Research On Obstacle Detection Technology Based On 3D Data

The environment perception as the first part of automatic transmission,its core is to allow autonomous land vehicle(ALV)to accurately perceive and understand the driving situation of the vehicle itself and the surrounding environment.One of the basic tasks of environment perception is obstacle detection.This paper takes the obstacle detection of ALV as the research object,and adopts local height difference and multi-frame fusion technology to realize positive obstacle detection;adopts side-mounted 32-line Li DAR to achieve the detection of negative obstacles in unstructured environments;proposes the inverse depth line for point cloud segmentation.The detailed research work is summarized as follows:(1)In order to detect positive obstacles in unstructured environments,a positive obstacle detection method based on local height difference and multi-frame fusion is proposed.In data processing,the Li DAR data is rasterized to avoid processing a large amount of point-level data,thereby achieving the purpose of reducing computing time.Furthermore,local edge detection is performed on the height dimension to initially detect the positive obstacle grid.Using GPS positioning information and IMU posture information,using historical positive obstacle feature points to supplement the current positive obstacle feature points.By counting the number of occurrences of obstacles in consecutive frames,noise removal is performed.Finally,the positive obstacle is detected robustly.(2)In order to detect negative obstacles in unstructured environment,two 32-line Li DARs mounted on both sides of the top support are set to detect negative obstacles,and the feasibility of the system to detect negative obstacles is analyzed.By analyzing the scanning characteristics of the side-mounted 32-line Li DAR,the distance tomographic characteristics and density characteristics of the negative obstacle point cloud are proposed.The negative obstacle feature point group is extracted based on these two characteristic values.Then the geometric relationship between the three points in the negative obstacle feature point group under different terrains is analyzed.Looser constraints are set for the feature point group of negative obstacles,which greatly reduces the rate of missing detection of negative obstacles.Finally,the DBSCAN clustering algorithm is used to cluster the negative obstacle feature points to obtain the number and category of negative obstacles.(3)Aiming at the problem of point cloud segmentation,an inverse depth line(IDL)model is proposed.A 2D structure array is designed to store and index the point cloud,so that the point cloud changes from disorder to order,which is convenient for indexing and searching.Then,based on the coordinate conversion principle of Li DAR,IDL model is derived to estimate the road plane.Finally,an inverse depth segmentation(IDS)algorithm is proposed to classify obstacles and traversable regions.