Accurate Stereo Visual Odometry Based on Keypoint Selection

Feature association is a core issue in feature-based Visual Odometry methods. In this paper, we present a novel approach for stereo Visual Odometry, based on a careful feature selection. The proposed method relies on a circular matching for feature selection using spatial and temporal information. The process combines the Illumination Normalized SAD metric for stereo matching and the KLT algorithm for feature tracking. In both approaches, we explore the epipolar geometry constraints to get a fast and accurate feature correspondence. Experimental results demonstrate that our method achieves a local accuracy comparable to state-of-the-art techniques on the KITTI benchmark. Furthermore, even without global optimizations, the proposed method demonstrated to be accurate for long term tracking.