带有深度特征重平衡网络的多目标跟踪方法

Multiple object tracking method with deep feature rebalancing network

针对基于联合检测嵌入范式的多目标跟踪方法中,检测与re-ID任务间冲突导致系统性能劣化的问题,首先设计一种用于富集多层语义信息并能针对不同分支倾向重构特征图的网络,有效缓解检测与re-ID分支在优化中对特征信息需求的恶性竞争;其次采用一种强化的关联策略,该策略将检测信息更深入地引入到关联流程中,旨在为更多检测结果提供关联机会,同时抑制环境干扰在关联中带来的长期损害,有效降低关联过程中误关联和漏关联的发生.实验结果表明,所提出的方法相对于当前的先进方法展现了强大的潜力,在MOT17测试集上取得了75.7%MOTA、73.4%IDF1及60.0%HOTA的性能.

Aiming at the decreasing of system performance caused by conflicts between detection and re-ID branches in the multiple object tracking method based on the joint detection and embedding paradigm,we design a network for concentrating multi-level semantic information and constructing differentiated feature maps for different branch.This network effectively relieves the vicious competition between detection and re-ID branches for feature information demands.Secondly,a modified association strategy is adopted,which introduces further information from detection branch into the association process.This strategy provides more opportunities for detections to associate while suppressing the long-term damage caused by environmental noise,effectively reducing the occurrence of misassociation and missed association.The experiments show that the method in this paper has great potential compared with the current advanced methods,and achieves the performance of 75.7%MOTA,73.4%IDF1 and 60.0%HOTA on the MOT17 test set.