原文
- Bashar M, Islam S, Hussain K K, et al. Multiple object tracking in recent times: A literature review[J]. arXiv preprint arXiv:2209.04796, 2022.
- https://arxiv.org/pdf/2209.04796
参考文章
- 多目标跟踪最新综述(基于Transformer/图模型/检测和关联/孪生网络)
文章目录
- 1 MOT MAIN CHALLENGES
- 2 MOT APPROACHES
- 3 MOT BENCHMARKS
- 4 MOT METRICS
- 5 APPLICATIONS
- 6 FUTURE DIRECTIONS
1 MOT MAIN CHALLENGES
(1)Occlusion
(2)Challenges for Lightweight Architecture
(3)Some Common Challenges
- inaccurate object detection
- motion blurring
- ID Switching
2 MOT APPROACHES
(1)Transformer
transformer专门处理序列信息,所以transformer可以完美地完成逐帧处理
Meinhardt T, Kirillov A, Leal-Taixe L, et al. Trackformer: Multi-object tracking with transformers[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022: 8844-8854.
(2)Graph Model
从连续帧中检测到的一组目标被视为一个节点,两个节点之间的链接被视为一个边缘。通常情况下,数据关联是通过应用匈牙利算法来完成的
Dai P, Weng R, Choi W, et al. Learning a proposal classifier for multiple object tracking[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021: 2443-2452.
(3)Detection and Target Association
(4)Attention Module
Wan J X, Zhang H, Zhang J, et al. Dsrrtracker: Dynamic search region refinement for attention-based siamese multi-object tracking[J]. arXiv preprint arXiv:2203.10729, 2022.
(5)Motion Model
(6)Siamese Network
Xing D, Evangeliou N, Tsoukalas A, et al. Siamese transformer pyramid networks for real-time UAV tracking[C]//Proceedings of the IEEE/CVF winter conference on applications of computer vision. 2022: 2139-2148.
(7)Tracklet Association
Yu E, Li Z, Han S. Towards discriminative representation: Multi-view trajectory contrastive learning for online multi-object tracking[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022: 8834-8843.
3 MOT BENCHMARKS
4 MOT METRICS
(1)MOTP
Multiple Object Tracking Precision
(2)MOTA
Multiple Object Tracking Accuracy
(3)IDF1
The Identification Metric
precision、recall、F1
(4)Track-mAP
(5)HOTA
Higher Order Tracking Accuracy
(6)LocA
Localization Accuracy
(7)AssA: Association Accuracy Score
(8) DetA: Detection Accuracy
(9) DetRe: Detection Recall
(10)DetPr: Detection Precision
(11) AssRe: Association Recall
(12) AssPr: Association Precision
(13) MOTSA: Multi Object Tracking and Segmentation Accuracy
(14)AMOTA: Average Multiple Object Tracking Precision
5 APPLICATIONS
(1)Autonomous Driving
(2)Pedestrian Tracking
(3)Vehicle Surveillance
(4) Sports Player Tracking
(5)Wild Life Tracking
6 FUTURE DIRECTIONS
(1)Multiple object tracking under multiple cameras
(2)Class-based tracking system can be integrated with multiple object tracking.
针对性
(3)3D tracking
(4) transformer can be used more specifically in solving different MOT tasks
(5)lightweight architecture
(6)real-time tracking
(7) Quantum computing can be used in MOT as well