Human lower-extremity movement classification based on biomechanical sensor data Machine learning ap

Abstract

Keywords: Biomechanics; Biomedical Signal Processing; Human Movement Classification; Machine Learning; Wearable Sensor

Introduction

Between 2015 and 2050, the World Health Organization (WHO) and United Nations (UN) estimated that the proportion of the global population over 60 years old will approximately double from 12% to 22% [1]. In recent years, the concept of developing novel methods and technologies that recognize daily human activities using data from wearable biomechanical sensors has gained popularity because of an increased older population. This approach has emerged because of the potential applications of such methods in various fields, including sports science, physical therapy, rehabilitation and assistive robotic device design [2-4]. In this study, we address recent studies that have classified human lower-extremity movements based on biomechanical sensor data using machine learning (ML) algorithms and highlight the benefits and drawbacks of these approaches. Initially, we provide a brief overview of the biomechanical sensors frequently used to collect information on lower-extremity activities in humans. Secondly, the sliding time window method and feature extraction methods used in human activity recognition studies are introduced and then machine learning applications in related work are presented. Finally, we focus on previous work that makes use of wearable sensor combination in studies of human movement recognition. Wearable biomechanical sensors can be utilized to classify multiple movement patterns including walking speed (WS) in addition to estimating significant movement metrics such spatiotemporal parameters [5, 6] or comparing gait variability between healthy and pathological gait [7, 8]. Several wearable biomechanical sensors, such as accelerometers, gyroscopes, magnetometers, electromyography (EMG) and goniometers (GON) have been employed to record information about human movement pattern classification. A common combination of these sensors is the inertial measurement unit (IMU), which can provide a more comprehensive view of movement by tracking both angular and linear motions. Because of their mobility, affordability, and ability to supply precise measurements of human movement in practical settings, IMUs have grown in popularity as tools for gathering biomechanical data [9-11]. The effectiveness of movement and muscle activation during lower-limb activity depends on how a person maintains their posture. Kinematic data are gathered through motion capture cameras and digital goniometers, which measure joint angles and distance [12-14]. In terms of multiple movement patterns and gait phases, biomechanical study [15] have demonstrated considerable biomechanical variances in how human muscles and joints work. Based to this context, a wearable system’s perception of the human body is its most fundamental capability to carry out the assigned task [16, 17], making human movement recognition systems comprehensive and integrated

systems. The signal processing methods and sensor combinations applied must be compatible to clearly recognize human movement. Nagaraj et al. applied the EMG signal and angular acceleration from motion cameras as input to the ANN model for the estimation of the lower extremity joint angle of the athletes during exercise and compared it with the experimental results measured with a digital goniometer [13]. Lencioni et al. statistically analyzed the human movement during straight walking and stair climbing at different speeds with the biomechanical parameters calculated from the force plate, motion capture camera and EMG data, and compared the results [15]. A popular time series analysis method for recognizing lower extremity activity is the sliding window method. The length of the selected window affects how well this method recognizes the human movement. While processing takes more time when a wider window length is used, recognition success declines because the motion pattern is not adequately recognized for narrower window lengths. The length of the sliding window was chosen to range from 0.08 seconds to 2.5 seconds and even 30 seconds in the motion recognition tests that have been published in the literature [18]. Noor et al. proposed an adaptive method for selecting the temporal frame length for activity recognition using IMU sensors [19]. Wang et al. evaluated different window lengths for recognizing human motion patterns and observed that pattern recognition was acceptable at lengths of 0.5 s and 2.5 s–3.5 s [20]. One of the main challenges in developing such methods are selecting appropriate features from sensor data that can accurately represent different types of human movements. Machine learning algorithms that can easily learn from appropriate extracted features by ML engineers and classify activities based on them, have been used to overcome this challenge [9-11]. Huynh and Tran, in their study on human fall detection, proposed to apply frequency domain features using FFT (Fast Fourier Transform) of IMU sensor data [21]. Shawen et al., in their machine learning classification study of Parkinson’s disease detection, suggested using the time domain and frequency domain mean, standard deviation, skew, kurtosis features of IMU sensor data [22]. In another study, in which continuous-time walking speed determination algorithms and sensors for robotic knee and ankle prosthesis were evaluated, 8 time series features of each gait phase including minimum, maximum, mean, standard deviation, start value, end value, signal magnitude area, and signal energy has been used [23]. Recent studies have used machine learning algorithms including Support Vector Machines (SVM), K-Nearest Neighbor (KNN), and Artificial Neural Networks (ANN) to classify human lower-extremity movements using information gathered from biomechanical sensors [23-26]. One study classified ambulatory activities of daily living using a SVM algorithm based on information gathered from IMU and EMG sensors. The SVM algorithm’s classification accuracy of 0.943 demonstrates the promise of

