| Engineering vehicles often travel on unstructured terrain.Therefore,fast and accurate reconstruction of complex unstructured terrain and recognition of its features are the premise for the realization of autonomous safe and stable driving of Engineering vehicles.This paper focus on Lidar perception of unstructured terrain,aiming at providing reference and basis for the research of unstructured terrain perception of driverless vehicles.According to the characteristics of unstructured terrain,this paper primarily solved the problems of accurate three-dimensional point cloud reconstruction of unstructured terrain,efficient preprocessing of point cloud data and accurate recognition and classification of terrain obstacle.The specific research work of the paper is as follow:(1)Realize the accurate 3-D point cloud reconstruction of unstructured terrain.According to the characteristics of unstructured terrain point clouds with uneven density and strong noise,a two-step registratalgorithm ion method including rough registration and accurate registration is proposed in this paper.Rough registration uses multisensor data combined with improved dead-reckoning to achieve efficient rough matching.Then an improved ICP algorithm based on neighborhood information is developed for accurate registration.Moreover,a multi-sensor data acquisition and processing system is set up to acquire and process all kinds of data in real time,which combined with the two-step registration method to realize the three-dimensional point cloud reconstruction of terrain.(2)Research on the pretreatment method of unstructured terrain point cloud.In order to improve the quality of terrain point clouds and enhance the effect of feature extraction,a combined filtering method is first constructed to realize the density homogenization,denoising and outlier removal of point clouds.Then an algorithm is proposed for extracting the boundary of the void region of point cloud.Finally,based on the idea of triangle construction,filling points of void region are generated from the boundary.(3)Extraction,segmentation and classification of obstacles in unstructured terrain from the terrain point cloud.Firstly,through the feasibility analysis and the introduction of Gauss kernel function estimation,the image edge detection and optimization method is applied to extract point cloud obstacles.Then,considering the interference of noise,the hypervoxel segmentation method is integrated into the Euclidean clustering method to realize the segmentation of single obstacle.Finally,based on the analysis of obstacle section,a fast and accurate method for judging convexity and concavity of obstacle is proposed.In summary,this paper covers the whole process of terrain perception from terrain data acquisition,preprocessing to feature extraction.The key problems such as inaccurate terrain data,excessive interference information and difficulty in extracting and classifying obstacles caused by the characteristics of unstructured terrain are solved,and the potential dangerous areas in unstructured terrain are identified accurately,which provides a new idea and reference for the unstructured terrain perception of unmanned vehicles. |
