Feature Tracking Using Epipolar Geometry for Ego-Motion Estimation

This paper proposes a method to improve ego-motion determination using latent semantic indexing for feature description and enhancing feature matching via an adaptive procedure. Binocular stereo images are processed using SURF to obtain feature descriptors, which are analyzed semantically using SVD to generate new feature descriptors. Matching criteria for image pairs are established based on epipolar geometry; the criteria are then adaptively adjusted between successive image pairs to achieve more reliable matches. Four types of experiments have been conducted to evaluate the proposed method. Experiment results demonstrated the effectiveness of the proposed approach in providing smaller numbers of initial match candidates with higher probability of correct matches, thus improving the efficiency and accuracy of ego-motion estimation.