高级检索

基于姿态特征融合的短道速滑运动员多目标追踪

Pose Feature Fusion Based Short Track Speed Skaters Multi Objective Tracking

  • 摘要: 多目标追踪技术对为短道速滑比赛进行数据分析, 运动员提供辅助技术支撑具有重要意义. 短道速滑运动员滑行过程中存在尺度变化大, 频繁遮挡, 运动模糊以及外观相似等复杂情况, 使得追踪过程面临更为严峻的挑战. 为此, 构建一个短道速滑场景下速滑运动员比赛的多目标追踪数据集SSSMOT, 并基于姿态信息提出了一个新颖的多目标追踪方法. 该方法首先从Anchor、损失函数、NMS等方面优化了Yolov5检测模型, 提高检测准确度; 其次, 设计全新的特征提取网络P-RNet, 依据姿态信息的指引针对性地提取特征, 提高特征鲁棒性; 最后, 使用姿态关键点改进数据关联匹配方法, 一定程度上缓解因运动员外观相似导致匹配错误的问题. 本文在SSSMOT数据集以及SKMOT数据集进行实验, 大量的实验表明所提方法的有效性以及优越性.

     

    Abstract: Multi-objective tracking technology is of great significance for data analysis for short track speed skating competitions and for athletes to provide auxiliary technical support. The tracking process is more challenging due to the large scale changes, frequent occlusions, motion blurring and similar appearance of the skaters during the skating. To this end, a multi-objective tracking dataset SSSMOT is constructed for speed skaters in the short track speed skating scene, and a novel multi-objective tracking method is proposed based on pose information. The method firstly optimizes the Yolov5 detection model in terms of Anchor, loss function, and NMS to improve the detection accuracy; secondly, a new feature extraction network P-RNet is designed to extract features based on the pose information to improve the feature robustness; finally, the data association matching method is improved by using the pose key points to alleviate the matching error caused by the similar appearance of athletes to a certain extent. Finally, Using pose key points to improve the data association matching method can alleviate the problem of matching errors caused by similar appearance of athletes to a certain extent. In this paper, experiments are conducted on SSSMOT and SKMOT datasets, and the effectiveness and superiority of the proposed method are demonstrated in a large number of experiments.

     

/

返回文章
返回