Research Article
EmbeddingwithBounding Box Contracting for Multi-object Tracking
@INPROCEEDINGS{10.1007/978-3-031-36011-4_6, author={Like Zhang and Wenjing Kang and Guangdong Zhang}, title={EmbeddingwithBounding Box Contracting for Multi-object Tracking}, proceedings={6GN for Future Wireless Networks. 5th EAI International Conference, 6GN 2022, Harbin, China, December 17-18, 2022, Proceedings, Part I}, proceedings_a={6GN}, year={2023}, month={7}, keywords={Multi-Object Tracking Object Detecting Embedding Methods}, doi={10.1007/978-3-031-36011-4_6} }- Like Zhang
Wenjing Kang
Guangdong Zhang
Year: 2023
EmbeddingwithBounding Box Contracting for Multi-object Tracking
6GN
Springer
DOI: 10.1007/978-3-031-36011-4_6
Abstract
The development of 5G/6G network can achieve high data transmission speed, which promotes the wide application of remote video monitoring. Multi-object tracking (MOT) aims at detecting and tracking all the objects of interesting categories in videos. Appearance and motion information of each object are significant clues utilized for finding associations between detections and tracks. Many approaches model each object appearance through bounding box region, which is vulnerable to background noise and motion deformation. In this paper, we alleviate this problem, via embedding with object bounding box contracting. We also integrate an online tracking by detection model, comprehensive use of appearance and motion information for data association. Object bounding box contracting is introduced to relieve the impact of interference and obtain high-quality re-ID embeddings. Experimental results based on the MOT17 benchmark show that the integrated tracker with bounding box contracting for embedding achieves 80.6 MOTA, 79.4 IDF1 and 64.4 HOTA.
