基于自校准与异构网络的多目标跟踪算法  

Multi-target tracking algorithms based on self-calibrationand asymmetric convolution

作　　者：王玲 杜新兆 罗可心 王鹏[1] 赵领娣 WANG Ling;DU Xin-zhao;LUO Ke-xin;WANG Peng;ZHAO Ling-di(College of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,China)

机构地区：[1]长春理工大学计算机科学技术学院,吉林长春130022

出　　处：《计算机工程与设计》2023年第12期3662-3669,共8页Computer Engineering and Design

基　　金：中央引导地方科技发展基金吉林省基础研究专项基金项目(202002038JC)。

摘　　要：为解决多目标跟踪算法在遮挡场景下导致的身份切换等问题,提升算法跟踪精度,提出一种融合自校准与异构卷积的离线图跟踪网络(self-calibrated convolutions and asymmetric convolution track, SCAACTrack)。利用融合自校准卷积网络与非对称结构进行目标外观特征提取,提升算法行人重识别能力。通过采用不同帧之间目标外观特征、时间和空间3种维度进行图神经网络建模,引入基于时间感知的消息传递网络加强多目标跟踪流式守恒约束。实验结果表明,与传统的多目标跟踪算法MPNTrack、Tracktor、KCF等模型相比,该模型跟踪效果更有效。To solve the problem of identity switching caused by the multi-target tracking algorithm in the occlusion scene,and improve the tracking accuracy of the algorithm,an offline graph tracking network(self-calibrated convolutions and asymmetric convolution track,SCAACTrack)was proposed.The integration of self-calibration convolution network and asymmetric structure was used to extract the appearance features of the target,the pedestrian recognition ability of the algorithm was effectively improved.The graph neural network was modeled using the three dimensions of target appearance characteristics,and a time-aware message passing network was introduced to strengthen the streaming conservation constraints of multi-target tracking.Experimental results show that compared with the traditional multi-target tracking algorithms MPNTrack,Tracktor,KCF and other models,the tracking effect of this model is more effective.

关 键 词：多目标跟踪 身份切换 异构卷积 自校准卷积 离线图跟踪网络 特征提取 图神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]