Research On High Precision Semantic Point Cloud Mapping For Autonomous Driving

Traditional navigation maps can not meet the needs of autonomous driving.As a necessary part of autonomous driving,HD Map(High Definition Map)has gradually become a consensus in the industry,which has higher precision,more dimensions,up-to-date and other advantages.Not only can sensors perform the function of sensing the surrounding environment,but high-precision maps can also provide driving system with more forward-looking information indication and redundancy to help vehicles to match and locate.So that the driving system can sense a wider range of traffic situation,to ensure the safety of autonomous driving.The point cloud map is favored by the self-driving industry because it is not affected by environmental light,and its accurate reconstruction of the environment.However,due to the huge amount of data,point cloud map has great obstacles in map storage and online matching location.In recent years,the research on point cloud map has gradually turned to explore the semantic structure of point cloud map,hoping to bring about the innovation of map data storage and matching location technology.Based on the analysis of the key requirements,this paper proposes a high-precision semantic point cloud mapping method for autonomous driving.The main research work of this paper includes:First of all,the requirements of autonomous driving for high-precision semantic point cloud map are analyzed systematically.On this basis,a set of high-precision semantic point cloud map construction scheme is designed: it covers the definition of high-precision semantic point cloud map from the concept and data structure.The key requirements of the analysis,to the specific implementation of sensor configuration designing,installation,and then to mapping raw data collection,sensor data fusion and post-processing,map generation and delivery.Secondly,for multi-lidar panoramic camera as a multi-sensor system,this paper covers the joint calibration of multi-lidar,the joint calibration of panoramic camera and lidar and the joint calibration of panoramic camera and RTK-GPS.Thirdly,because there are still many problems in the segmentation of panoramic images by single-camera image semantic segmentation methods,this paper proposes a method of generating semantic panoramic images based on panoramic stitching.Furthermore,based on the precise joint calibration,the semantics of the image is transferred into the point cloud.At the same time,the error of semantic transmission may occur because of the difference between the laser radar vision and the panoramic camera view.The method based on probabilistic graph optimization is used to improve the robustness of semantic transmission.Finally,due to the traditional triangulation of traffic facilities map generation accuracy is not high,most of the error is about 1m,this paper generates traffic facilities map based on semantic point cloud extracting traffic facilities with post-processing.