Research And Application Of Key Techniques Of Sports Evaluation Based On Human Body Posture Estimation

In recent years,with the rapid development of computer vision applications,the use of computer vision for online sports competitions and exercise to improve the quality of physical education has received widespread attention.However,current AI-based motion recognition frameworks focus on classifying different movements during sports,while lacking related research and practical applications for further operations such as counting,filtering improper movements,and providing suggestions.This thesis aims to explore a posture estimation-based motion detection framework and counting algorithm,and apply it to the field of sports motion detection and recognition.To address the above problems,this thesis first proposes a posture estimation-based motion detection framework and counting algorithm,which utilizes deep learning technology to achieve fine-grained segmentation and recognition of sports movements,thereby improving the recognition accuracy of specific sports movements.Secondly,this thesis designs an end-cloud combined motion detection system for actual application scenarios in university physical fitness tests,implements relevant algorithms,and figures out counting and filtering of sports movements in scenarios such as sit-ups and pull-ups.Finally,by optimizing network models and inference acceleration methods,this thesis improves the end-cloud combined system and successfully achieves cloud-based network training and deployment of lightweight models for real-time inference tasks on edge devices,further improving the efficiency of sports motion detection and counting tasks.The research results of this thesis are expected to provide technical support and reference for online sports competitions and exercise scenarios.