基于Transformer的跨空间特征关联多目标追踪

Multi-Object Tracking with Cross-Spatial Feature Association Based on Transformer

多目标追踪算法在复杂场景中常会出现目标识别不精确与追踪效果差的问题,在追踪目标数量多且相互遮蔽情况严重的环境中追踪目标丢失现象更为明显。为此,提出一种基于Transformer跨空间特征关联的多目标追踪算法,利用Transformer结构提取全局特征与多头注意力机制的优势提升目标特征的提取能力。此外,为解决追踪目标之间相互遮蔽而导致的追踪目标丢失问题,利用互注意力机制映射追踪目标与干扰目标之间的特征进行增强与抑制,以提高追踪算法的准确性和可靠性;同时基于追踪目标与干扰目标之间的特征相似度确定特征增强与抑制的程度。在MOT16与MOT17数据集上进行实验,所提算法分别取得了58.81%和60.05%的多目标追踪准确性,相较其他主流算法性能更优。

Multi target tracking algorithms often suffer from inaccurate target recognition and poor tracking performance in complex scenes.In environments with a large number of tracked targets and severe mutual occlusion,the phenomenon of target loss is more pronounced.To this end,a multi-target tracking algorithm based on Transformer cross spatial feature association is proposed,which utilizes the advantages of Transformer structure to extract global features and multi head attention mechanism to improve the extraction ability of target features.In addi⁃tion,to solve the problem of tracking target loss caused by mutual occlusion between tracking targets,a mutual attention mechanism is used to map the features between tracking targets and interfering targets for enhancement and suppression,in order to improve the accuracy and reli⁃ability of the tracking algorithm;At the same time,the degree of feature enhancement and suppression is determined based on the feature simi⁃larity between the tracking target and the interfering target.Experiments were conducted on the MOT16 and MOT17 datasets,and the pro⁃posed algorithm achieved multi-target tracking accuracy of 58.81%and 60.05%,respectively,with better performance compared to other mainstream algorithms.