基于卷积混合注意力机制的多目标跟踪算法  

Multi-target tracking algorithm based on convolutional hybrid attention mechanism

作　　者：郭崇 刘晟 张文波[1] 朱宏博 GUO Chong;LIU Sheng;ZHANG Wen-bo;ZHU Hong-bo(School of Information Science and Engineering,Shenyang Ligong University,Shenyang 110159,China;School of Computer Science and Engineering,Northeastern University,Shenyang 110169,China)

机构地区：[1]沈阳理工大学信息科学与工程学院,沈阳110159 [2]东北大学计算机科学与工程学院,沈阳110169

出　　处：《控制与决策》2025年第4期1127-1135,共9页Control and Decision

基　　金：国家自然科学基金项目(62102272);辽宁省博士科研启动基金计划项目(2023-BS-130)。

摘　　要：基于检测的多目标跟踪方法在复杂场景问题上达到了较好的效果,但已有研究大多关注于时空特征关联而忽视了提高检测性能所能带来的全局跟踪收益.据此,提出一种卷积混合注意力机制,该模块结合动态稀疏通道注意力和空间位置注意力:在处理通道注意力时,整合空间上下文信息,动态调整通道权重;在处理空间注意力时,结合不同通道特征评估空间区域的重要性,旨在优化注意力分配并提升检测精度.进一步地,提出一种两阶段多目标跟踪方法——CHAMTrack,通过在运动目标检测阶段使用该注意力机制,增强算法在复杂场景中对关键信息的捕捉能力,提升不同尺度目标的跟踪效果,降低跟踪过程中漏检和ID切换的发生率.在MOT17和MOT20数据集上的实验结果表明,CHAMTrack在MOTA指标上分别提升28%和20.5%,显著提升了多目标跟踪算法在复杂场景中的效果和鲁棒性.Currently,detection-based multi-target tracking methods have achieved better results in complex scene problems,but most of the existing research focuses on spatio-temporal feature correlation and neglects the global tracking benefit that can be brought by improving detection performance.Accordingly,this paper proposes a convolutional hybrid attention mechanism,which combines dynamic sparse channel attention and spatial location attention:when dealing with channel attention,it integrates spatial context information to dynamically adjust the channel weights;when dealing with spatial attention,it combines different channel features to evaluate the importance of spatial regions,aiming at optimising the allocation of attention and improving the detection accuracy.Further,this paper proposes a two-phase multi-target tracking method,CHAMTrack,to enhance the algorithm's ability to capture key information in complex scenes,improve the tracking effect of targets at different scales,and reduce the occurrence rate of omission and ID switching in the tracking process by using the attention mechanism in the detection phase of moving targets.The incidence of missed detection and ID switching during the tracking process is reduced.The experimental results on MOT17 and MOT20 datasets show that the CHAMTrack improves the MOTA metrics by 2.1%and 1.3%,and the IDSw.metrics by 28%and 20.5%,which significantly improves the effectiveness and robustness of the multi-target tracking algorithms in complex scenes.

关 键 词：计算机视觉 多目标跟踪 特征增强 通道注意力机制 空间注意力机制 卷积混合注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]