Scene Understanding And Semantic Mapping For Unmanned Ground Vehicles Using 3D Point Clouds

The perception and understanding of the surrounding environment are the foundation of UGV navigation and mapping. The purpose of this paper is to describe the surrounding environment of UGV with 3D laser point clouds and use semantic description to build large-scale outdoor topological semantic map. In order to realize the on-line semantic mapping, we must have a real-time scene understanding algorithm. In this paper, we propose the optimal depth and vector length graph model for 3D point cloud data. The model can transform 3D point clouds into 2D gray images, which are suitable for a variety of laser scanning methods, and have better features than traditional models. Understanding based on ODVL graph model is fast and accurate.ODVL images are divided into super pixels, and the 20 dimensional texture features are extracted. Local point clouds'shape description and orientation are added to the super pixels' features. Gentle-AdaBoost algorithm is used to classify the super pixels. Via analyzing point clouds with lower confidence, we complete the re-classification of low confidence regions. Then this paper propose a correction method based on semantic constraints, extracting the surface area of regions with higher confidence, and correcting the wrong points surrounded.Through the understanding of the scene, the semantic description can be obtained. Before semantic mapping, we need to extract the category, size, location and other information of each individual object. In this paper,3D point clouds of the left and right side of UGV are divided into the scene nodes according to the mapping rules, and the tracks of UGV are divided into the road nodes by the rules. Topological map of the outdoor environment is obtained by generating topological relations between the scene nodes and the road nodes. With driving, there are new environment to be perceived and understood. The nodes information in the semantic map needs to be updated in real time, and the topological structure will be adjusted. Finally, a topological semantic map for large-scale outdoor environment is generated. The accuracy of our algorithm is 0.868. Compared to the grid map and other maps, Semantic map contains a large amount of semantic information, and the generation process of it is fast and efficient.