Segmentation and Quality Assessment of Continuous Fitness Movements Based on Vision

Zeying Li, Hongtao Chen, Jing Cai, Yanbing Xue

Abstract: With the growing popularity of a healthy lifestyle, fitness activities have garnered widespread attention. However, traditional fitness approaches face challenges such as constraints of time and location, high costs, and varying quality of instructors, with the incorrect execution of movements potentially leading to physical injuries. To address these issues, this study proposes a framework based on vision for the segmentation and quality assessment of continuous fitness movements, designed to provide technical assistance in daily fitness routines to optimize and correct exercise methods. This framework utilizes video recordings, analyzing the three-dimensional positions of skeletal points to transform them into angle features, and then employs deep learning methods for segmenting movements and assessing their quality, significantly lowering the technical threshold for application. Our core technologies include: a movement conversion module, action segmentation based on wavelet transform technology, and a quality scoring system through deep neural networks that integrates time pyramids, Transformer, and MLP networks. Furthermore, to validate the effectiveness of our framework, we have also established a dataset of squat movements with expert ratings. The aim of this research is to provide an efficient and user-friendly fitness assessment tool to reduce the health risks associated with improper exercise execution and promote greater efficiency and safety in fitness routines.