基于长短期超图神经网络匹配的多目标跟踪  

A multi-object tracking method based on long-term and short-term hypergraph neural network matching

作　　者：郭文[1] 刘其贵 王拓 丁昕苗[1] GUO Wen;LIU Qi-gui;WANG Tuo;DING Xin-miao(School of Information and Electronic Engineering,Shandong Technology and Business University,Yantai 264005,China)

机构地区：[1]山东工商学院信息与电子工程学院,山东烟台264005

出　　处：《控制与决策》2025年第3期853-862,共10页Control and Decision

基　　金：国家自然科学基金项目(62072286,61876100,61572296);山东省研究生教育创新计划项目(SDYAL21211).

摘　　要：针对联合检测与跟踪范式中存在的检测特征和Re-ID特征相互竞争的问题以及在复杂场景下难以保持被遮挡目标视觉一致性关系的问题,提出一个端到端的超图神经网络关联的多目标跟踪方法(HGTracker).首先,HGTracker设计一个增强的空间金字塔池化网络(ESPPNet)模块用来提高目标检测骨干网络的检测能力,该模块通过聚合不同维度的特征来适应跟踪过程的不同任务,有效地缓解一阶段跟踪方法中检测任务与Re-ID任务相互竞争的问题.其次,提出一个基于长短期超图神经网络的数据关联模块,通过设计长期超图神经网络和短期超图神经网络来分别关联未被遮挡和被遮挡的检测视觉特征,将数据关联问题转化为轨迹超图与检测超图之间的超图匹配问题,跟踪器将轨迹片段信息与当前检测帧信息之间的关系建模为超图神经网络,在严重遮挡的情况下保持了视觉轨迹的一致性.通过一系列的对比实验,所提出的HGTracker跟踪方法相比于FairMOT跟踪方法,在MOT17数据集上HOTA值由59.3%提高至61.4%,IDF1值由73.7%提高至79.3%,MOTA值由72.3%提高至76.9%;在MOT20数据集上,HOTA值由54.6%提高至57.9%,IDF1值由61.8%提高至73.1%,MOTA值由67.3%提高至75.1%.Addressing the issues of competition between detection features and Re-ID features in joint detection and embedding multi-object tracking methods,as well as difficulties in maintaining visual consistency for occluded targets in complex scenes,we propose an end-to-end hypergraph neural network matching tracking method,named HGTracker.Firstly,the HGTracker introduces an enhanced spatial pyramid pooling networks(ESPPNet)module to enhance the detection capability of the target detection backbone network.This module aggregates features from different dimensions to adapt to different tasks in the tracking process,effectively alleviating the issue of competition between detection and Re-ID tasks in one-stage multi-object tracking methods.Secondly,it introduces a short-term and long-term hypergraph neural network matching module,which designs long-term and short-term hypergraph neural networks to associate unoccluded and occluded detection visual features.It transforms the data association problem into a hypergraph matching problem between trajectory hypergraphs and detection hypergraphs.The tracker models the relationship between trajectory segment information and the current detection frame information as a hypergraph neural network,maintaining visual trajectory consistency under severe occlusion.Through a series of comparative experiments,compared to the FairMOT tracking method on the MOT17 dataset,the proposed tracking method increases the HOTA value from 59.3% to 61.4%,the IDF1 value from 73.7% to 79.3%,and the MOTA value from 72.3% to 76.9%;on the MOT20 dataset,the HOTA value is increased from 54.6% to 57.9%,the IDF1 value from 61.8% to 73.1%,and the MOTA value from 67.3% to 75.1%.

关 键 词：多目标跟踪 超图神经网络匹配 视觉一致性关系 数据关联 联合检测与跟踪范式 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]