Sigma J Eng Nat Sci, Vol. 42, No. 5, pp. 1378−1390, October, 2024

machine-learning methods [27]. Another study analyzed the importance of several sensor types and locations. Camargo et al. observed that biomechanical sensors such as IMU, GON, and EMG are crucial for classification, goniometers provide the majority of accuracy for estimation of stair/ramp, and a single IMU for speed estimation on the level ground dramatically reduced the need for all additional sensor types and locations [28]. In addition, Dong et al. have classified the gait phase and pattern with high accuracy using multi-channel EMG and IMU sensor data and machine learning algorithms [29]. Human movement recognition studies using wearable sensor data and machine learning tools are a very popular new research area. The most difficult issues for many applications are the precise classification and prediction of human lower-extremity movements. While understanding the user experience can improve the adaptability of rehabilitation devices to changing environments during the design phase, it can also considerably help physicians in clinical practice monitor patient activities in-depth. A few related works present studies such as fall detection [21], classification of patients with balance disorders such as Parkinson’s disease [22], the development of robotic motion assist devices [23]. Bhakta et al. evaluated the machine learning models they developed to readily predict walking speed on the data of subjects walking on a robotic knee-ankle prosthesis. They showed that using machine learning models provides high accuracy with low error rate, by applying the performance of their proposed model in various static walking speeds and dynamic speed trials [23]. Wearable biomechanical sensors, whose resolution and usage requirements have increased, are still limited in application because the algorithms used in human motion pattern recognition systems increase the computational load. This study proposes a machine learning approach-based method for human gait pattern recognition with features extracted from single and multiple combinations of IMU, GON and EMG wearable sensor signals. The sliding windows method approach was used in feature extraction and the effects of different window sizes on the classification performance were examined in the experiments. The primary aim of our study was to investigate the performance of widely used machine learning classifiers to recognize human gait pattern of single and multiple wearable biomechanical sensor signals. Our objective was to determine the appropriate temporal gait features from wearable sensors subject to machine learning algorithms: i) build signal processing methods for multiple biomechanical sensor channels, ii) observe sliding window length impacts on human movement pattern classification, iii) compare ML classification performance of single and multiple wearable sensor models.

Materials And Methods

A graphical representation of the proposed methodology is shown in Figure 1. Initially, the open-source dataset of Camargo et al. was used in this study, which included lower-extremity biomechanical and wearable sensor signals [30]. Then, biomedical signal processing and feature extraction methods are outlined for classifying the data from wearable sensors. Finally, machine learning techniques were presented to classify human lower-extremity activities.

Figure 1. The overall procedure for the classification of human gait. In this dataset, level-ground locomotion of the human lower extremity was performed at three distinct walking speeds using data from four IMU, three goniometer, and eleven EMG wearable sensors. Demographic information of 22 healthy individuals in this dataset is shown in Table

1. With respect to each subject’s preferred speed, a total of

thirty repetitions of level-ground walking have been performed at each of three self-selected speeds: fast (F), normal (N), and slow (S). Fast, normal, and slow walking speeds (mean ± standard deviation) have been calculated using the average pelvic velocities of the subjects and were observed to be 1.45±0.27 m/s, 1.17±0.21 m/s, and 0.88±0.19 m/s, respectively. The wearable biomechanical sensors have been placed unilaterally on the right side of the subjects when dataset was created. At a sampling frequency of 200 Hz, IMU

Table 1. Demographic information of subjects in this dataset Gender

Sigma J Eng Nat Sci, Vol. 42, No. 5, pp. 1378−1390, October, 2024

data has been collected from the foot, shank, thigh, and trunk sensors along with a 3-axis accelerometer (Accx, Accy, and Accz) and 3-axis gyroscope (Gyrox, Gyroy, and Gyroz). Goniometer angle data has been collected from the ankle in sagittal and frontal planes, the knee in sagittal planes and, the hip in sagittal and frontal planes, at a sample frequency of 1000 Hz. EMG data has been recorded from Gastrocnemius, Tibialis anterior, Soleus, Vastus medialis, Vastus lateralis, Rectus femoris, Biceps femoris, Semitendious, Gracilis, Gluteus medius and Right external oblique lower extremity muscles with a sampling frequency of 1000 Hz. In signal preprocessing, down-sampling method was applied to balance the sampling frequency of GON and EMG signals with the sampling frequency of IMU, and the sampling frequency of all wearable signals used in the study was obtained as 200 Hz. Each wearable sensor data varied in duration and also included redundant parts that were at the beginning and

