| In the past three decades,visual SLAM has become an increasingly popular research topic to meet the application requirements of mobile robots.As an indispensable part of visual SLAM,map construction has always been a research hotspot of many scholars at home and abroad.For different application scenarios,the map forms used in map construction are also different.With the rapid development of computer vision,3D maps are widely used because they contain abundant information.ORB-SLAM2 system based on feature points is a relatively perfect visual SLAM system,which can collect image information from the surrounding physical environment in real time and display it by constructing a three-dimensional dense point cloud.However,due to the large number of dense point clouds,it is difficult for SLAM system to complete real-time processing of large scenes.Therefore,it is an urgent problem to find a portable,real-time updating and easy-to-expand map form.Aiming at the problems of efficiency and redundancy in the construction of dense point cloud map,an octree map form based on tree hierarchy is introduced,and a process of constructing lightweight 3D spatial map combined with orb-slam2 key frame and Octree Data structure is designed,so that the constructed map has more comprehensive and fully compressible description of 3D spatial state The advantages of high flexibility and can be used for navigation processing.Finally,comparative experiments are designed for the mapping effects of octree with different types and resolutions.The experiments show that the octree map occupies very little space compared with the original point cloud,and the description of the scene is relatively good.Aiming at the problem that most filtering methods can not effectively and comprehensively remove noise in complex environment,firstly,the main noise factors in the process of point cloud map generation are analyzed,and then a combined filtering method integrating voxel,statistics and radius filtering is proposed based on PCL point cloud library to comprehensively remove the noise points generated in the process of point cloud generation.In order to optimize the order of combined filtering,comparative experiments with different specifications and parameters are set for the table point cloud model to determine the optimal filtering combination order.The experiments show that the filtering method greatly reduces the number of point clouds without destroying the original main structure,which lays a foundation for the subsequent lightweight map construction.In view of the problem that some noise points in the octree map affect the space occupied by the overall map,the filtering part of orb-slam2 map building is improved by using the best order combined filtering.In the ROS environment,the process of constructing octree map based on orb-slam2 system nodes is designed,which makes the constructed octree map more lightweight and can serve higher-level applications.Finally,the effectiveness of the proposed mapping method is verified on the standard data set,and the combined filtering experiment is carried out for the mapping effect of the overall system.The results show that the proposed scheme can obtain better map construction effect under different motion trajectories.At the same time,the combined filtered octree map does not destroy the original environmental structure,Compared with point cloud,the map is cleaner and more accurate,which provides a basis for subsequent t three-dimensional navigation. |
