A Comprehensive Review on Human Activity Recognition from Conventional Techniques to Deep Learning

Abstract

Human Activity Recognition (HAR) is a dynamic research area in computer vision because of its applications, for example, video surveillance, human-computer interaction, and healthcare.  Distinguishing human activities from video comprises one of the most challenging issues, mainly because of the nature of the real-world activities and vast volumes of information. The strategies present in the literature for HAR are extensively grouped into two classes: single-layered (for simple activities) and hierarchical approaches (for complex movements). Deep learning is one of the hierarchical strategies to perceive the actions of human beings and providing better results by extracting the features by automatic knowledge from the input data. Yoga has become a notable discipline around the globe that keep individuals in excellent physical and mental health. The specialists of Yoga are more, and most men and aged persons perform Yoga for more time, and the classification of these activities from the videos is essential. Smart IoT based yoga frameworks required for the individuals who need to rehearse Yoga at home. There exist methods such as RGB/Kinect camera-dependent, or sensor-based or gadget we can wear relies on Yoga pose identification approaches with reasonable accuracy with a bounded dataset with fewer parameters. These methods provide poor results for real-time streaming data, and still, we can improve the efficiency with advanced deep learning architectures. Furthermore, the present techniques focused on a limited number of yoga poses, yet to be handled with fine-grained classification by creating a vast benchmark dataset. As per the best of our knowledge, the comprehensive review of HAR strategies on yoga pose recognition also missed in the literature. Hence, this paper provides a summary of different approaches, difficulties, and issues of HAR systems.