ICA-Enhanced YOLOv5-AdaBoost Framework for Player Localization in Semi-Automatic Offside Detection

Abstract

As the rules of football matches become more and more strict, the traditional position information capture technology cannot accurately carry out the problem of athletes' position information capture. To address this problem, this study proposes an adaptive sample size method based on dynamic sample weight optimization. This method focuses on improving the AdaBoost algorithm's sample weighting mechanism and combining YOLOv5's object detection capability with the Empire Competition algorithm's global optimization characteristics to create an athlete position information capture platform. The experimental results showed that in the campus football game dataset, the average absolute error value of the proposed algorithm was 0.086, and the root mean square error was 0.049, which was 0.211 and 0.119 lower than YOLOv4, respectively. Under 100 sets of experimental datasets, the average accuracy of the proposed algorithm reached 96.18%, which is 5.93% higher than the YOLOX Nano algorithm. In the SoccerReplay dataset, the capture platform designed by the research had an occupancy rate of 5.139% and a packet loss rate of 2.367%. These rates were reduced by 19.753% and 35.06%, respectively, compared to YOLOv5s. The above results show that the study of the mention capture technique can capture the positional information of the athletes more accurately and with higher capture accuracy in football SEMI-automatic offside detection.