Human Activity Recognition using Deep Learning: Past, Present and Future

Abstract

The study of human activities has always incite researchers. The human actions are acquired from raw time-series signals using smartphones and wearable devices' integrated sensors. Human Activity Recognition (HAR) has wide applications in rehabilitation centres, geriatric care houses, orphanages, and public places with large crowds. The essential stages of HAR architecture are feature detection, feature selection, and feature extraction. The majority of existing HAR algorithms implement conventional machine learning techniques for detection and classification. However, the desired computational results are not obtained for complex human activities. Thus, deep learning techniques have attracted attention of researchers. The deep learning techniques are widely used in various applications involving feature extraction, classification and detection. The deep learning techniques automatically extract features and reduces the computational complexity. This paper discusses various existing benchmark datasets to evaluate the performance of HAR models. A comprehensive review of sensor-based human activity recognition using convolution neural network-based models, long short-term memory-based models, and hybrid models is presented. Finally, the challenges and significance of human activity recognition is discussed.