基于激光雷达的远距离运动车辆位姿估计  

Estimation of Long Distance Moving Vehicle Pose Based on LIDAR

作　　者：张名芳[1] 付锐[1] 石涌泉[1] 程文冬[2] 

机构地区：[1]长安大学汽车学院,陕西西安710064 [2]西安工业大学机电工程学院,陕西西安710032

出　　处：《公路交通科技》2017年第12期131-139,共9页Journal of Highway and Transportation Research and Development

基　　金：国家自然科学基金项目(61473046);中央高校基本科研业务费专项资金项目(310822151028;310822172001);陕西省自然科学基金项目(2016JQ5096);长江学者和创新团队发展计划项目(IRT1286)

摘　　要：为了解决激光雷达扫描远距离运动车辆产生的点云稀疏导致位姿特征难以提取的问题,提出了一种远距离运动车辆位姿估计方法。首先利用时空连续性提取远距离运动车辆。然后利用最小二乘拟合得到稀疏点云水平面二维投影近似拟合直线对,依次在不同角度的垂直正交直线对上对稀疏点云的二维投影进行一维向量估计的装箱过程,基于目标车辆与激光雷达间相对位置的观测角函数最大化匹配滤波响应,进而利用全局优化算法对投影点概率分布与匹配滤波运算得到的代价函数作离散卷积,寻优比较得到单帧拟合最优矩形。最后结合连续帧平移约束进行多帧拟合,优化当前帧目标车辆拟合矩形的位姿。利用仿真和真实场景下采集的目标车辆点云数据进行算法验证分析。结果表明:在点云稀疏的情况下,当远距离目标车辆做直线运动时,提出的多帧拟合方法得到的位姿参数均方根误差低于单帧拟合和已有的RANSAC拟合方法;当远距离目标车辆做曲线运动时,提出的单帧拟合和多帧拟合方法得到的位姿估计结果较为接近,且误差明显低于已有的RANSAC拟合方法;对于不同相对距离下采集的目标车辆点云,提出的单帧拟合和多帧拟合位姿估计方法的适应性优于已有的RANSAC拟合方法。In order to solve the problem of the difficulty in extracting the pose feature from sparse point cloud produced by LIDAR scanning long-distance moving vehicle, a method for pose estimation of long distance moving vehicles is proposed. First, the long distance moving vehicle is extracted with the temporal-spatial continuity. Then, the approximate fitting line pair of the 2D projection on the horizontal plane for the sparse point cloud is obtained by using the least square fitting, and the binning process of 1D vector estimation for the 2D projection of sparse point cloud on the perpendicular orthogonal straight lines at different angles in order is conducted, the matched filtering response is maximized based on the viewing angle function between the target vehicle and LIDAR, the global optimization algorithm is further used to do the discrete convolution on the cost function from the projective point probability distribution and the matching filtering, and the optimal rectangle of single frame fitting is achieved by the optimization. Finally, the multi-frame fitting is conducted based on the translational restraint of consecutive frames, and the fitting rectangle pose of the target vehicle at the current frame is optimized. The point cloud dataset of the target vehicle is collected in the simulation and real scenes to verify and analyze the proposed algorithm. The result shows that （1） when the long distance target vehicle is moving in a straight line under the condition of sparse point cloud, the root mean square errors of the pose parameters obtained from the multi-frame fitting are lower than the errors of single frame fitting and the existing RANSAC fitting method; （2） when the long distance target vehicle does curve motion, the pose estimation results obtained from the proposed single frame fitting and muhi-frame fitting methods are approximate, and their errors are significantly lower than that of the existing RANSAC fitting method ; （3） for the point cloud of target vehicle collected at differen

关 键 词：汽车工程 位姿估计 全局优化 多帧拟合 稀疏点云 匹配滤波 

分 类 号：U463.675[机械工程—车辆工程]