基于互相关注意力的链式帧处理多目标跟踪算法

Multi-object Tracking Based on Cross-correlation Attention and Chained Frames

多目标跟踪的一阶段方法因其在推理速度方面的优势逐渐成为主流。然而,与两阶段方法相比,其跟踪精度较差。一方面是因为采用单幅图像输入,目标间的关联性不强,容易导致目标丢失,另一方面忽视了检测和跟踪两个任务之间的差异性。为了减轻上述限制,提出了一种基于互相关注意力的链式帧处理多目标跟踪算法(MOT-CCC)。MOT-CCC将连续的两帧图片作为输入,将目标关联问题转化为两帧检测框对回归的问题,增强了目标间的关联性;采用互相关注意力模块将检测任务和身份识别任务解耦,以平衡并减少这两个任务之间的竞争。此外,所提算法将目标检测、特征提取和数据关联3个模块融合到一个网络中,实现了端到端的优化,提高了跟踪准确性,减少了跟踪耗时。在MOT16和MOT17基准测试中,MOT-CCC比原有的基准CTracker算法的MOTA提高了1.3%,FP减少了13%。

The one-stage method of multi-object tracking(MOT)has gradually become the mainstream of MOT due to its advantages in reasoning speed.However,compared with the two-stage method,its tracking accuracy is poor.One reason is that the target is easy to be lost due to the use of single frame input that cause the correlation between the targets is not strong,the other is that the difference between the two tasks of detection and tracking is ignored.In order to alleviate the limitations,a multi-object tracking algorithm based on cross-correlation attention and chained frames(MOT-CCC)is proposed.MOT-CCC takes two consecutive frames as input,and converts the target association problem into a two-frame detection frame pair regression problem,which enhances the correlation between targets.The cross-correlation attention module decouples the detection task and the identification task to balance and reduce the competition between the two tasks.In addition,the proposed algorithm integrates the three modules of target detection,feature extraction and data association into one whole network to achieve end-to-end optimization,which improves tracking accuracy and reduces tracking time.In the MOT16 and MOT17 benchmark tests,compared with the benchmark CTracker algorithm,the MOTA of MOT-CCC increases by 1.3%and the FP decreases by 13%.