Study Of Filtering Algorithms Using Mobile Laser Scanning Point Clouds In Urban Area

Mobile laser scanning(MLS),as a mobile data collector which can acquire highresolution 3D model of the real world,has become an indispensable technology in driverless system,auto-navigation and road-level mapping.Point clouds filtering is a key step in LiDAR data processing and also the basis for objects recognition,terrain modeling and environmental perception.This thesis aims to accurately recognize ground surface from point clouds.To fulfill this objective,we proposed several new filtering algorithms which fully consider the features of vehicle-based data,including the large-scale coverage,abundant objects and point density variation.To evaluate the performance of these algorithm,they were applied to the data collected in real scenario.Additionally,we also explore the automatic extraction method of road surface from mobile LiDAR data from the perspective of functionality.This study provides several new ideas for ground type points recognition in urban.The main contents and conclusions of the thesis are as follows:(1)Regular voxel growth filtering.Octree structure is adopted to voxelize point cloud and form a 26 adjacency neighborhood space.Based on upward growth rule,ground points are filtered by setting rational threshold of elevation.The experimental result shows that this algorithm has merits of simple parameter,high precision and strong adaptability.(2)Super-voxel growth filtering.On the basis of regular voxelization,VCCS(Voxel Cloud Connectivity Segmentation)is applied to segment point clouds into supervoxels which is more close to objects boundary.Regional growth algorithm with ground features are added to get a final connective surface.Compared with regular voxel based filtering,this proposed algorithm is invulnerable to topographical change and has an outstanding performance in conserving objects.(3)Ground recognition through supervision classification.Geometric features are calculated based on supervoxels and SVM model is chosen as the classier to label each units.The method has achieved a good result with accuracy over 87%.Meanwhile,this method is invulnerable to occlusion and point density variation.(4)Road segmentation from high-density point coulds.At first,the road is segmented into tiles.For each tile,plane detection is carried out to give a rough road extraction.Then,both geometric and intensity information of points are utilized to recognize the road boundary.After fitting these discrete boundary points,the final road surface is determined.Experiment shows that the proposed method can extract road accurately in high way with no curb.