Research On LiDAR-based Obstacle Detection Methods For Intelligent Vehicle

Intelligent vehicles integrate artificial intelligence,advanced sensor technology,computer control technology and so on,which perceives the environment by the sensors and on-board processor,controls the vehicle according to the perceived information to achieve autonomous driving,provides a novel solution for road traffic problems.Light detection and ranging(LiDAR)sensors play an indispensable role in environmental perception,because of their high ranging accuracy and the ability of obtaining precise position of targets,which can build more accurate environmental model around vehicles.Therefore,it is of great significance to detect the obstacle around the vehicle,extract the road boundary,detect the dynamic obstacle and determine the drivable area around the vehicle.Aiming at the problems of obstacle detection in environment perception for intelligent vehicles,this thesis focuses on the following several aspects:Firstly,the research on front road drivable area detection and road boundary extraction methods was described.Two-dimensional LiDAR was used to detect front road drivable area,the LiDAR scanning points were segmented according to the continuity in the road surface and the discontinuity between road surface and obstacles or road edge,in order to choose the road points to form a travelable area.The road boundary was detected using the three-dimensional LiDAR data based on the active contour model.The 3D LiDAR data was firstly projected to occupancy grids to transform the occupancy grids into binary image,and balloons Snake model was used to extract image boundary,thus extracting the road boundary.The methods of point clouds ground segmentation was then studied to solve the problem of rapid obstacle detection with 3D LiDAR data.Random sample consensus algorithm was used to segment the point cloud,with plane representing the ground model and the lowest point of the grid fitting the ground.To solve the problem of segmentation in slope or other uneven ground,two-dimensional Gaussian process regression incremental sample consensus algorithm was then used to select ground seed points from the lowest points in each grid,and the ground seed set was used to segment all the scanning points,thus to realize point clouds ground segmentation in different scenarios.Finally,the non-ground point is taken as the obstacle point to establish local obstacle grids map on the basis of the point clouds segmentation,to study the problem of dynamic obstacle detection.Local obstacle grids map was used to construct global static obstacle grids map based on Bayes rule,local obstacle grids map and global static obstacle grids map was compared to get the possible dynamic obstacle grids,which were then clustered based on the distance between the grid center to get the target,the minimum rectangular envelope method was used to extract the target centroid,and Bayes rule and target tracking method were combined to detect dynamic obstacles,to obtain the global static obstacle grid,position and velocity information of dynamic obstacle.A large amount of data was collected in the campus environment using the modified intelligent vehicle,the methods studied in this thesis were used to process these data,and the results verify the feasibility and effectiveness of the methods.