Mismatch Removal of Visual Odometry using KLT danger-points tracking and suppression

Visual odometry (VO) is a technique to transform front-end visual observation to pose transformation. In simultaneous localization and mapping (SLAM) based on visual odometry, mismatch of features can lead to high uncertainty and inaccurate state estimation. Although RANSAC (RANdom SAmple Consensus) can reject the outlier with iterative sampling among all feature points, it only eliminate the mismatch instead of finding a better match to replace it. In this paper, we introduce an algorithm to reject the mismatch in visual odometry and find a better match if possible. Our approach start with a self-match of latest camera frame in order to detect the danger-point probably leading to mismatch for every feature. KLT (Kanade-Lucas-Tomasi) optical flow tracking method is used to predict the motion of danger-point in next frame, where we form a danger-area of mismatch. We additionally apply suppression in this area by adding an extra Hamming distance in Gaussian distribution to the points in the area. Therefore, mismatch can be removed with extra Hamming distance added. We integrate the algorithm on ROS (Robot Operating System) and record a series of video data sets. Then we apply our algorithm to the video stream and successfully remove the mismatch difficult to be rejected by RANSAC.