Research on omnidirectional ORB-SLAM2 for mobile robots

This paper proposes a method of using omnidirectional images as visual input to ORB-SLAM2 system. With it, the system can track ORB feature points in the whole 360°. Hence, compared to traditional monocular camera SLAM systems, the result will be much more robust since more ORB feature points can be extracted. To implement this scheme, the main task lies in processing the distorted omnidirectional image. In this paper, we describe an algorithm of turning the omnidirectional images into correct inputs to the ORB-SLAM2. It includes ORB feature extraction algorithm on sphere surface, area selection and image transformation. Additionally, to reduce the algorithm's time cost, we also add some new methods like pre-rotation estimation and optimization in feature points matching.