基于激光雷达和视觉信息融合的车辆识别与跟踪  

Sensor Fusion of LiDAR and Vision for Vehicle Detection and Tracking

作　　者：宫铭钱 冀杰[1] 种一帆 陈琼红 Gong Mingqian;Ji Jie;Chong Yifan;Chen Qionghong(Southwest University,Chongqing 400715)

机构地区：[1]西南大学,重庆400715

出　　处：《汽车技术》2020年第11期8-15,共8页Automobile Technology

基　　金：国家自然科学基金项目(61304189);中央高校基本业务费专项资金重点项目(XDJK2019B053);山区复杂道路“人-车-路”系统协同与安全重庆市重点实验室开放基金(2018HVRC02)。

摘　　要：为提高自动驾驶系统对车辆目标的识别和跟踪精度,提出一种基于激光雷达和单目视觉的信息融合框架。利用改进的DBSCAN算法对障碍物点云数据进行聚类,采用更快速区域卷积神经网络(Faster R-CNN)识别单目视觉图像中的车辆目标,融合两种传感器的信息获得完整的车辆目标信息,使用联合概率数据关联(JPDA)算法实时跟踪目标车辆的运动状态。试验结果表明,该算法可有效避免传感器杂波的干扰,能够适应车辆目标的数量变化并及时更新航迹信息,具有较高的精度和鲁棒性。A sensor fusion framework based on LiDAR and monocular vision is proposed to improve the accuracy of target detection and tracking for autonomous driving system.Firstly,the improved DBSCAN algorithm is developed to extract the obstacles by clustering point cloud data.Meanwhile,the vehicle targets in the monocular visual image are detected by Faster R-CNN.The complete information of vehicle targets is obtained by fusion of the data from above 2 sensors and the JPDA algorithm is applied to track the movement of detected target vehicles in real time.The simulation results show that the proposed algorithm can effectively avoid the interference of sensor clutter,adapt to the change of the number of vehicle targets and update the track information in time with high precision and robustness.

关 键 词：激光雷达 视觉 信息融合 DBSCAN 更快速区域卷积神经网络 联合概率数据关联 

分 类 号：U463.6[机械工程—车辆工程]