毫米波雷达和视觉融合的车辆鲁棒跟踪方法  

Robust Vehicle Tracking Method Based on Radar and Vision Fusion

作　　者：郭熙 胡广地 王志琛 周红梅[1] GUO Xi;HU Guang-di;WANG Zhi-chen;ZHOU Hong-mei(School of Mechanical Engineering,Southwest Jiaotong University,Sichuan Chengdu 610031,China)

机构地区：[1]西南交通大学机械工程学院,四川成都610031

出　　处：《机械设计与制造》2024年第7期253-259,共7页Machinery Design & Manufacture

基　　金：四川省科技计划项目(2019YFH0096);四川省氢能源与智能汽车重大科技专项(2019ZDZX0028)。

摘　　要：针对单传感器实际工作环境中的漏检错检问题,提出了一种具有鲁棒性的摄像头毫米波雷达并联融合识别追踪车辆方法。构建基于毫米波雷达多维数据的目标识别网络R-DenseNet,毫米波雷达数据经过初筛后输入R-DenseNet神经网络得到毫米波雷达识别结果;摄像头的图像数据输入YOLO-v4tiny网络得到摄像头识别结果。对原始目标采用融合规则进行融合,筛选掉雷达错检点和噪点;对融合成功目标点采用真实目标存活判断方法进行持续追踪,捕捉摄像头毫米波雷达错检漏检的情况,使用Kalman滤波、YOLO重识别等方法补齐数据。实验结果证明:该融合方法可以有效排除错误目标,对于跟踪期间单传感器漏检、错检及两传感器同时漏检的情况具有较好的鲁棒性,跟踪较为稳定。Aiming at the problem of missed detection and false detection in the actual working environment of a single sensor,a robust vehicle tracking method based on radar and vision fusion is proposed.Firstly,a target recognition network called RDensenet based on multi-dimensional data which are outputted by millimeter-wave radar is constructed.To obtain the radar identification results,multi-dimensional data detected by millimeter-wave radar is input to R-Densenet neural network.Then,The image taken by camera is input into the YOLO-V4tiny network to get the camera recognition results.The original target points are fused by the fusion rule,this step is to filter out the radar false detection points and noise points.Then the successful fusion points are continuously tracked by using the real target survival judgment method.During the tracking process,false detection and missed detection of millimeter-wave radar or camera are captured and the lost data are supplemented by Kalman filter,YOLO re-identification and other methods.The results of the experiment show that the fusion method can effectively eliminate the false target,and has good robustness for single sensor missed detection,single sensor false detection and two sensors missed detection at the same time during tracking,so the tracking is relatively stable.

关 键 词：车辆跟踪 毫米波雷达 视觉 信息融合 时空配准 

分 类 号：TH16[机械工程—机械制造及自动化] TP[自动化与计算机技术]