Scholarlink

<p>To address the problem that existing motion detection models cannot effectively integrate athlete posture and visual information when processing high-speed, highly competitive handball match videos, this paper proposes a multimodal motion detection model incorporating bidirectional cross attention. First, an improved Transformer feature extraction network is constructed, the core of which lies in the hybrid encoder fusing local and global information, and capturing motion features of different ranges through a multi-scale feed-forward network. Simultaneously, an interaction module based on focused bidirectional cross attention is designed to establish a dynamic interaction path between visual and posture modalities, achieving deep information complementarity. Experimental outcomes demonstrate that the improved Transformer network achieves an average F1 score and accuracy of 85.0% and 86.0% in five tests, with standard deviations of 0.16 and 0.14, respectively. After iteration, the mean accuracy @ intersection–union threshold of the improved Transformer network is approximately 53%, which is 6 percentage points higher than the standard Transformer. The mean average precision of the final model at @0.5 performance fluctuation center remained stable around 88.0%, within the range of 84.8% to 91.3%. When the input was 32 frames, its F1 score and accuracy reached 82.47 and 87.35% respectively. This model can effectively improve the accuracy and robustness of motion detection in complex sports scenarios. The research provides reliable technical support for real time tactical analysis of highly competitive sports.</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

Detecting real time motion in a handball game using a bidirectional cross attention mechanism and a transformer

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月17日

查看全文
  • Lu Yan,
  • Yan Si

摘要

To address the problem that existing motion detection models cannot effectively integrate athlete posture and visual information when processing high-speed, highly competitive handball match videos, this paper proposes a multimodal motion detection model incorporating bidirectional cross attention. First, an improved Transformer feature extraction network is constructed, the core of which lies in the hybrid encoder fusing local and global information, and capturing motion features of different ranges through a multi-scale feed-forward network. Simultaneously, an interaction module based on focused bidirectional cross attention is designed to establish a dynamic interaction path between visual and posture modalities, achieving deep information complementarity. Experimental outcomes demonstrate that the improved Transformer network achieves an average F1 score and accuracy of 85.0% and 86.0% in five tests, with standard deviations of 0.16 and 0.14, respectively. After iteration, the mean accuracy @ intersection–union threshold of the improved Transformer network is approximately 53%, which is 6 percentage points higher than the standard Transformer. The mean average precision of the final model at @0.5 performance fluctuation center remained stable around 88.0%, within the range of 84.8% to 91.3%. When the input was 32 frames, its F1 score and accuracy reached 82.47 and 87.35% respectively. This model can effectively improve the accuracy and robustness of motion detection in complex sports scenarios. The research provides reliable technical support for real time tactical analysis of highly competitive sports.