at the end of the signal without carrying any information. For this purpose, wearable biomechanical sensor signal processing steps were applied to the signals using MATLAB (The MathWorks, 2021b). As shown in Figure 2, raw wearable sensor signals to clipped with a 5 second (s) duration to standardize signal lengths collected from movement repetitions among different subjects. Sliding window method, where window size and overlap size are the two main factors affecting segmentation performance, is very popular in the context of human activity recognition and aim to perform the sensor data segmentation preprocessing step. The length of the sliding time window will affect the movement recognition results. If the time window is too large, it may contain information from multiple activities, reducing the responsiveness of the recognition system and increasing the computational load. Conversely, if the time window is too small, some activities may be split into multiple consecutive windows, the

Figure 2. Wearable biomechanical sensors time series signal channels. (a) Foot IMU sensor, (b) Shank IMU sensor, (c) Thigh IMU sensor, (d) Trunk IMU sensor, (e) Angles of GON sensor, (f) EMGs of lower extremity muscles.

Sigma J Eng Nat Sci, Vol. 42, No. 5, pp. 1378−1390, October, 2024

Table 2. The pseudocode of proposed sliding window feature extraction method Algorithm Sliding window feature extraction for IMU time series data. Input: Fs, t, winlen, Wearable_Data Output: Wearable _labels 1:

Store Wearable _ labels[i]= Wearable _Data.fileName for classfication labels

Extract features of segData segfts (:, j, k) = [timeSeriesFts(segData), timeFreqFts(segData)]

Store average of all segments based on features Wearable _labels{i,2} = mean(segfts) end

recognition task is repeated too often but high recognition results are not obtained. As listed in Table 2, the sliding window algorithm calculates the number of segments (nSeg) by first dividing the signal time length (t) by the sliding window length (winlen). Each channel here represents the wearable sensor degrees of freedom shown in Figure 2. For each channel, time series and time-frequency domain features of each segment were then computed. These sequential tasks were processed for all segments in a channel, and then the average of the segments was determined. Several window lengths of between 100 (0.5s) and 1000 (5s) was analyzed using IMU and GON sensor signals in this study. In addition, in the study, EMG sensors were also included in the experiments to evaluate the motion recognition performance of wearable sensor combinations and the classification performances of the sensor combinations were compared.

The features were extracted from each signal segment segmented by the sliding windows method, and the average of these features obtained from the segments was applied as an input to machine learning. While mean, maximum (max), standard deviation (std), minimum (min), median absolute deviation (mad), interquartile range (iqr), and area of under curve (AUC) were used as the signal time series features, the time-frequency domain features were max, skewness, kurtosis and AUC calculating Fast Fourier Transform (FFT) of each wearable signal channel [21-23]. The degrees of freedom and number of extracted features for single and multiple wearable sensors are given in Table 3. The extracted feature matrix and classification procedures compose the two sections of the ML classification procedure. Initially, the extracted feature matrices were constructed to the type of the wearable sensors based on

Table 3. Extracted feature matrix of wearable sensor signals for ML models Sensor combination

Imu-Gon-Emg

Sigma J Eng Nat Sci, Vol. 42, No. 5, pp. 1378−1390, October, 2024

Esknn

the sliding window length. In this study, there was a total of 675 class labels, consisting 225 ‘F’, 221 ‘N’, and 229 ‘S’ labels. Then, ML classification section were evaluated using MATLAB (The MathWorks, R2021b) Classification Learner Toolbox. The classification process was performed using three different ML algorithms, which are frequently employed in human movement classification problems: Ensemble Subspace K-nearest neighbors (ESKNN), Quadratic Support Vector Machine (QSVM) and Wide Neural Networks (WNN). The K-Nearest Neighbors (KNN) non-hierarchical clustering approach detects input data characteristics based on likely similarity with their neighbors. A neighborhood has been set to contain k points, and the closest neighbors of a query point are determined using a distance metric such as the Euclidean distance. Support vector machine (SVM) is a supervised classifier that transforms human gait data, which is fundamentally nonlinear data, into a higher dimensional feature space via different kernel approaches including linear, polynomial, and Gaussian RBF. This classification is achieved by finding the most appropriate separating hyperplane that provides the balance between the maximum margin width and the minimum classification error between the predicted classes. A multilayer feed forward neural network known as an ANN consists of a collection of interconnected neurons, with connections between units occurring solely through hidden layers from input to output layers. All ML models were analyzed with the training hyper parameters included in Table 4 with repeated 10-fold cross-validation to avoid over-fitting. Performance metrics such as accuracy, receiving operating curve-area under curve (ROC-AUC) and confusion matrix are used to determine how well the analyzed ML classifiers perform on the training, test and validation set. When a model correctly classifies positive and negative classes, the outputs are True Positive (TP) and True Negative (TN), respectively. Similarly, False Positive (FP) and False Negative (FN) are the model’s outcomes when the positive and negative classes are misclassified, respectively. Classification accuracy is calculated by dividing the sum of the TP and TN by the total samples.

