Research On Unstructured Road Boundaries Detection Methods Based On LiDAR

Although driverless technology has been successfully applied in structured road scenarios such as highways and closed parks,there are still technical problems in the identification and processing of unstructured roads.As an important part of unstructured road detection,road boundaries detection poses new technical challenges to the traditional unmanned environment perception based on computer vision.LiDAR,as an important sensor,has more advantages than vision in acquiring three-dimensional information of the road.Considering the road boundaries detection accuracy,point cloud processing time and economic cost,this paper proposes an unstructured road boundaries detection method based on 16-beam LiDAR 3D point cloud to achieve reliable boundaries detection.The main contents of this paper include:(1)The existing road detection methods based on LiDAR 3D point cloud are summarized.The advantages and disadvantages of the methods to filter out road points and methods to extract road edge points are analyzed in detail.The existing technical problems are pointed out that the unstructured road boundaries are blurred and the bumpy roads lead to the road points cannot be completely filtered out,which affects the next step,extraction of edge points.Much interference points that cannot be filtered make the fitting error of the final boundary lines larger.(2)Aiming at the problem that road points cannot be completely filtered,a double-grid algorithm that can adaptively generate thresholds is proposed.This method divides the 3D point cloud within the range into large and small grids according to the plane coordinates,and uses the constraints within the grids and between adjacent grids to filter out the road points.The threshold of the constraint condition is adaptively generated in the region by the improved Otsu method,which optimizes the filtering effect of the road point cloud.(3)Aiming at the low efficiency of the unstructured road boundaries detection algorithm,an improved KD-tree retrieval algorithm is proposed to extract road edge points.The non-road point cloud in each row of grids is divided into left and right subspaces,and then the candidate road edge points are further extracted according to the distance from the remaining points to the vehicle’s forward direction.Finally,the improved Random Sample Consensus algorithm is used to complete the road boundaries line fitting,which improves the accuracy of unstructured road boundaries detection.On this basis,the point cloud data of three different unstructured road scenes collected by 16-beam LiDAR are used for experimental verification.The percentage of frames that successfully detected boundary lines in each dataset was counted,and the detection rate reached 90.04%.The problem that road points cannot be accurately filtered due to blur and obstacle interference.The improved edge point extraction and edge line fitting method greatly reduces the influence of interference points on the results,and realizes reliable detection of unstructured road boundaries detection with low cost.