基于局部特征关联性的足球追踪及其可视化研究  

Research on soccer ball tracking and visualization based on local feature correlation

作　　者：谷鹏 薛振宇 董培梅 胡森伟 GU Peng;XUE Zhenyu;DONG Peimei;HU Senwei(College of Information Engineering,China Jiliang University,Hangzhou 310018,China)

机构地区：[1]中国计量大学信息工程学院,浙江杭州310018

出　　处：《中国计量大学学报》2024年第4期655-663,691,共10页Journal of China University of Metrology

基　　金：浙江省自然科学基金项目(No.LQ23F010014)。

摘　　要：目的:提高运动视频中足球的追踪准确度,对其运动轨迹进行可视化。方法:提出一种基于局部特征关联性的算法。使用YOLOv5进行足球检测,在其采样过程中建立输入输出之间的局部特征映射,并保留通道维度内所有信息,提高对足球的检测精度。在卡尔曼滤波中使用双特征关联性匹配取代交并比,对足球轨迹进行校正。利用尺度不变特征变换(scale-invariant feature transform,SIFT)对视频相邻帧进行特征点抓取,使用特征点匹配计算单应性矩阵,实现场地拼接及足球位置映射。结果:在SoccerNet-v2数据集上,对足球检测的AP为93.11%,检测精度比优化前提高3.27%,对足球追踪的精度为80.33%。结论:算法能够对足球进行准确追踪,可对其轨迹及关键帧所发生的事件进行展示,拥有良好的可视化效果。Aims:This paper aims to study the enhancement of the tracking accuracy of a soccer ball in motion videos and visualize its complete trajectory accurately.Methods:The algorithm leverages YOLOv5 was proposed to detect and establish mappings of local features between input and output graphs.The information across channel dimensions during the sampling phase was preserved to improve the detection accuracy of soccer balls.The dual feature correlation matching was used instead of intersection and union ratio in Kalman filtering to correct football trajectories.By utilizing scale invariant feature transformation(SIFT),the feature points from adjacent frames of the video were captured.And by using feature point matching,the homography matrix was calculated to splice and map the soccer ball position.Results:On the SoccerNetV2 dataset,the AP for football detection was 93.11%.Compared to the pre-optimized model,the detection accuracy increased by 3.27%.The tracking accuracy for football reached 80.33%.Conclusions:The model can accurately track the soccer ball and display its trajectory and key events in each frame with superior visualization effects.

关 键 词：足球追踪 卡尔曼滤波 图片拼接 足球检测 

分 类 号：TB391.41[一般工业技术—材料科学与工程]