Results And Discussion

The classification accuracies of the ML classifiers using different sliding window length based on IMU and GON sensor for gait classification are shown in Figure 3(a) and in Figure 3(c), respectively. Similarly, AUC-ROC scores of ML classifiers using IMU and GON sensor models are represented in Figure 3(b) and in Figure 3(d). The ML classification accuracies of IMU-based sensor and GON-based sensor were increasing with increased window lengths. The highest accuracies were 0.954 for IMU sensor and 0.828 for goniometer sensor using window length of 1000 (winlen-1000) and QSVM classifier. Similarly, the ROC-AUC score of IMU-based sensor and GON-based sensor were gradually increased with increased window lengths. In case of QSVM and window lengths of 1000, just as the ROCAUC score of IMU-based sensor reached 0.99, the ROCAUC score of GON-based sensor reached 0.97. It can be seen from these graphs that the classifiers have different classification accuracy and ROC-AUC score for IMU and GON sensor models. QSVM has higher accuracy in gait speed recognition compared to other classifiers. In particular, the accuracy in the experiment using sliding window length of 1000 based on the IMU sensor and QSVM classifier is the highest. In IMU sensor, the accuracy was found to be higher than those with the GON sensor. In gait recognition with QSVM classifier using IMU sensor, winlen-1000 (0.954), winlen-500 (0.913) and winlen-250 (0.861) models are seen as the models showing the highest accuracy. Likewise, ROC-AUC scores were also ranked by the accuracy of the models. When the ML classification performance were evaluated using wearable sensors, the overall (all ML classifiers, sensor types and window lengths) accuracies and AUC-ROC scores were 0.954 and

Imu-Gon-Emg

achieved in IMU-GON (0.961), IMU-EMG (0.895) and GON-EMG (0.776) sensor models, respectively. Finally, the accuracy of the IMU-GON-EMG model, in which all three sensors are included, is 0.919. The findings of this study showed that IMU sensor models improved the classification performance in level-ground gait pattern recognition, and their use together with GON sensor models contributed positively to this performance. It has been found that EMG sensor models show lower classification performance compared to IMU sensor models. It can be seen that the proposed method gave results that were roughly comparable to those of the studies mentioned in previous studies, when the wearable sensor combination models and only the IMU sensor model are compared in human gait pattern recognition. Dong et al. evaluated the performance of four algorithms as SVM, ANN, AlexNet and LeNet5 in recognizing gait phases and patterns. Their experimental results showed that with the model using the EMG and IMU signal together, the four algorithms were able to achieve a recognition accuracy of 0.977 for gait phases and an average recognition accuracy of over 0.992 for gait patterns. In the model in which they used only the IMU sensor with the SVM algorithm, they reached 0.941 and 0.987 accuracy for gait phases and gait pattern, respectively [29]. With the same dataset we used in the study, Camargo et al. performed speed estimation in level-ground locomotion using dataset, including EMG, GON, and IMU data. Even with a single IMU, it has been demonstrated that using an IMU sensor for speed estimation rather than an EMG or GON sensor decreases model error. They revealed that ML-based SVM models for determining walking speed had the lowest classification error rate. Additionally, they demonstrated in their study which sensor type is significant for various walking area conditions, indicating that mechanical sensors like IMU and GON are more significant for classification than EMG sensors in estimating walking speed on level ground [28]. These results point out the possible utility of walking speed as biomechanical indicators, from which it can be

Sigma J Eng Nat Sci, Vol. 42, No. 5, pp. 1378−1390, October, 2024

extended using the sliding time-window method and different machine learning classifiers proposed in this study. The gait recognition approaches in this study include the single use of biomechanical sensor models as well as the multiple use of their different combination models. Table 8 demonstrates that the highest classification accuracy is 0.961 and 0.954, respectively, when the model including IMU and GON sensor and the single IMU sensor model are used to classify human gait with the QSVM classifier. In addition, in the classification of gait pattern with the ESKNN classifier, the lowest accuracy was found to be 0.548 and 0.776, respectively, when the EMG sensor model and the model including the EMG and GON sensor were used. This result can be attributed to the efficiency of the IMU sensor model in correctly addressing temporal gait parameters, and also that EMG sensor models with machine learning approach lag behind IMU and GON sensor models in gait recognition.

