Road Environment Perception Based On 3D Lidar

Since the birth of the vehicle,people have begun to endless fantasy of unmanned driving.For unmanned driving,the difficulty is not to control the vehicle but how to let the computer replace the driver to observe the road environment.This is the goal of road environment perception technology.Specifically,road environment perception technology should be able to capture road environment information,segment ground information and various obstacles in the road and determine the location of its categories and obstacles.The improvement of road environment perception technology has great value for the research and practical application of unmanned technology.The paper designs a road environment perception system based on 3D lidar,including 3D point cloud information collection,point cloud information preprocessing,point cloud segmentation based on connected domain markers,and point cloud classification algorithm based on SVM.Based on improved Unet neural network voxel segmentation algorithm and obstacle location algorithm based on bounding box generation.The point cloud preprocessing part includes three parts:region of interest division,voxelization and ground point cloud segmentation.Massive point cloud data is processed through the division and voxelization of regions of interest,and the amount of data is greatly reduced.Then,based on the existing ground point cloud segmentation algorithm based on random sampling consistency and voxel grid elevation difference,the ground point cloud segmentation algorithm based on double threshold is designed according to its characteristics and point cloud information.The point cloud segmentation and filtering of the ground and slope in the road environment is realized.The segmentation of the point cloud of obstacles is implemented by two methods.One is the classification algorithm of obstacle point cloud segmentation based on connected domain markers and SVM.Combined with the characteristics of point cloud image,the multi-scale connected domain labeling algorithm is used to complete the point cloud segmentation of obstacles and reduce the over-segmentation phenomenon.Then,the SVM classifier is trained by selecting the geometric shape features of the obstacle point cloud and the laser intensity to realize the classification of the segmented obstacle point cloud.The second is a voxel segmentation algorithm based on the improved Unet neural network model.Analyze the characteristics of the two segmentation models Unet and Deeplab commonly used in the field of color images and their difficulties in point cloud images.Then combine the characteristics of the point cloud image to improve the Unet network model.At the same time,the binary image optimization method is designed to optimize the voxel segmentation effect of point cloud images.Point cloud obstacle location uses a bounding box based obstacle location algorithm.Firstly,the existing algorithm based on minimum rectangle and distance-based bounding box generation is analyzed,and its advantages and disadvantages are analyzed.Combined with the characteristics of obstacle point cloud,a boundary frame generation algorithm based on complement is designed.The positioning of the obstacle is achieved by calculating the geometric center of the bounding box.Finally,the road environment perception system based on 3D laser radar can realize the collection of road environment point cloud information on campus,the segmentation and classification of pedestrians,vehicles and other obstacles in the road environment,and the positioning of pedestrians and vehicle obstacles.The accuracy rate is as high as 92.8%.