Improved RGB-D Camera-based SLAM System for Mobil Robots

Abstract

The paper presents an improved Simultaneous Localization and Mapping (iSLAM) and 3D reconstruction system for a mobile robot carrying an RGB-D camera. The developed system generates a three-dimensional point cloud from the color and depth camera data of the RGB-D camera. The matching is performed on successive point clouds that partially overlap. After feature detection on the color camera images, the method selects the 3D points in the successive point clouds that belong together. During the iterative multi-step matching algorithm based on Singular Value Decomposition (ISVD), it minimizes the matching error between the selected point clouds and deletes non-real point pairs in the process. Our previous SLAM method has been improved in several ways. On the one hand, a conditional averaging filter (Distance Image Filter: DIF) was created for the depth camera data to reduce the noise. The matching algorithm iteratively determines the matching transformation and the estimated displacement from the feature points of several recent point cloud segments. It defines a parameter for the accuracy and quality of each matching and includes the sub-results of pairs in the final displacement estimate by weighting these accuracy parameters. In this way, the algorithm yields significantly improved accuracy values, which in all cases are of comparable magnitude to those of methods in the literature, and for some published test sets exceed their characteristics. Since parallel programming methods are used to run the fits to previous states, the operation runtime remains fast. If the robot returns to its previous location, the improved loop closure detection method detects this fact, refines the estimates, and improves the accuracy. Finally, the proposed system also produces a 3D point cloud of the environment.