Video based exercise recognition and correct pose detection

Abstract

Human pose estimation has gained significant attention from researchers of the present era. Personal exercise sessions can be monitored and supervised with the help of pose recognition. Existing work on exercise classification primarily relies on external or wearable sensors for recognizing poses. However, such sensors often fail to differentiate amongst similar exercises. Some essential extensions of human pose estimation are activity detection and activity prediction. In this paper, we first classify an individual�s exercises and then predict whether the pose corresponding to an exercise is correct or not. The tasks mentioned above are performed with the help of 2-dimensional pose coordinates. We have used an RGB camera to capture the poses during exercises performed by individuals. We formulate our model with 2D coordinates obtained from the 2D pose. We consider 2D coordinates of 18 joints of a human body as the primary features to classify different exercises and predict correctness about the poses. We have developed a benchmark dataset consisting of human subjects of various age groups with varying heights. An accuracy of 97.01% has been obtained, and it is better than existing work when tested on our dataset. � 2022, The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature.