models.

May combine two claims.

Time-domain features are more commonly used than frequency-domain features because they are less computationally expensive due to the transform required for frequency-domain features.

“Time-domain features are the most used because they are cheaper than the frequency domain features because of the transform from time to frequency domain [35].”

• cites [35] sourceNot checked yet

  Not checked in the free run.

Classified as probably not a substantial claim.

One of the main limitations of HAR algorithms is the lack of standardized methodologies that can generalize to heterogeneous sets of activities performed by diverse users.

“One of the main limitations of HAR algorithms is the lack of standardized methodologies that can generalize to heterogeneous set of activities performed by a diverse set of users.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Unsupervised clustering techniques are particularly suitable when working with datasets lacking labels or when the measure of similarity or dissimilarity between classes is a primary outcome.

“These techniques are particularly suitable when working with datasets lacking labels or when the measure of similarity/dissimilarity between classes is a primary outcome [41]–[43].”

• cites [41]–[43]Not checked yet

  Not checked in the free run.