Visual Odometry Drift Reduction Using SYBA Descriptor and Feature Transformation

Abstract

Feature point matching is a critical step to visual odometry (VO) computation and many other vision applications. Frame-to-frame ego-motion drift caused by feature mismatching is the main challenge for VO. This paper presents a VO algorithm that uses a newly developed feature descriptor called synthetic basis descriptor to obtain accurate feature matching and reduce the drift. An initial estimate of the camera motion is calculated using matching feature pairs. Feature points in the current frame are then transformed to the next frame using this initial estimate of camera motion. The sample means between the matched points and the transformed points in the next frame are used to obtain the final estimate of camera motion to reduce the drift or re-projection error. Our algorithm uses a sliding window approach to extend feature transformation into subsequent frames to overcome the limitation of the short baseline nature of VO. The accuracy of the proposed system is evaluated and compared with competent VO methods along with ground truth (GPS + inertial measurement unit data).