基于多层激光雷达的道路与障碍物信息提取算法

发布时间：2018-11-01 20:16

【摘要】：无人驾驶车辆是智能交通系统的重要组成部分,主要包括环境感知、规划决策和控制执行等子系统。多层激光雷达凭借其测量精度高、数据多、速度快、鲁棒性强等优点在无人驾驶车的环境感知系统中得到了广泛应用。论文对基于多层激光雷达提取无人驾驶车周围的道路与障碍物信息进行研究,主要内容如下:(1)根据路沿数据点特征从众多的激光雷达数据中提取出路沿数据集。应用相似性度量方法对COBWEB算法进行改进,以提高路沿数据集聚类分析的准确率。提出多层融合表示规则能够融合多层激光雷达数据、剔除干扰路沿、分清左右路沿,并通过最小二乘法拟合出最终路沿,将道路分割为可行驶区域和不可行驶区域。在可行驶区域内根据三维激光雷达不同扫描层之间数据点的相对位置关系提出道路坡度检测算法,能够判别平坦路、上坡路和下坡路等路况信息。(2)为剔除路面数据点,应用多层激光雷达数据的三维信息建立三维局部栅格地图。为解决动态环境下应用DS证据理论(Dempster-Shafer Theory,DST)融合局部地图与全局地图时的不匹配问题,本文提出首先根据无人驾驶车辆的运动速度、运动方向等信息将局部栅格地图进行位置估计后,再使用DST融合规则将两地图进行融合,提高了基于DS证据理论建立栅格地图的精确度。(3)利用DST中的冲突系数检测动态障碍物,并采用膨胀、侵蚀算法组合而成的闭运算填补障碍物的漏洞和裂缝。针对经典区域标记算法重复访问堆栈和大量冗余邻域搜索等问题进行改进,并应用改进的八邻域区域标记算法对动态障碍物进行聚类分析,以提取障碍物的长度、宽度和中心位置等静态信息。(4)考虑到Kalman滤波器具有出色的稳定性,本文提出基于Kalman滤波器的障碍物动态信息提取方法。并提出应用Kalman滤波器为每个被跟踪目标建立一个随目标中心位置、长度、宽度和航向角实时变化的可变跟踪门,增加了障碍物目标跟踪系统的自适应能力。针对最近邻数据关联算法在密集环境下容易产生错误跟踪等问题,提出基于多特征马氏距离改进的最近邻数据关联算法,能够在密集环境下准确地为多被跟踪目标匹配最优关联目标。最后,通过实车试验验证了上述方法能够稳定、准确、快速地完成检测路沿信息、道路坡度信息、检测并跟踪测障碍物目标、提取目标的动静态信息等工作。
[Abstract]:Driverless vehicle (UAV) is an important part of Intelligent Transportation system (its), which includes environment perception, planning decision and control execution. Multi-layer lidar has been widely used in the environment sensing system of driverless vehicles because of its advantages of high measurement accuracy, high speed, high speed and high robustness. In this paper, the road and obstacle information around driverless vehicle is extracted based on multilayer lidar. The main contents are as follows: (1) the road edge data sets are extracted from many lidar data according to the road edge data points. The similarity measure method is used to improve the COBWEB algorithm in order to improve the accuracy of data cluster analysis along the road. It is proposed that the multilayer fusion representation rule can fuse the multilayer lidar data, eliminate the interference path edges, distinguish the left and right edges, and fit the final path edge by the least square method, and divide the road into drivable and non-drivable areas. According to the relative position relation of data points between different scanning layers of 3D lidar, a road slope detection algorithm is proposed in the traveling region, which can distinguish the road condition information such as flat road, uphill road and downhill road. (2) in order to eliminate the road surface data points, Three-dimensional local raster map is built by using three-dimensional information of multi-layer lidar data. In order to solve the mismatch problem when using DS evidence theory (Dempster-Shafer Theory,DST) to fuse the local map with the global map in dynamic environment, this paper proposes a method based on the velocity of driverless vehicle. The location of the local raster map is estimated by moving direction and other information, and then the two maps are fused by using the DST fusion rule. The accuracy of building grid map based on DS evidence theory is improved. (3) the collision coefficient in DST is used to detect the dynamic obstacle, and the closed operation composed of expansion and erosion algorithm is used to fill the hole and crack of the obstacle. Aiming at the problems of repeated access stack and redundant neighborhood search in classical region marking algorithm, the improved eight-neighborhood region marking algorithm is applied to cluster analysis of dynamic obstacles to extract the length of obstacles. Static information such as width and center position. (4) considering the excellent stability of Kalman filter, an obstacle dynamic information extraction method based on Kalman filter is proposed in this paper. A variable tracking gate based on the center position, length, width and heading angle of each tracked target is proposed by using Kalman filter, which increases the adaptive ability of the obstacle target tracking system. Aiming at the problem that the nearest neighbor data association algorithm is easy to generate error tracking in dense environment, an improved nearest neighbor data association algorithm based on multi-feature Markov distance is proposed. It can accurately match the optimal target association for multiple tracked targets in dense environment. Finally, the method is proved to be stable, accurate and fast in detecting road edge information, road slope information, detecting and tracking obstacle targets, and extracting static and dynamic information of targets.
【学位授予单位】：北京工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U495;U463.6;TP391.41

【参考文献】