摘要
足球视频中的跨镜头目标跟踪技术对于实现整场比赛的全局连续分析具有重要意义,然而,当前算法大多局限于单镜头跟踪,难以应对足球比赛中频繁的镜头切换,导致跟踪中断和对象失配,无法满足全局数据分析的需求.针对这一关键问题,本文提出了一种跨镜头全局跟踪算法,通过单镜头跟踪增强和跨镜头关联匹配两方面的创新优化,显著提升了足球视频分析的完整性和准确性.在方法层面,首先设计了基于球员状态的关系交互模型,有效解决了运动模糊、外观相似和遮挡等问题,设计了基于重叠窗口的层级关联策略,解决逐帧关联容易轨迹碎片化的足球跟踪难题;其次创新构建了融合球员位置和姿态的长跨度关联匹配方法,提出边界轨迹衔接后处理方法,实现跨镜头精准匹配.实验方面,构建了全新的足球视频跨镜头跟踪数据集,所提方法在跨镜头关联上取得86.7%的准确率和 97.0%的召回率,全局跟踪方面在 HOTA 指标上提升 14.0%,此外,所提方法在 SoccerNet-Tracking 和 SportsMOT 公开数据集上达到最优性能.研究成果为体育智能化发展提供了重要的技术支撑.
Cross-shot object tracking is essential for enabling comprehensive match analysis in soccer video sequences.While current tracking methods predominantly operate within individual camera shots,they exhibit significant limitations when confronted with the dynamic multi-shot nature of broadcast soccer footage.Frequent shot transitions often result in tracking discontinui-ties and identity mismatches,ultimately compromising the integrity of global match analytics.To address this key challenge,this paper proposes a cross-shot global tracking algorithm,featu-ring innovative optimizations in both single-shot tracking enhancement and cross-shot object asso-ciation,significantly improving the completeness and accuracy of soccer video analysis.Methodo-logically,the study first introduces a player-state-based relational interaction model to effectively handle motion blur,appearance similarity,and occlusion,and introduces a hierarchical associa-tion strategy with overlapping time windows to mitigate trajectory fragmentation in frame-by-frame tracking.Second,we develop a long-range cross-shot association method that fuses player positional and pose features,along with a boundary trajectory stitching post-processing step,to achieve precise cross-shot matching.For experimentation,a novel cross-shot soccer video track-ing dataset was constructed.The proposed method achieves 86.7%accuracy and 97.0%recall in the cross-shot association,while improving the HOTA metric by 14.0%in global tracking.It al-so demonstrates state-of-the-art performance on public datasets such as SoccerNet-Tracking and SportsMOT.These research outcomes provide critical technical support for the advancement of intelligent sports analytics.