Research On Obstacle Perception Method Based On Lidar

With the continuous development of intelligent driving technology,intelligent vehicles put forward more accurate requirements for the perception of driving environment information.As the ’ eyes ’ of intelligent vehicles,the environment perception system needs to provide accurate obstacle target state information in real time to ensure the safe driving of intelligent vehicles.Lidar has important research significance and application value in environmental perception because it can provide three-dimensional point cloud information in real time,and has the characteristics of wide detection range and strong anti-interference ability.Obstacle detection and tracking based on lidar point cloud is one of the key problems in the research of dynamic environment perception of intelligent vehicles.Therefore,according to the requirements of the environment perception system and the working characteristics of the lidar,a perception system for obstacle detection and tracking based on lidar is constructed.Aiming at the problems of under-segmentation of adjacent obstacles,over-segmentation of obstacles at different distances,and multi-target error-associated tracking in laser radar obstacle detection,this dissertation improves the obstacle perception algorithm based on laser radar,and verifies the feasibility and timeliness of the algorithm through real vehicle experiments.The main research contents are as follows :(1)Lidar calibration and point cloud preprocessing.In this dissertation,the multi-plane calibration method is used.The lidar scans three planes at the same time.The transformation matrix is obtained by sensing three different attitudes of the plane,and the external parameter calibration of multiple lidars is completed.According to the three different attitudes of the sensing plane,the transformation matrix is solved to complete the external parameter calibration of the lidar.In order to improve the computational efficiency,direct filtering and voxel filtering are used to reduce the amount of point cloud.(2)Ground point cloud segmentation.Aiming at the fact that the current ground segmentation algorithm cannot correctly segment the slope road surface,this dissertation proposes an effective ground point cloud segmentation method.According to the scanning characteristics of the lidar,a variable-size ring model is constructed,and the ground point cloud is obtained by plane extraction,and the extracted ground point cloud is finally estimated by designing the ground likelihood estimation.The real vehicle verification shows that the method can effectively segment the ground point clouds of different road conditions such as slopes,with high robustness and improved real-time performance.(3)Obstacle detection.Aiming at the problems of under-segmentation of adjacent obstacles and over-segmentation of obstacles at different distances in current obstacle detection algorithms,this dissertation proposes a Multi-constraint adaptive Density-Based Spatial Clustering of Application with Noise(MCA-DBSCAN).Firstly,the point cloud is stored as a distance image.In the first stage,the connected neighborhood labeling method is used for rough segmentation.In the second stage,the MCA-DBSCAN clustering algorithm is used to finely segment the multi-obstacle point cloud cluster obtained by rough segmentation to obtain a single obstacle point cloud cluster.The Oriented Bounding Box(OBB)is used to extract the shape features and heading of the obstacle to complete the obstacle detection.Real vehicle experiments show that the method can correctly segment obstacles at different distances and adjacent obstacles.Compared with the Grid-based DBSCAN(G-DBSCAN)clustering algorithm,the detection recall rate of MCA-DBSCAN clustering algorithm is increased by 9.9 %,and the computational efficiency of the algorithm is increased by 42.5 %.(4)Dynamic obstacle target tracking.Aiming at the complex and changeable driving environment,there will be problems such as error data association and computational explosion in target tracking.In this dissertation,a tracking gate based on multi-feature of the target is used to screen out the measurements that may be associated with the target.The joint probability data association algorithm and the extended Kalman filter algorithm are used to predict and estimate the dynamic obstacle state,and a trajectory tracker is designed to achieve stable tracking of obstacles.The stability and real-time performance of the tracking algorithm are proved by real vehicle experiments.