Study Of Visual Simultaneous Localization And Mapping Based On Kinect

While intelligent mobile robots need to meet the demand of navigation and localization in the environment by themselves, the simultaneous localization and mapping is believed to be key to realize fully automated guided robots. Recent years, the research and application of intelligent mobile home service robot become a research hotspot. To realize the advanced features such as recognition and scraping of indoor objects, the applications always need to get 3D map of the environment with texture and colour.RGB-D cameras like Kincet are widely used in indoor environment rebuilding for they can provide both vision and depth imformation of the scene. Most existing simultaneous localization and mapping systems with RGB-D sensors are constructed by frontend of graphic processing and backend of pose graph optimization.This thesis analyses and studies the key points which influence the performance of RGBD SLAM system on the basis of the exis ting solutions. In the processing of frontend, the system error of the depth data will greatly influence accuracy of the image motion estimation, while distribution of edge constraint force and accuracy of the loop closure are very important to the whole optimization results.Considering that error of the depth imformation gained by RGB-D sensors increases as the distance become larger, in this thesis the depth information is used within the 3D points alignment of RANSAC. The weight of closer points is improved within the statistic of RANSAC inl iers. And it dynamically adjusts threshold for judging inliers according to the depth imformation. Thus the accuracy and stability of interframe motion estimation. For the backend, weights of the constraint edges between poses in the graph are reasonably allocated afetr calculateing information matrix of edge by measuring the accuracy of corresponding motion estimation. Meanwhile, A model to eliminate wrong loop closure detection is put forward which can improve the optimization result. Finally the experiment of indoor 3D mapping in the laboratory is carried out to test the feasibility of the RGB-D SLAM system. Also in this thesis use RGB-D benchmark datasets are used to conduct a comparative experiment, which aims at verifying the feasibility to improving the SLAM system with dynamic 3D points alignment, model for calculating information matrix and model of eliminating wrong loop closure.