Real-Time Posture Correction in Gym Exercises: A Computer Vision-Based Approach for Performance Analysis, Error Classification and FeedbackHitesh Kotte; Milos Kravcik; Nghia Duong-Trung
In: Khaleel Asyraaf Mat Sanusi; Bibeg Limbu; Jan Schneider; Milos Kravcik; Roland Klemke (Hrsg.). Proceedings of the Third International Workshop on Multimodal Immersive Learning Systems (MILeS 2023). International Workshop on Multimodal Immersive Learning Systems (MILeS-2023), located at Eighteenth European Conference on Technology Enhanced Learning (EC-TEL 2023), September 4-8, Aveiro, Portugal, Pages 64-70, Vol. 3499, CEUR Workshop Proceedings, Aachen, 10/2023.
Developing psychomotor skills is a major challenge, but modern technology offers novel solutions that can provide valuable support to trainees. This paper proposes a pioneering strategy that uses computer vision methods to monitor performance and provide real-time feedback on posture during fitness exercises, allowing for instant self-correction and motivation even without professional guidance. Our system utilizes a versatile learning framework to analyze live expert demonstrations or recorded video content. We leverage the deep learning YOLOv7-pose model to identify human keypoints and combine it with a human topology-oriented tracking procedure. Our system delivers immediate feedback to rectify posture by collecting comprehensive tracking data. Notably, we capitalize on transfer learning techniques to avoid extensive model retraining. To demonstrate the usefulness of our method, we benchmarked it to professional fitness videos and evaluated it with five inexperienced participants. The results showed a positive reaction from the participants, suggesting improvements to the user interface.