Map Building And Obstacle Detection Based On Vehicle-mounted Multiple Lidars

The lidar based obstacle detection and map building are important components of environmental perception module in self-driving car system.The current popular Velodyne HDL64E lidar has the disadvantage of being bulky and expenive.In order to improve this problem,this thesis proposes a method of combining multiple small lidars for map building and obstacle detection.In this thesis,the scanning precision of Velodyne HDL32E and VLP16 lidar in different installation modes is analyzed,and the method of combination installation and calibration is proposed.To obtain the car trajectory and surrounding map,a lidar based SLAM subsystem is used,in which the edge-keypoints and planar-keypoints are extracted and used for inter-frame nearest neighbor matching.By minimizing the correspponding matching error,the inter-frame transformation is calculated,in addition the map registration and loop closure are used to reduce cumulative error.Besides of car trajectory and surrounding map,the SLAM subsystem outputs distortion-free point clouds which can be used in multi-frame fusion to obtain more accurate and denser point cloud which is helpful in positive and negative obstacle detection.In this thesis,the obstacle distribution is represented by the popular grid map,and the positive obstacles are classified by means of analyzing the point cloud distribution feature.In addition,three local structural features of negative obstacle are proposed,and negative obstacles are detected by extracting and clustering candidate negative segments.Qualitaitve experiments on real world dataset show that the lidar based SLAM can obtain accurate car trajectory and high precision point cloud map.Besides,quantative experiments on rural environment show that the multi-lidar combination and milti-frame fusion improve point cloud density significantly,and improve the detection of positive and negative obstacles.