足球场内球员多目标跟踪算法  

Multi-object tracking algorithm for players on football field

作　　者：张俊杰[1] 李博正 曾丹[1] ZHANG Jun-jie;LI Bo-zheng;ZENG Dan(School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China)

机构地区：[1]上海大学通信与信息工程学院,上海200444

出　　处：《计算机工程与设计》2023年第7期2148-2155,共8页Computer Engineering and Design

摘　　要：为解决足球比赛场内球员的多目标跟踪任务中,因场外人员对跟踪的干扰,球员频繁地运动、互相遮挡,以及摄像镜头复杂地移动等情况,造成的跟踪准确度低、跟踪目标ID(identity)保持能力弱的问题,提出一种适用于足球场内球员跟踪的多目标跟踪数据集和多目标跟踪算法。通过条件生成对抗网络分割出球场区域,筛选出球场内的基于YOLOX框架的目标检测结果;在数据关联阶段,设计一种融合IoU(intersection over union)与欧式距离的代价矩阵进行目标间的相似性度量;利用足球比赛上场人数存在上限的先验条件,弹性约束跟踪目标ID的增长。实验结果表明,针对足球场内球员的跟踪问题,该算法能够在多目标跟踪准确度、跟踪目标ID保持能力上有极大提高。Considering the interference of off-field personnel,frequent movements,the occlusion among players,as well as the complex movement of cameras,the performance of state-of-the-art MOT(multi-object tracking)model for players on the football field is somewhat compromised,resulting in low tracking accuracy and weak identity retention ability.To address above issues,a MOT framework was proposed with a newly collected dataset targeted at football matches.More specifically,the football field was segmented using conditional generative adversarial network to filter the YOLOX-based detection results.To associate these filtered results,a cost matrix combining IoU(intersection over union)and Euclidean distance was designed to measure the similarity among objects.Moreover,since the total number of players appeared on the field is often limited,the growth of trac-king object ID is elastically constrained by this prior.Extensive experimental results demonstrate that the proposed model signi-ficantly improves the tracking accuracy as well as the tracking object ID retention ability.

关 键 词：足球比赛 多目标跟踪 目标检测 生成对抗网络 跟踪目标身份约束 数据关联 代价矩阵 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]