Table 6.

Comparison of various methods to automatically detect human activity monitoring using wearable sensor technology.

Ref	ML Model/NN Type	Details	Epochs	No. of Participants	Test for Analysis	Results
[166]	(CNN) and (LSTM-RNN)	TensorFlow is used to implement the NN.	40	22	Accuracy (84%)	CNNs may perform better than LSTM-RNN for real-time datasets.
[167]	CNN with the Deep Q Neural Network (DQN) model compared with LSTM models and DQN	CCR, EER, AUC, MAP and the CMC.		50	Classification accuracy (98.33%)	CNN model performing better than the LSTM model.
[176]	1-D Convolutional neural network (1-D CNN)—a RNN model with LSTM	3+3 C-RNN designed for data processing.	1000	80	Accuracy (90.29%)	Model works well for lower sampling rates. However, for large data set accuracy is getting lower.
[135]	Hierarchical Dirichlet process (HDP) model to detect human activity levels	SVM		27	Precision of 0.81 and recall of 0.77.	(HDP) model that can infer the number of levels automatically from a sliding window time duration.
[168]	Apriori Algorithm and Pattern Recognition (PR) Algorithm	New algorithm for PR is designed and implemented in MATLAB.		9	Standard deviation of Predicted v/s Actual Graph (Standard Deviations were around 2.6 for PR-Algorithm and 3.32 for Apriori algorithm).	PR algorithm indicated better prediction than the Apriori algorithm.
[177]	Hierarchical Dirichlet Process Model (HDPM)	Feed forward neural network.	50	201	Simple accuracy (sitting—78.60%, standing—9.45%, walking—26.87%)	The physical activity levels are automatically learned from the input data using the HDPM.
[169]	HAR method based on U-Net	CNN	100	266,555 samples and 5026 windows	Accuracy and Fw-score (Max. Accuracy of 96.4% and Fw-Score of 0.965).	U-Net method overcomes the multiclass window problem inherent in the sliding window method and realises the prediction of each sampling point’s label in time series data.
[170]	InnoHAR—DL model	Combination of inception neural network and RNN structure built with Keras.		9	Opportunity, PAMAP2, and Smartphone datasets with F-scores of 0.946, 0.935 and 0.945, respectively.	Consistent superior performance and has good generalisation performance.
[171]	Deep Neural Network	Combination of convolutional and recurrent NN.		417	F1-Score in between 0.8–0.9 for different activities.	Simulated sensor data demonstrates the feasibility of classifying athletic tasks using wearable sensors.
[172]	Deep Neural Network	Fully connected CNN.	50	5 (20 actions per person)	cross validated accuracy for action classification. (Camera only—85.3% IMU only 67.1%, Combined—86.9%).	Action recognition algorithm utilising both images and inertial sensor data that can efficiently extract feature vectors using a CNN and performs the classification using an RNN.
[173]	Hybrid DL model	Combines the simple recurrent units (SRUs) with the gated recurrent units (GRUs) of neural networks.	50	1007	Accuracy (99.8%)	Deep SRUs-GRUs networks to process the sequences of multisensors input data by using the capability of their internal memory states and exploit their speed advantage.
[174]	CNN	Akamatsu Transform		120	Accuracy (85%)	Proposed a human action recognition method using data acquired from wearable sensors and learned using a Neural Network.
[178]	SVM, ANN and HMM, and one compressed sensing algorithm, SRC-RP	DL using MATLAB.		4 people with 5 different tests	Recognition accuracy for different datasets (Debora—93.4%, Katia—99.6%, Wallace—95.6%).	Three different ML algorithms, such as SVM, HMM and ANN, and one compressed sensing-based algorithm, SRC-RP are implemented to recognise human body activities.
[179]	ML	Ensemble Empirical Mode Decomposition (EEMD), Sparse Multinomial Logistic Regression algorithm with Bayesian regularisation (SBMLR) and the Fuzzy Least Squares Support Vector Machine (FLS-SVM).		23	Classification accuracy (93.43%).	A novel approach based on the EEMD and FLS-SVM techniques is presented to recognise human activities. Demonstrated that the EEMD features can make significant contributions in improving classification accuracy.
[180]	ML	WEKA		30	Accuracy (98.5333%)	Sensors on a smartphone, including an accelerometer and a gyroscope were used to gather and log the wearable sensing data for human activities.
[151]	Real-time Gesture Pattern Classification	Neural network-based classifier model.		1040	Accuracy (77%)	Human hand gesture recognition using manually collected data and processed by LSTM layer structure. Accuracy is denoted using unity visualisation.
[181]	Pattern Recognition Methods for Head Gesture-Based Interface of a Virtual Reality Helmet (VRH) Equipped with a Single IMU Sensor	Classifier uses a two-stage PCA-based method, a feedforward artificial neural network, and random forest.		975 gestures from 12 patients	Classification rate (0.975)	VRH with sensors are used to collect data. Dynamic Time Warping (DTW) algorithm used for pattern recognition.
[182]	Hand Gesture Recognition (HGR) System.	Restricted Coulomb Energy (RCE) neural networks distance measurement scheme of DTW.		252	Accuracy (98.6%)	Hand Gesture Recognition (HAR) system for Human-Computer Interaction (HCI) based on time-dependent data from IMU sensors.
[183]	Motion capturing gloves are designed using 3D sensory data	Classification model with ANN.		6700	Accuracy (98%)	Data gloves with IMU sensors are used to capture finger and palm movements.
[184]	Quaternion-Based Gesture Recognition Using Wireless Wearable Motion Capture Sensors	SVM and ANN		11	Accuracy (90%)	Multisensor motion capturing system that is capable of identifying six hand and upper body movements.