Vehicle Detection Of Autodriving Based On Multi-Line Lidar

With the development of technology such as depth sensor and target detection,more and more scientific research institutions and commercial companies are paying attention to the unmanned vehicle system,and the necessary environmental perception ability of the unmanned vehicle system is the guarantee of the safe driving of the vehicle.In the aspect of sensors commonly used in unmanned driving,LiDAR has the main advantages such as less environmental interference and more accurate depth information compared with cameras.In the aspect of the design of unmanned vehicle detection algorithm,the detection based on deep learning has the advantages of automatic learning and higher accuracy compared with the traditional detection.Therefore,the main content of this paper is to design the vehicle perception algorithm based on LiDAR and deep learning.There are three main tasks for the above:point cloud processing,data set establishment and deep learning network design.Point cloud data processing mainly removes point cloud distortion caused by vehicle motion.The correction is mainly based on the pose obtained by the OXTS system.It corrects the point cloud by interpolation.The task of building the data set is to label the data collected by the vehicle platform for training and deployment.The deep learning perception network adopts the current two-stage framework.In this dissertation,a LiDAR-Region Proposal Network and Shape-RCNN are designed to generate 3d proposals and conduct secondary classification and bounding box regression of the 3d proposals,respectively.LiDAR-RPN is to project the point cloud to the top view,and then use the designed feature extractor to extract the features of each grid.According to the transfer learning theory of deep learning,in this dissertation,the technology commonly used in 2D image detection is transferred to the top view detection of point cloud,which can quickly detect the potential target area in the point cloud.The recall rate of region proposal network is up to Recall₅₀^97%.In the second part,the Shape-RCNN takes voxel as input and uses the three-valued feature extractor trained in this paper to extract shape features.Experiments show that this detector can better extract 3D shape information of the target,and it can effectively and accurately regress the size of the target,yaw angle and space coordinates.