Conlusion

Human movement classification studies can readily be performed using machine learning algorithms and wearable biomechanical sensors. The technique used in this study may also make it possible to track stride length and walking speed on an individual basis for wearable assistive device designs and rehabilitative gait exercise programs. Consequently, there is a clear demand for applicable systems that can accurately categorize human movement and analyze biomechanical data. In this study, we present a wearable biomechanical sensor-based system for recognizing human walking movement. The proposed approach requires extracting the proposed features, namely time domain and frequency domain features, by segmenting using the sliding windows method from various biomechanical signals based on IMU, GON and EMG sensor data. These variables performed incredibly well in classifying human movement based on walking speeds when applied as inputs for machine learning classifiers. Furthermore, our experiment findings highlighted that the proposed approach was capable of distinguishing between healthy human subjects walking at three different speeds (Fast, Normal, and Slow) and that the IMU sensor and its sensor combination models were able to achieve an average accuracy of 0.932 for movement classification with the QSVM classifier using sliding window length of 1000. The methodology used in this study may be used as a model for future significant strategies for the complete diagnosis and phase assessment of gait disorders, as well as for the identification of gait biomarkers for assistive and rehabilitation wearable technologies. The outcomes of this study may also help assess the progression of gait disorders by extending machine learning approaches from the research area to the realm of biomedicine. Further study, the same lower limb biomechanics dataset would be utilized

to create popular deep learning models like 2D-CNN and LSTM to recognize human activities including ramp and stair movement.

Acknowledgments

We would like to thank Camargo J, et al. to provide open-source dataset of lower limb biomechanics in multiple conditions.

Authorship Contributions

HK, the machine learning analysis and model, biomedical signal processing, interpretation of the data, original draft preparation and revision of the manuscript. FK, supervision, the development of the machine learning model, and revising of the manuscript. All authors read and agreed to the published version of the manuscript.

Data Availability Statement

The authors confirm that the data that supports the findings of this study are available within the article. Raw data that support the finding of this study are available from the corresponding author, upon reasonable request.

Conflict Of Interest

The author declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Ethics

There are no ethical issues with the publication of this manuscript.

