基于视觉相机和激光雷达融合的无人车障碍物检测与跟踪研究  

Research on Obstacle Detection and Tracking of Autonomous Vehicles Based on the Fusion of Vision Camera and LiDAR

作　　者：魏超[1,2] 吴西涛 朱耿霆 舒用杰 李路兴 随淑鑫 WEI Chao;WU Xitao;ZHU Gengting;SHU Yongjie;LI Luxing;SUI Shuxin(Science and Technology on Vehicle Transmission Laboratory,Beijing Institute of Technology,Beijing 100081;Institute of Advanced Technology,Beijing Institute of Technology,Beijing 100081)

机构地区：[1]北京理工大学坦克传动国防科技重点实验室,北京100081 [2]北京理工大学前沿技术研究院,北京100081

出　　处：《机械工程学报》2025年第2期296-309,320,共15页Journal of Mechanical Engineering

基　　金：国家自然科学基金资助项目(51875039)。

摘　　要：为提高无人车障碍物检测跟踪的精度和稳定性,首先针对YOLO v5(You only look once version 5,YOLO v5)网络存在的语义信息和候选框信息丢失的问题,引入深度可分离空洞空间金字塔结构与目标框加权融合算法完成对网络的优化;其次针对单阶段障碍物点云聚类精度低的问题,设计一种考虑点云距离与外轮廓连续性的两阶段障碍物点云聚类方法并完成三维包围盒的建立;最后将注意力机制引入MobileNet使网络更加聚焦于目标对象特有的视觉特征,并综合利用视觉特征和三维点云信息共同构建关联性度量指标,提高匹配精度。利用KITTI数据集对构建的障碍物目标检测、跟踪与测速算法进行仿真测试,并搭建实车平台进行真实环境试验,验证所提算法的有效性和真实环境可迁移性。To improve the accuracy and stability of obstacle detection and tracking,depthwise separable atrous spatial pyramid pooling(DASPP)layer and weighted boxes fusion(WBF)algorithm are firstly introduced into you only look once version 5(YOLO v5)to tackle the problems of loss of semantic information and candidate box information,respectively.Then,a two-stage point cloud clustering method considering the point cloud distance and the continuity of the outer contour is proposed and a bounding box is established to improve the clustering accuracy of each target while ensuring the recall rate of obstacle targets.Finally,the convolutional block attention module(CBAM)is added into MobileNet to effectively extract the visual features of the obstacle target,visual features and 3D information are combined to establish correlation metrics and thus to improve tracking precision.Tests based on KITTI dataset and real environments show the effectiveness and transferability of the proposed algorithm.

关 键 词：视觉相机 激光雷达 目标检测 多目标跟踪 无人车 

分 类 号：U463[机械工程—车辆工程]