Full Range Stereo Vision System Based On A Single Camera Visual Odometry

Visual odometer can effectively overcome the deviation introduced by traditional odometer skid and other problems, which provides a foundation for visual navigation system and SLAM. But the recent vision-odometer system still has its limitation in its application in small payload intelligent robots, which is researched in this paper. An omni-stereo visual odometry in visual navigation is studied with a single-camera catadioptric omni-stereo system as imaging equipment.Based on the Matthies's vision-odometer framework (feature extraction, feature matching tracking, coordinate transform, the motion estimation), the main contribution of this paper is as follows:(1) The analysis of the imaging process is carried out for the single-camera omni-stereo imaging system, which consists of a conventional camera and two mirrors:one hyperboloidal mirror on the bottom and one hyperboloidal-planar combined mirror on the top. The unwrapping method for the acquired double-ring omni-directional image is derived.(2) Feature extraction methods, including Moravec, Harris and sift, are analyzed and compared based on experiments, so that the suitable one is used for the feature extraction and matching in the experiments with simulated images generated by POV-Ray.(3) Visual odometer is discussed in the proposed system. Based on the imaging process by the system, a method of 3D reconstruction corresponding to the feature in panoramic image is presented. The motion parameters are solved by the singular value decomposition method and the RANSAC method. The nonlinear least squares method is used to further optimize the motion parameters.The above algorithms are realized under the platform of Matlab, and tested with the image obtained by the POV-Ray simulation software. The results prove the effectiveness of the algorithms.