面向目标遮挡场景的车辆实时跟踪方法  

Real-time vehicle tracking method for target occlusion scenes

作　　者：李凯 林宇舜 LI Kai;LIN Yushun(School of Transportation and Civil Engineering,Fujian Agriculture and Forestry University,Fuzhou 350108,China)

机构地区：[1]福建农林大学交通与土木工程学院,福州350108

出　　处：《交通科技与经济》2022年第4期48-54,共7页Technology & Economy in Areas of Communications

基　　金：福建省自然科学基金项目(2020J05029)。

摘　　要：对视频车辆进行实时准确跟踪,可以为智能化交通管控提供重要信息。然而,在一些复杂的行车环境中,车辆之间的遮挡现象会严重影响跟踪性能,因此,为实现在复杂交通背景下车辆的实时鲁棒跟踪,提出一种基于卡尔曼滤波与动态卷积的轻量级重识别车辆跟踪方法。实验结果显示:该跟踪方法在公开的UA-DETRAC车辆数据集上的MOTA达到80.25%,比现有基于ResNet的跟踪方法提高2.42%,IDSW显著减少10%,且优化后的跟踪速度较之前提升25%。研究结果表明,该跟踪方法具有一定的实时性和抗遮挡性。Real-time and accurate tracking of video vehicles can provide important information for intelligent traffic control.However,in some complex driving environments,the occlusion phenomenon between vehicles can seriously affect the tracking performance.Therefore,in order to achieve real-time robust vehicle tracking in complex traffic backgrounds,a lightweight re-identification vehicle tracking method based on Kalman filter and dynamic convolution is proposed.The experimental results show that the MOTA of the tracking method on the public UA-DETRAC vehicle dataset reaches 80.25%,which is 2.42%higher than the existing ResNet-based tracking method,the IDSW is significantly reduced by 10%,and the optimized tracking speed is improved 25%.The research results show that the tracking method has certain real-time performance and anti-occlusion performance.

关 键 词：智能化交通管控 遮挡实时性 卡尔曼滤波 动态卷积 轻量级跟踪方法 车辆跟踪 

分 类 号：U491[交通运输工程—交通运输规划与管理]