An effective approach to SLAM toward autonomous operation for a Leg/Arm composite mobile robot in unknown environment based on RGB-D images

Abstract

In this paper, an effective approach to Simultaneous Localization and Mapping (SLAM) based on RGB-D images is presented toward autonomous operation of a Leg/Arm Composite Mobile Robot (LACMR), in which depth information and its effective features are utilized sufficiently so as to overcome some malpractice in conventional methods and enhance the performance of SLAM. Our scheme can be narrated as follows. Firstly, in view of some feature-less regions in RGB image of operating scene, relevant stable and reliable depth information is employed to extract effective features. Secondly, the pose estimation and fine tuning is carried out by combining with features extracted from RGB-D images and proper successive matching of features. Thirdly the pose-graph optimization is conducted through the evaluation and selection of keyframes and loop closure detection to correct the drift and fine tuning the estimated pose. Finally, a dense 3D map is generated with keyframe information to serve for the autonomous operation. The accuracy and performance of the proposed SLAM approach is verified and validated by the map building and trajectory results with the TUM RGB-D dataset and our physical experiment results on a LACMR.