常用框架有以下三种:
Separate Detection and Embedding (SDE- 物体检测,特征提取与物体关联),JOINT Detection and Embedding (JDE) (a)Deep sort :Simple Online and Realtime Tracking with a Deep Association Metric Nicolai Wojke, Alex Bewley, Dietrich Paulus. ICIP2017. cites:2523. https://github.com/nwojke/deep_sort (b) Tracktor: Tracking without bells and whistles. Philipp Bergmann, Tim Meinhardt, Laura Leal-Taixe. ICCV2019, cites:721. https://github.com/phil-bergmann/tracking_wo_bnw
Tracktor经过用前帧目标的bbox做为当前帧的初始bbox来回归,减小了FP,且省略了data association步骤,但是这里用的检测方法是二阶段的Faster-Rcnn,实时性?实验没有给出FPS数据,为了解决遮挡问题引起的IDswitch,因此作者在模型中加入ReID,即Tracktor++,这样就解决了遮挡问题,以及加入运动模型来解决计算可能带来的帧率降低所导致的先后两帧结果差距较大。 (c) RetinaTrack: Online Single Stage Joint Detection and Tracking. CVPR2020. Zhichao Lu, Vivek Rathod, Cited by 144. Ronny Votel, Jonathan Huang. https://github.com/Hanson0910/RetinaTrack
- 通过将分类、回归和特征提取设为三个分支任务,除了FPN之前的部分,三者的特征共享部分含有m1个3x3卷积;
- 对于每层特征图上每个特征点的k个anchor,全部预测分类、回归和特征,增加区分度。对于检测任务,分类和回归分支都包含m2个3x3卷积,而embedding分支则为m3个1x1卷积。
(c)Towards Real-Time Multi-Object Tracking. Zhongdao Wang, Liang Zheng, Yixuan Liu, Yali Li, Shengjin Wang. Cited by 556. 2020 ECCV. https://github.com/Zhongdao/Towards-Realtime-MOT
这里的做法是:将每个track ID当作一个类别,embedding阶段就是预测trackid类别,这种做法在现实使用时不太可行。 (c)Track to Detect and Segment: An Online Multi-Object Tracker. Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, Junsong Yuan. cvpr 2021. Cited by 160.
backbone基于CenterNet。后处理方法类似:CenterTrack,就是求两帧之间 实例 相似度的一系列过程 (c)Centtrack: Tracking Objects as Points. Xingyi Zhou, Vladlen Koltun, Philipp Krähenbühl. Cited by 634 https://github.com/xingyizhou/CenterTrack.git
标签:github,Cited,tracking,Tracking,https,Tracktor,com,调研 From: https://www.cnblogs.com/jianyingzhou/p/17320487.html