Road And Obstacle Information Extraction Algorithm Based On Multi-layer Lidar

Driverless car is a major component of intelligent transportation system, which includes environmental perception, planning decision, control execution and so on. Multi-layer Lidar has been widely used in driverless car with its high precision, large amount of data, fast speed and strong robustness. The information extraction of road and obstacle are studied based on multi-layer Lidar. The primary work of this dissertation are summarized as follows:(1) According to the characters of road edge data, road edge data set is extracted from all Lidar data. Similarity measure method is applied to improve the COBWEB algorithm for increasing the accuracy of cluster analysis of the set. Multi-layer fusion rule is proposed to fuse multi-layer Lidar data, eliminate interference road edges and distinguish left or right edge. Road is divided to drivable area and undrivable area by final road edges fitted out by least square method. In drivable area, road slope detection algorithm is proposed that based on the relative positional relationship between data of different scanning layers, which can distinguish flat road, uphill and downhill.(2) For excluding road surface data, apply three-dimensional information of Lidar to build 3D local grid map. When local map and global map are fused by Dempster-Shafer theory(DST) in dynamic environment, to solve the mismatch problem, we estimate firstly the location of the local grid map by speed and orientation of driverless car, and then use DST to fuse the two maps, which can improve the accuracy of grid map built by DST.(3) Obstacles are detected by analyzing conflict information in DST, and apply closing operation grouped by dilation and erosion algorithm to fill the gaps and cracks of obstacles. We improve the repeated stack visits and a large number of redundant neighborhood search in the original labeling algorithm, and apply the improved eight neighborhood labeling algorithm to cluster dynamic obstacles for extracting the length, width, center position and other static information of obstacles.(4) Considering the excellent stability of Kalman filter, we propose a dynamic information extraction method of obstacles based on Kalman filter and apply Kalman filter to build a variable tracking gate that can change with the length, width, center position and orientation of each tracked target. We apply Mahalanobis distance based on multiple features to improve the error tracking of nearest neighbor data association algorithm in the dense environment for exactly matching the optimal target with each tracked target.Finally, experimental results show that the proposed algorithm can stably and accurately detect the road edges and slope information, detect and track obstacles, extract static and dynamic information of targets, etc.