Identifying Competition Characteristics of Athletes Through Video Analysis
Yuzhong Liu, Tianfan Zhang, Zhe Li, Mengshuang Ma · International Journal of Advanced Computer Science and Applications · 2024
The vast repositories of training and competition video data serve as indispensable resources for athlete training and competitor analysis, providing a solid foundation for strategic competition analysis and tactics formulation. However, the effectiveness of these analyses hinges on the abundance and precision of data, often requiring costly professional systems for existing video analysis techniques. Meanwhile, readily accessible non-professional data frequently lacks standardization, compelling manual analysis and experiential judgments, thus limiting the widespread adoption of video analysis technologies. To address these challenges, we have devised an intelligent video analysis technology and a methodology for identifying athletes' competition characteristics. Initially, we employed target detection models, such as You Only Look Once (YOLO), renowned for their ease of deployment and low environmental dependency, to perform fundamental detection tasks. This was further complemented by the intelligent selection of standardized scenes through customizable scene rules, leading to the formation of a standardized scene dataset. On this robust foundation, we achieved classification and identification of competition participants as well as sideline recognition, ultimately compiling a comprehensive competitive dataset. Subsequently, we constructed an athlete posture estimation method utilizing OpenPose, aimed at minimizing interference caused by obstructions and enhancing the accuracy of feature extraction. In experimental validation, we gathered a diverse collection of table tennis competition video data from the internet, serving as a validation dataset. The results were impressive, with a detection success rate for standardized scenes exceeding 94% and an identification success rate for competitors surpassing 98%. The accuracy of posture reconstruction for obstructed individuals exceeded 60%, and the effectiveness of identifying athletes' main features exceeded 90%, convincingly demonstrating the effectiveness of the proposed video analysis method.