Visual SLAM by Single-Camera Catadioptric Stereo

Abstract

In this paper, we present a robust simultaneous localization and mapping (SLAM) framework using a single camera catadioptric stereo system composed of vertically aligned two hyperboloidal mirrors and a CCD camera. In the SLAM, conventional stereo cameras have some problems due to their narrow field of view, for example, the data association problem which occurs in the featureless homogeneous regions or dynamic environments where moving persons or objects exist and error accumulation when the robot moves for a long time. However, a single camera catadioptric stereo system which gives not only a full horizontal field of view but the 3D locations of the landmarks helps to solve the above problems. For more accurate motion estimation, we propose the outlier detection algorithm to eliminate mismatched or incorrectly tracked features. We also propose the rectification algorithm that makes the mirrors and a camera parallel each other to satisfy single viewpoint (SVP). We analyze the proposed methodology through various experiments and have shown the robustness of the proposed SLAM algorithm