ENHANCE STEREO VISUAL ODOMETRY PERFORMANCE BY REMOVING UNSTABLE FEATURES

Abstract

Visual odometry includes two important stages: 1) feature extraction and 2) pose estimation. The performance of visual odometry is dependent on the quality of features including the number of features, the percentage of the correct matching, and the location of detected features. Usually, RANSAC method has been used in pose estimation to remove outlier and select a good set of features that provide higher accuracy. However, in the case the higher wrong matches, the RANSAC seems to be failing. This article proposes the removing unstable feature method by deep learning-based object detection. The proposed method evaluated on the KITTI dataset shows a higher accuracy 6 - 8% compared to the conventional method.