Comprehensive machine and deep learning analysis of sensor-based human activity recognition

Human Activity Recognition (HAR) is a crucial research focus in the body area networks and pervasive computing domains. The goal of HAR is to examine activities from raw sensor data, video sequences, or even images. It aims to classify input data correctly into its underlying category. In the current study, machine and deep learning approaches along with different traditional dimensionality reduction and TDA feature extraction techniques are suggested to solve the HAR problem. Two public datasets (i.e., WISDM and UCI-HAR) are used to conduct the experiments. Different data balancing techniques are utilized to deal with the problem of imbalanced data. Additionally, a sampling mechanism with two overlapping percentages (i.e., 0% and 50%) is applied to each dataset to retrieve four balanced datasets. Five traditional dimensionality reduction techniques in addition to the Topological Data Analysis (TDA) are utilized. Seven machine learning (ML) algorithms are used to perform HAR where six of them are ensemble classifiers. In addition to that, 1D-CNN, BiLSTM, and GRU deep learning approaches are utilized. Three categories of experiments (i.e., ML with traditional features, ML with TDA, and DL) are applied. For the first category experiments, the best-reported scores concerning the WISDM dataset are accuracy and WSM of 99.10% and 86.61%, respectively. When concerning the UCI-HAR dataset, the best-reported scores are accuracy and WSM of 100% and 100%, respectively. For the second category experiments, the best-reported scores concerning the WISDM dataset are accuracy and WSM of 95.34% and 89.62%, respectively. When concerning the UCI-HAR dataset, the best-reported scores are accuracy and WSM of 96.70% and 92.57%, respectively. For the third category experiments, the best-reported scores concerning the WISDM dataset are accuracy and WSM of 99.90% and 99.76%, respectively. When concerning the UCI-HAR dataset, the best-reported scores are accuracy and WSM of 100% and 100%, respectively. After concluding the final results, the suggested approach is compared with 6 related studies utilizing the same dataset(s).