Research On Obstacle Detection And Tracking Algorithm Based On 3D Laser Point Cloud

Intelligent vehicles have become a research highlight in the field of vehicle engineering and new motive force of rapid growth in motor industry,it is also the trend of future automotive technology development,which puts forward higher requirements for the accurate acquisition of environmental information.Environment perception is the foundation of intelligent driving system and the precondition for path planning and vehicle control.The environment perception system is mainly composed of obstacle detection and tracking.In order to solve the problem of point cloud motion distortion in traditional obstacles detection and tracking algorithms,the difficulty of distinguishing adjacent obstacles and the easy splitting of distant obstacles,and the problem of false association in multi-obstacle tracking,obtain more accurate information about the surrounding environment and improve the driving safety of intelligent vehicles.In this paper,the traditional obstacle detection and tracking algorithm is improved,and the effectiveness and real-time performance of the improved algorithm are verified by real vehicle experiments.The main research contents include the following parts:(1)The working principles and types of lidar and inertial measurement unit are expounded,and perform joint calibration in time and space,analyze the principle of point cloud motion distortion and its impact on obstacle detection,the inter-frame motion of lidar is compensated by the method of fusing high-frequency IMU,and the motion distortion of point cloud is corrected.(2)Established the lidar coordinate system,vehicle coordinate system and the coordinate conversion relationship between them,and the point cloud data in the lidar coordinate system is converted into the vehicle coordinate system;In order to decrease the amount of point cloud data and raise the real-time performance of the system,the point cloud data is preprocessed,and the Ray Ground Filter algorithm is used to solve the under-segmentation problem in ground segmentation;For the traditional DBSCAN clustering algorithm,the parameter threshold is fixed,which makes it difficult to distinguish near-adjacent obstacles or long-distance obstacles are easily split into multiple obstacles and the problem that the algorithm takes a long time when the number of point clouds is large,this paper adopts an improved clustering algorithm of DBSCAN fusion region growth,and uses the L-shaped fitting method to complete the point cloud 3D bounding box fitting,the accuracy of this method for obstacle detection is compared with the traditional algorithm,it is improved by 6%,and the real-time performance is improved by 13%.(3)Aiming at the problem of large amount of data association calculation and easy to produce false association in multi-obstacle tracking,this paper improves the joint probability data association algorithm,and more parameter features of obstacles are introduced to further screen effective measurement values and reduce reduce the amount of calculation;The interactive multi-model is used to deal with the tracking problem of different motion models of obstacles,and the optimal state estimation of obstacles is obtained according to the interaction of different models,at the same time,the Unscented Kalman filter is used to update the motion state of obstacles,and the dynamic tracking of multiple obstacles is completed.(4)In this paper,the built Haval H7 intelligent vehicle is used as the experimental platform to collect data.Through a series of real vehicle experiments: point cloud dedistortion experiments,data preprocessing experiments,obstacle clustering experiments,data association experiments and state update experiments,the effectiveness and real-time performance of the improved obstacle detection and tracking algorithm in this paper are verified.