Monocular visual odometry based on optical flow and feature matching

Abstract

To solve the problem of real-time precise localization in CPS-denied places, a monocular visual odometry method based on optical flow tracking and feature matching is proposed. To speed up traditional pose estimation algorithm, the image sequences are classified into key frames and non-key frames. The conventional pipeline of feature detection and matching is utilized to process key frames, while utilizing Lucas Kanade optical flow to track the correspondences in non-key frames. To improve the robustness of the visual odometry method, a RANSAC-based outlier rejection scheme is applied in the phase of pose estimation. Moreover, a Kalman Filter based on the dynamic equation is designed to optimize the pose estimation. Experimental results demonstrate that proposed method can acquire the high accuracy of feature matching, while highlighting the real-time performance of optical flow tracking, which can meet the needs of real-time accurate localization in cities.