Realtime Indoor Workout Analysis Using Machine Learning & Computer Vision

These days there are thousands of workout videos available on the internet. Samsung Health
²
[1] provides a dedicated section called programs containing short workout videos for various exercises. The goal is to assist people perform these workouts independently on their own. A common observation is that even people who visit gym regularly find it difficult to perform all steps (body pose alignments) in a workout accurately. Continuously doing an exercise incorrectly may eventually cause severe long term injuries. To help solve this problem and provide assistance in form of a visual feedback while performing a workout, we propose a system to analyze user's body posture during a workout and compare it to a professional's reference workout. We represent human body as a collection of limbs and analyze angle between limb pairs to detect errors and provide corrective action to the user. Our system builds on the latest advancements using deep learning for human body pose estimation. We use techniques for time series data alignment like DTW [2] (Dynamic Time Warping) along with optical flow tracking to synchronize user/reference videos. We are able to detect and locate errors in user's activity (pose) very effectively based on some threshold deviation between the limb angles. The system in future can be extended to be used by physicians to monitor patient's recovery following an injury.