Research On Indoor Scene Location And Dense Mapping Based On 3D Vision

Simultaneous Location and Mapping(SLAM)can be described as: a mobile robot starts to move from an unknown location in a completely unknown environment,and uses its own sensors to perform data on the surrounding unfamiliar environment during the movement.Collecting and estimating the position so as to realize self-positioning and constructing incremental map.With the characteristics of low cost,large amount of information,and the ability to extract semantic information,visual SLAM has more advantages than laser SLAM.However,visual SLAM has many problems in the process of real-time positioning and mapping,such as the phenomenon of feature point extraction and clustering,less information,and lack of loop detection.In response to the above technical issues,based on the RGB-D camera image acquisition,this paper carries out the following research work:(1)Construct an image pyramid,use the quadtree method to homogenize the feature points,solve the redundancy problem of the bunching phenomenon in the ORB feature point extraction process,divide the image of each layer,and count each The number of feature points in the grid.If the number of feature points contained in the divided network is greater than 1,the grid will continue to be divided,the grid with the number of feature points is 0 will be deleted,and the grid with the number of feature points equal to 1 will not be continued.Grid division,count the number of feature points of all grids until the threshold condition is met.Chapter 3 will introduce this method in detail.(2)Non-linear optimization based on the pose graph is added to the back end to reduce the cumulative error caused by the front end solving the pose over time,and obtain a globally consistent trajectory.(3)To solve the problem of insufficient space occupation information in the navigation task of the sparse point cloud map,construct a dense map construction.(4)Aiming at the problems of information redundancy in the navigation tasks of dense point cloud maps,an octree map conversion method suitable for robot navigation is realized.Through experiments,it can be seen that after the feature points are homogenized by the quadtree division grid,the extracted feature points are more uniform than before the improvement,which solves the problem of extracting feature points redundancy.After adding the nonlinear optimization based on the pose map at the back end,the cumulative error of the trajectory becomes smaller,and a globally consistent trajectory map is obtained.Finally,after the optimized dense map is transformed into an octree map,a map with a smaller storage capacity is obtained.