Research On Improvement Of RGBD-SLAM With Corresponding Spheres Extending

In recent years, there has been a great interest in SLAM(simultaneous localization and mapping). A SLAM system solves the problem that a robot needs to know its location in the world in order to navigate from an unknown location in an unknown environment by building a incremental map and estimating the robot's trajectory at the same time[1]. The ability to do real-time 3D(Three-Dimension) mapping is the basis of a robot to perform tasks—using carried sensors, quickly generating related map and then completing the task. This thesis studies the real-time 3D reconstruction algorithm based on an RGBD camera in SLAM. The main research results are as follows:1. The general framework of SLAM system based on an RGBD camera is analyzed in detail. The block diagram mainly includes three modules: Front end — data acquisition based on an RGBD camera and the point clouds sequence registration, back-end —the result optimization of the registration by the closed-loop detection and graph optimization, and rendering — the creation of a 3D occupation probability map(that is, where is free and where is occupied) to navigate and complete the task.2. The searching and matching of the corresponding observation points are studied. We should get enough corresponding points before solve the rigid transformation between a pair of point clouds. We can detect feature points in different point cloud as observation points. Though there is no good algorithm on the 3D feature detection and feature description, we can get the corresponding 3D feature points indirectly by 2D feature points, and then we utilize RANSAC(random sampling consistency) estimation algorithm for punishing matching points.3. The Corresponding Spheres Extending, which is suitable for RGBD-SLAM system, is implemented. This module mainly utilizes the space structure of point cloud to punish the error matching points and extend the rest, which eventually improves the registration results. The experimental results show that this method is feasible.4. This paper puts forward CSE-RGBDSLAM(RGBD-SLAM System Combined with Corresponding Spheres). When the texture information in target scene is very little or similar objects occur repeatedly, acquired image data have little feature points or feature matching gets errors. Therefore, CSE-RGBDSLAM is presented. In the corresponding feature matching module of this system, a large number of corresponding points uniformly distributed are got by punishing and extending the matching feature points. And then we carry out RANSAC algorithm on the rough corresponding observation points. Finally, a large number of fine corresponding points are achieved.