Research On Lidar-based Obstacle Detection Technology For Intelligent Vehicles

With the development of artificial intelligence technology,smart driving technology has received a lot of attention.An intelligent vehicle is an integrated system that combines functions such as environmental awareness,decision planning and multi-level assisted driving.Real-time detection of obstacles in the surrounding environment has been the focus of research in intelligent vehicle environment sensing technology.This paper focuses on LIDAR-based obstacle detection technology for the surroundings of smart cars.The working principle of LIDAR is analysed,and LIDAR calibration,acquisition of point cloud data,point cloud data pre-processing,obstacle detection and classification are studied and implemented.The main points are as follows:(1)LIDAR and other equipment were selected to build the autonomous driving platform and to complete the calibration of the LIDAR on the vehicle coordinate system.Data collection in the campus environment using the autopilot platform provides the data base for the following work.(2)Obstacle clustering via point cloud processing algorithms.First the body point cloud is removed and imported into the vehicle model,a Grid Map is created to delineate the ROI area and statistical filtering is used to remove the noise points.An algorithm for ground plane fitting is used to address the problem of under-segmentation of existing point cloud ground segmentation algorithms for sloped road conditions.The ground plane fitting algorithm fits some points with the lowest height value to a ground plane,and compares the projected distance between the remaining points and the fitted ground plane with the set threshold to distinguish whether the points are non-ground points.Aiming at the problem that the classical DBSCAN clustering algorithm has poor clustering effect when the data distribution is uneven,an improved DBSCAN clustering algorithm is proposed.Different parameters are used depending on the distance,thus improving the clustering of obstacle point clouds.Finally the minimum external enclosing frame is drawn for the obstacle point cloud.(3)Connecting traditional point cloud processing algorithms with neural networks for obstacle classification.The point cloud data,which has been ground segmented by conventional algorithms,is fed into the Apollo LIDAR based sensing module model via the ROS communication mechanism.The classification of surrounding obstacles into cars,bicycles and pedestrians is achieved and the distance between the center of the obstacle and the LIDAR can be detected.(4)Designing straight and curved road experiments in a campus environment and calculating the accuracy verifies that both the ground plane fitting algorithm and the improved DBSCAN clustering algorithm outperform the classical algorithm.The feasibility and stability of the obstacle classification algorithm was verified by designing real-world experiments under different weather conditions to calculate the accuracy,recall and F1＿Score.