基于车载多激光雷达的地图构建与障碍物检测

发布时间：2018-04-27 07:29

  本文选题：激光雷达 + 同时定位与建图　； 参考：《浙江大学》2017年硕士论文

【摘要】：基于激光雷达的障碍物检测与地图构建是无人驾驶系统中环境感知的重要组成部分。当前主流采用的Velodyne HDL64E激光雷达具有体积大、价格高的缺点,为改善这一问题,本文提出使用多个小激光雷达组合的方式进行地图构建与障碍物检测。本文分析了 VelodyneHDL32E和VLP16激光雷达在不同安装方式下的扫描精度,提出了多个激光雷达之间的组合安装方式和标定方法。为得到车体运动轨迹和地图,本文实现了以激光雷达为传感器的SLAM子系统,即从点云中提取“线特征点”和“面特征点”用于帧间最近邻匹配,通过最小化匹配误差求出无人车帧间运动量,此外通过地图配准和闭环优化两个步骤减小累积误差。除了车体轨迹和地图,SLAM子系统输出的去畸变点云可进行多帧融合获得更加精确、致密的点云,有助于正、负障碍物检测。本文采用主流的栅格属性地图表示障碍物分布,通过分析正障碍物点云空间分布特征对正障碍物分类。此外,本文提出了负障碍物的三个局部结构特征,通过检测负障碍物候选线段,并对候选线段聚类处理获得负障碍物区域。实际数据集上的定性实验表明基于激光雷达的SLAM系统能够获得准确的车体轨迹和高精度点云地图。此外在越野环境下的定量试验表明多激光雷达组合与多帧融合的方法能显著提升点云密度,提高正障碍物和负障碍物检测效果。
[Abstract]:Obstacle detection and map construction based on lidar is an important part of environment perception in unmanned systems. The current mainstream Velodyne HDL64E lidar has the disadvantages of large volume and high price. In order to improve this problem, this paper proposes to use multiple small lidar combinations to construct maps and detect obstacles. In this paper, the scanning accuracy of VelodyneHDL32E and VLP16 lidar under different installation modes is analyzed, and the combined installation mode and calibration method between several lidar radars are put forward. In order to obtain the motion track and map of the car body, a SLAM subsystem based on lidar sensor is implemented in this paper, namely, "line feature points" and "surface feature points" are extracted from the point cloud for the nearest neighbor matching between frames. By minimizing the matching error, the motion amount between the frames of the unmanned vehicle is obtained, and the cumulative error is reduced by map registration and closed-loop optimization. In addition to the car-body trajectory and map slam subsystem output dedistorted point clouds can be fused in multiple frames to obtain more accurate and dense point clouds, which is helpful for both positive and negative obstacle detection. In this paper, the main raster attribute map is used to represent the distribution of obstacles, and by analyzing the spatial distribution characteristics of point cloud of positive obstacles, the positive obstacles are classified. In addition, three local structural features of negative obstacles are presented in this paper. The candidate segments of negative obstacles are detected, and the negative obstacle regions are obtained by clustering the candidate segments. The qualitative experiments on the actual data set show that the SLAM system based on lidar can obtain accurate vehicle track and high accuracy point cloud map. In addition, quantitative experiments in off-road environment show that the combination of multi-lidar and multi-frame fusion can significantly improve the point cloud density and improve the effectiveness of both positive and negative obstacle detection.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN958.98
，

本文编号：1809814