基于图网络最优匹配的多目标跟踪算法  

A Multi-object Tracking Algorithm Based on Optimal Matching of Graph Networks

作　　者：孙波 任劼[1] 吴涛 SUN Bo;REN Jie;WU Tao(School of Electronics and Information,Xi'an Plolytechnic University,Xi'an 710048,China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048

出　　处：《长江信息通信》2022年第4期32-35,共4页Changjiang Information & Communications

基　　金：陕西省自然科学基础研究计划(2022JM-394)资助;陕西省教育厅科研计划项目(19JK0364)资助。

摘　　要：为解决多目标跟踪任务中复杂场景下因检测器漏检或频繁遮挡导致特征表达信息不足和数据关联不正确的问题。文章提出了一个双线多目标跟踪方法;使用外观与运动的多特征信息解决特征表达信息不足的问题;而数据匹配采用端到端的图网络(GNN)进行图最优匹配,并结合传统匈牙利算法优化数据关联。实验在MOT数据集上与近年来经典跟踪方法进行了比较,在MOT17数据集上综合性能指标MOTA、IDF1均高于该性能指标排名第二的方法,MOTA提高了3%,IDF1提高了5%;并通过消融实验验证了多特征信息对多目标跟踪性能提高的有效性。In order to solve the problems of insufficient feature expression information and incorrect data association caused by missing detection or frequent occlusion of detectors in complex scenes in multi-object tracking tasks.This paper proposes a two-line multi-objecttracking method,weuse multi feature information of appearance and motion to solve the problem of insufficient feature expression information;we matching data adopts end-to-end graph network(GNN)for graph optimal matching,and optimizes data association combined with traditional Hungarian algorithm.Experiments in the MOT data set compared with the classic tracking method in recent years,in the MOT17 data set,the comprehensive performance index MOTA,IDF1 are higher than the second-ranked method of the performance index,MOTA increased by 3%,IDF1 increased by 5.%,and through the melting experiment to verify the effectiveness of multi-feature information to multi-object tracking performance.

关 键 词：机器视觉 多目标跟踪 卷积神经网络 光流 图网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]