基于Transformer和位置约束的端到端多目标追踪算法  

End-to-end multi-object tracking with location constraintusing Transformer

作　　者：吴悦 雒江涛 张攀 任媛 WU Yue;LUO Jiangtao;ZHANG Pan;REN Yuan(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China;Electronic Information and Networking Research Institute,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]重庆邮电大学电子信息与网络工程研究院,重庆400065

出　　处：《重庆邮电大学学报（自然科学版）》2023年第3期563-570,共8页Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)

摘　　要：为了应对计算机视觉中多目标追踪任务的挑战。针对网络中使用的锚点造成重识别训练模糊的问题,通过可生成参考点的可变形Transformer,提出了一个端到端的多目标追踪系统(tracker Transformer,TKTR)。由骨干网络模块提取特征图,将其送入可生成参考点的可变形Transformer架构,以检测目标并生成代表检测框中心的参考点;利用参考点对候选目标进行空间位置约束并计算检测框的交并比来关联目标。实验结果表明,TKTR利用Transformer的查询特征向量对目标进行位置约束,提高了追踪精度,并且降低了ID切换的指标。To address the challenges of multi-object tracking in computer vision,we propose a novel end-to-end system called TKTR.The system addresses the issue of blurred re-identification training caused by the anchors used in the network.It employs a deformable Transformer that can generate reference points to detect and track multiple objects.Features are extracted from a backbone network module,and are passed through the Transformer architecture to detect the objects and generate reference points representing the center of the detected bounding boxes.The reference points are then used to impose spatial constraints on candidate objects and compute the intersection-over-union to associate the objects.Experimental results have shown that TKTR achieves competitive results in the multi-object tracking accuracy(MOTA)metric compared to other state-of-the-art methods,with reduction in the ID switching metric compared to models that directly represent appearance features using target query feature vectors.

关 键 词：多目标追踪 TRANSFORMER 位置约束 端到端 

分 类 号：TN929[电子电信—通信与信息系统]