Physical Exercise Classification from Body Keypoints Using Machine Learning Techniques

In today’s world, the importance of physical workout and exercise is being realized by the human community. However, there are a lot of challenges faced by people, in terms of convenience and cost, as they need a trainer to monitor their workouts. This research work aims to alleviate this challenge and proposes a methodology that automates the physical workout monitoring process. The proposed framework can take live video streams and identify the type of exercises performed by a person using machine learning techniques. This task is achieved by identifying keypoints from the human pose caught in the live video and classifies it to a physical exercise. The methodology also offers utilities like counter and feedback for physical exercises classified. A comparative evaluation of different machine learning techniques is experimented for physical exercise classification. It is observed from the evaluation that the Random Forest technique gives significant values in terms of precision 89% and recall 89%, as well as in F1-score with 89%.