Scholarlink

<p>With the increasing penetration of computer vision technology into the field of sports, intelligent video analysis has gradually become an important means to improve the understanding of the game, and the tracking and discrimination of football players has become a core task in the sports data analysis system. However, existing target detection and multi-target tracking methods still show significant limitations in scenarios such as small target recognition, dense occlusion, and high-speed motion. Therefore, this study proposes a football player tracking and discrimination method that integrates an improved You Only Look Once version 5 (YOLOv5) and a depth-distance intersection over union (DIoU) strategy. The method improves detection accuracy through backbone network simplification, multi-scale output reconstruction, and attention enhancement, and strengthens motion prediction and cross-frame correlation stability by combining unscented Kalman filter (UKF) and DIoU matching mechanism. Experimental results show that the proposed method achieves an average precision of 90.56% ± 0.25 and a recall of 87.32% ± 0.37 at the detection end, with an average precision improvement of nearly 10% points for small targets. The tracking method achieves a multi-target tracking accuracy of 76.84% ± 0.38 and an identity retention F1 score of 78.56% ± 0.39, with a 35.4% reduction in identity information switching frequency. The end-to-end team classification accuracy reaches 91.46% ± 0.28, with a 51.6% decrease in incorrect team classification rate, maintaining optimal performance even in high-speed and dense scenarios. This demonstrates that the research method combines high-precision detection with robust tracking capabilities, effectively improving the level of intelligent processing of football player behavior and matches.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Football player tracking and discrimination by integrating improved YOLOv5 and deep DIoU strategy

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月3日

查看全文
  • Chunyan Na

摘要

With the increasing penetration of computer vision technology into the field of sports, intelligent video analysis has gradually become an important means to improve the understanding of the game, and the tracking and discrimination of football players has become a core task in the sports data analysis system. However, existing target detection and multi-target tracking methods still show significant limitations in scenarios such as small target recognition, dense occlusion, and high-speed motion. Therefore, this study proposes a football player tracking and discrimination method that integrates an improved You Only Look Once version 5 (YOLOv5) and a depth-distance intersection over union (DIoU) strategy. The method improves detection accuracy through backbone network simplification, multi-scale output reconstruction, and attention enhancement, and strengthens motion prediction and cross-frame correlation stability by combining unscented Kalman filter (UKF) and DIoU matching mechanism. Experimental results show that the proposed method achieves an average precision of 90.56% ± 0.25 and a recall of 87.32% ± 0.37 at the detection end, with an average precision improvement of nearly 10% points for small targets. The tracking method achieves a multi-target tracking accuracy of 76.84% ± 0.38 and an identity retention F1 score of 78.56% ± 0.39, with a 35.4% reduction in identity information switching frequency. The end-to-end team classification accuracy reaches 91.46% ± 0.28, with a 51.6% decrease in incorrect team classification rate, maintaining optimal performance even in high-speed and dense scenarios. This demonstrates that the research method combines high-precision detection with robust tracking capabilities, effectively improving the level of intelligent processing of football player behavior and matches.