抗遮挡的行人多目标跟踪算法  

Pedestrian multiobject tracking algorithm with anti-occlusion

作　　者：张国印[1] 王传博 高伟[1] ZHANG Guoyin;WANG Chuanbo;GAO Wei(College of Computer Science and Technology,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学计算机科学与技术学院,黑龙江哈尔滨150001

出　　处：《智能系统学报》2024年第5期1248-1256,共9页CAAI Transactions on Intelligent Systems

摘　　要：为了解决在复杂场景下行人相互遮挡导致跟踪系统精度降低的问题,提出了基于FairMOT的抗遮挡多目标跟踪算法(multiple obeject tracking algorithm with anti-occlusions,AOMOT)。首先通过轻量化平衡模块,解耦不同层次的语义信息,减少检测任务和重识别任务的语义冲突,降低重识别任务的性能提升对检测任务的影响。其次应用自注意力结构提取行人的外观特征,加强局部窗口下的类内特征的区分度,增强行人身份信息的匹配一致性并减少身份标识的频繁切换。最后优化身份关联算法,挖掘低置信度目标中的被遮挡对象,将其重新纳入目标身份关联并更新其重识别特征。实验结果表明,AOMOT相比原有FairMOT在MOT17数据集中高阶跟踪精度提升1.5百分点,身份F1分数提升3百分点,身份切换数量降低32%。A multiobject tracking algorithm with anti-occlusion(referred to as AOMOT),which is based on the Fair-MOT framework,is proposed to improve the accuracy of tracking systems in crowded pedestrian scenes.First,the lightweight balance module decouples semantic information at different levels to minimize semantic conflicts between detection and recognition tasks and decrease the impact of performance improvement in re-identification tasks.Second,the self-attention structure is adopted to extract pedestrian appearance features and improve the discrimination of intra-class features under local windows.The matching consistency of pedestrian identity information is enhanced,and frequent switching of identity signs is reduced.Finally,the identity association algorithm is optimized to mine occluded objects in low-confidence targets,reincorporate them into the target identity association,and update their recognition features.Experimental results show that,compared with the original model in the MOT17 dataset,the improved model enhances the higher-order tracking accuracy by 1.5 percentage points,improves identity F1 score by 3 percentage points,and reduces identity switching by 32%.

关 键 词：计算机视觉 行人跟踪 目标检测 重识别 关联算法 抗遮挡 自注意力 特征提取 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]