Research On Dynamic Obstacle Detection Of 3D Lidar Based On Intelligent Vehicle

With the continuous innovation of autonomous driving technology,intelligent vehicles have become a research hotspot in today’s automotive industry.Among them,environmental perception,as the core technology of autonomous driving,is the key to realizing the automatic driving function of intelligent vehicles.With its high resolution,long detection distance and strong anti-interference ability,3D lidar is widely used in various intelligent vehicles and has become an important part of the intelligent vehicle environment perception system.It is aimed at the undersegmentation and oversegmentation problems of 3D lidar point cloud preprocessing,as well as the low detection rate of dynamic obstacles.In this thesis,the traditional pretreatment and dynamic obstacle detection methods are improved,and the improved detection results are applied to dynamic obstacle tracking,which verifies the effectiveness of the proposed method through real vehicles.Firstly,aiming at the undersegmentation and oversegmentation problems of the traditional laser point cloud ground segmentation method of 3D lidar in the complex road environment,this thesis presents a region-based ground segmentation method to segment and filter the ground point cloud in different arc-shaped areas.Experiments show that compared with the GPF method and the RANSAC method,the areabased ground segmentation method improves the segmentation rate by an average of 8ms and 19ms,respectively,and the method also shows strong robustness in segmentation efficiency.Secondly,after completing the ground point cloud segmentation,in order to optimize the traditional Euclidean clustering method for clustering non-ground point clouds,a Euclidean clustering method based on adaptive parameter joint angle constraint is presented to cluster non-ground point clouds.Different area thresholds are delineated by 3D lidar ray distance,and the angle threshold is superimposed for constraint,and the 3D bounding box fitting of the point cloud is performed by L-shaped fitting.Experiments show that the detection rate of vehicles is increased by 10.32% and 15.87% compared with the traditional Euclidean clustering method and DBSCAN clustering method,and the detection rate of pedestrians is increased by 5.91% and 10.53%,respectively.Finally,this thesis aims at the problem that there are many dynamic obstacle targets under urban road conditions,and it is easy to follow the target.The joint JPDA-UKF-IMM method is used to predict and track dynamic obstacles,the joint probabilistic data association(JPDA)method is used to track dynamic obstacles in complex environments,the motion state of obstacles is predicted and tracked by lossless Kalman filter(UKF),and the tracking problems of different obstacles are processed by IMM interactive multi-model,and the effectiveness of the tracking method in this thesis is verified by experiments.