基于轨迹掩膜的在线多目标跟踪方法  

Online multi-object tracking method based on trajectory mask

作　　者：余明骏 刁红军[1] 凌兴宏[1,2,3] YU Mingjun;DIAO Hongjun;LING Xinghong(School of Computer Science and Technology,Soochow University,Suzhou 215006,Jiangsu,China;School of Computational Science and Artificial Intelligence,Suzhou City University,Suzhou 215104,Jiangsu,China;Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education,Jilin University,Changchun 130012,Jilin,China)

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006 [2]苏州城市学院计算科学与人工智能学院,江苏苏州215104 [3]吉林大学符号计算与知识工程教育部重点实验室,吉林长春130012

出　　处：《山东大学学报（工学版）》2023年第2期61-69,共9页Journal of Shandong University（Engineering Science）

基　　金：符号计算与知识工程教育部重点实验室(吉林大学)开放课题项目(93K172021K08);江苏高校优势学科建设工程资助项目(PAPD)。

摘　　要：针对现有多目标跟踪方法易受到遮挡、运动模糊等问题干扰的情况,提出基于轨迹掩膜的在线多目标跟踪方法(online multi-object tracking method based on trajectory mask,OMTMTM)。提出轨迹掩膜生成算法,利用前一帧跟踪轨迹结果生成轨迹掩膜,设计轨迹掩膜网络对轨迹掩膜提取多维度特征,包含目标可见区域的估计值、大致位置及形状等信息;将该特征与基础骨干网络提取的原始图像特征融合后进行多目标检测跟踪。OMTMTM的目标跟踪器具备先验判断能力,可实现遮挡情况下的准确跟踪;OMTMTM利用目标跟踪轨迹的时空信息,恢复出部分漏检或低置信待检目标,使轨迹掩膜更加合理,有利于后续跟踪。对OMTMTM的性能进行多维度评估,并结合基线模型进行对比分析。试验结果表明,OMTMTM具有先进的多目标跟踪性能。To address the situation that existing multi-object tracking methods were easily disturbed by problems such as occlusion and motion blur,an online multi-object tracking method based on trajectory mask was proposed,called OMTMTM.A trajectory mask generation algorithm was proposed,which used the previous frame tracking results to generate trajectory mask,and designed a trajectory mask network to extract multi-dimensional features,including the estimated value,approximate location and shape of the object visible area.The features were fused with the original image features extracted by the basic backbone network for multi-object detection and tracking.The object tracker of the OMTMTM had a priori judgment ability and could achieve accurate tracking in the case of occlusion.The OMTMTM used the spatiotemporal information of the object tracking trajectory to recover some missed or low-confidence objects.It made the trajectory mask more reasonable and was conducive to subsequent tracking.The performance of the OMTMTM was evaluated in multiple dimensions through comparing with the baseline models.The experimental results showed that the OMTMTM could achieve advanced multi-object tracking performance.

关 键 词：多目标跟踪 目标检测 轨迹掩膜 视频流 遮挡 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]