RGB-D sensor based real-time 6DoF-SLAM

Abstract

A robust 6 Dof SLAM algorithm using an RGB-D sensor has been developed, which can be used in autonomous navigation and localization for robots in an indoor scenario. The SLAM algorithm developed is a graph-based approach in which the nodes of the graph represents the position and orientation of the camera, and the edges of the graph represents the constraints between the nodes. To simplify calculation and achieve real-time application, SURF matching and RANSAC are first applied to consecutive frames from the RGB camera to obtain good corresponding keypoints. Afterwards, ICP is applied to the point cloud of these keypoints, calculating the transformation matrix, and thus obtaining visual odometry. The nodes, edges, and scenery data are then recorded in a graph. Finally, ELCH is applied to eliminate dead-reckoning upon loop detections to ensure global consistency. The resulting 3D model of the surrounding is presented in a point cloud format, consisting of both color and depth information.