Semantic Segmentation And Map Generation Based On 3D Point Cloud

The development of intelligent traffic system(ITS)is facing significant changes due to the in-depth integration of self-driving with cloud computing and artificial intelligence.For the core technologies of self-driving include perception,localization,high-definition map,planning and intelligent control.Among them,the high-definition map is the core technology for L4 and above self-driving.Scholars at home and abroad have made abundant accomplishments in the research of automatic generation methods for highdefinition map,however,the problems of low automation and slow update have not yet been solved.To address the automatic generation of environment layers in high-definition maps based on 3D point cloud,this paper uses a deep learning based on point cloud semantic segmentation method for large scene point clouds and extracted features through random sampling and local feature aggregation to conduct accurate segmentation for large scene point clouds.To address automatic generation of road layers in high-definition maps,this paper uses a registration algorithm based on normal distributions transform(NDT)and graph optimization method for the scene mapping,according to the segmentation results of ground points,it solves road edge error point detection and correction through line fitting and distance compensation.To address map storage,this paper uses opening source Open Street Map(OSM)format for vector storage of extracted road elements and validated.The specific research in this paper mainly contains four aspects as follows:(1)Automatic generation of environment layers in high-definition map based on point cloud semantic segmentation is realized.To address feature loss in point cloud feature extraction of existing methods,a method combining local feature aggregation and random sampling for local feature extraction is developed after reviewing and analyzing the existing literature in the field of point cloud semantic segmentation,and a semantic segmentation network for large scene point clouds is designed.Meanwhile,a campus environment dataset is made by collection,annotation and pre-processing data.The automatic segmentation for unknown campus scenes is conducted by the trained model.(2)Automatic generation of road layers in high-definition map is completed.Pose graphs of key frames are constructed by NDT registration algorithm.The point cloud frames are repeatedly observed to construct an initial global map with point cloud registration technology.To eliminate accumulates errors during matching process,a graph optimization method is used for the backend optimization.After extracting the map road point clouds through semantic segmentation,a line fitting method is used to extract road edges,and a method combining line fitting and distance compensation is used to conduct automatic location and correction of edge misdetection points and obstacle points.(3)Vectorized storage of road layer data is realized.The extracted road edge points are fitted and calculated to obtain the intermediate points,which are stored in vector format after transformed from Li DAR coordinate system to WGS84(World Geodetic System)coordinate system according to the format requirements of OSM map,and the road information is further consummated with JOSM software.(4)To verify the effectiveness of the methods used in this paper,those are compared with existing semantic segmentation methods under publicly available datasets and campus dataset collected at anytime.To verify the accuracy of the road information,the road visualization and map information distribution are completed under ROS operation system.