References

1. Celik Y, Stuart S, Woo WL, Godfrey A. Gait analysis [19] Noor MHM, Salcic Z, Kevin I, Wang K. Adaptive in neurological populations: Progression in the use sliding window segmentation for physical activity of wearables. Med Eng Phys 2021;87:9−29. [CrossRef] recognition using a single tri-axial accelerometer.
2. Teufl W, Lorenz M, Miezal M, Taetz B, Fröhlich Pervasive Mob Comput 2017;38:41−59. [CrossRef] M, Bleser G. Towards inertial sensor based mobile [20] Wang G, Li Q, Wang L, Wang W, Wu M, Liu T. Impact gait analysis: Event-detection and spatio-temporal of sliding window length in indoor human motion parameters. Sensors (Basel) 2018;19:38. [CrossRef] modes and pose pattern recognition based on smart-
3. Yang M, Zheng H, Wang H, McClean S, Hall J, phone sensors. Sensors (Basel) 2018;18:1965. [CrossRef] Harris N. A machine learning approach to assessing [21] Huynh QT, Tran BQ. Time-frequency analysis of gait patterns for complex regional pain syndrome. daily activities for fall detection. Signals 2021;2:1−12. Med Eng Phys 2012;34:740−746. [CrossRef] [CrossRef]
4. Khan MH, Farid MS, Grzegorzek M. Spatiotemporal [22] Shawen N, O'Brien MK, Venkatesan S, Lonini L, features of human motion for gait recognition. Signal Simuni T, Hamilton JL, et al. Role of data measure- Image Video Process 2019;13:369−377. [CrossRef] ment characteristics in the accurate detection of
5. Ferreira GA, Teixeira JLS, Rosso ALZ, de Sá AMF. Parkinson's disease symptoms using wearable sen- On the classification of tremor signals into dyskine- sors. J Neuroeng Rehabil 2020;17:1−14. [CrossRef] sia, Parkinsonian tremor, and Essential tremor by using machine learning techniques. Biomed Signal [23] Bhakta K, Camargo J, Compton W, Herrin K, Young Process Control. 2022;73:103430. [CrossRef] A. Evaluation of continuous walking speed determi-
6. Ma C, Li D, Pan L, Li X, Yin C, Li A, et al. Quantitative nation algorithms and embedded sensors for a pow- assessment of essential tremor based on machine ered knee & ankle prosthesis. IEEE Robot Autom learning methods using wearable device. Biomed Lett 2021;6:4820−4826. [CrossRef] Signal Process Control 2022;71:103244. [CrossRef] [24] Panyakaew P, Pornputtapong N, Bhidayasiri R.
7. Nasrabadi AM, Eslaminia AR, Bakhshayesh PR, Using machine learning-based analytics of daily Ejtehadi M, Alibiglou L, Behzadipour S. A new activities to identify modifiable risk factors for scheme for the development of IMU-based activity falling in Parkinson's disease. Parkinsonism Relat recognition systems for telerehabilitation. Med Eng Disord 2021;82:77−83. [CrossRef] Phys 2022;108:103876. [CrossRef] [25] Trabassi D, Serrao M, Varrecchia T, Ranavolo
8. Aliman N, Ramli R, Haris SM, Amiri MS, Van M. A A, Coppola G, De Icco R, et al. Machine learning robust adaptive-fuzzy-proportional-derivative con- approach to support the detection of Parkinson's troller for a rehabilitation lower limb exoskeleton. disease in IMU-based Gait analysis. Sensors (Basel) Eng Sci Technol Int J 2022;35:101097. [CrossRef] 2022;22:3700. [CrossRef]
9. Nagaraj G, Mir BA, Gomathy B, Leelavathy S, Sengupta [26] Xi X, Tang M, Miran SM, Luo Z. Evaluation of fea- A, Ahmad SS. Artificial Neural Network to Predict ture extraction and recognition for activity moni- Swinging of Lower Limb in Jumping Jack Exercise. toring and fall detection based on wearable sEMG In: 2023 International Conference on Intelligent Data sensors. Sensors (Basel) 2017;17:1229. [CrossRef] Communication Technologies and Internet of Things [27] Zhou B, Wang H, Hu F, Feng N, Xi H, Zhang Z, et al. (IDCIoT). IEEE; 2023. p. 914−918. [CrossRef] Accurate recognition of lower limb ambulation mode
10. Vijayvargiya A, Khimraj, Kumar R, Dey N. Voting- based on surface electromyography and motion data based 1D CNN model for human lower limb activity using machine learning. Comput Methods Programs recognition using sEMG signal. Phys Eng Sci Med. Biomed 2020;193:105486. [CrossRef] 2021;44:1297−1309. [CrossRef]
11. Lencioni T, Carpinella I, Rabuffetti M, Marzegan A, Kanwar B, Young A. A machine learning strategy Ferrarin M. Human kinematic, kinetic and EMG for locomotion classification and parameter estima- data during different walking and stair ascending and descending tasks. Sci Data 2019;6:309. [CrossRef] tion using fusion of wearable sensors. IEEE Trans
12. Dhiman C, Vishwakarma DK. A review of state- Biomed Eng 2021;68:1569−1578. [CrossRef] of-the-art techniques for abnormal human activity [29] Dong D, Ma C, Wang M, Vu HT, Vanderborght B, recognition. Eng Appl Artif Intell 2019;77:21−45. Sun Y. A low-cost framework for the recognition [CrossRef] of human motion gait phases and patterns based
13. Kenan E, Kutlu MÇ, Barış B. Comparison of gesture on multi-source perception fusion. Eng Appl Artif classification methods with contact and non-contact Intell 2023;120:105886. [CrossRef] sensors for human-computer interaction. Sigma J [30] Camargo J, Ramanathan A, Flanagan W, Young A. Eng Nat Sci 2022;40:219−226. A comprehensive, open-source dataset of lower
14. Ma C, Li W, Cao J, Du J, Li Q, Gravina R. Adaptive limb biomechanics in multiple conditions of stairs, sliding window based activity recognition for ramps, and level-ground ambulation and transi- assisted livings. Inf Fusion 2020;53:55−65. [CrossRef] tions. J Biomech 2021;119:110320. [CrossRef]

Share and Cite