An Improved Method of Vehicle Ego-motion Estimation Based on Stereo Vision

Abstract

Intelligent vehicle is a senior development form in intelligent transportation system. Acquisition of accurate positioning information with cameras is a core technology in intelligent vehicles. In this paper, we propose a novel method to recover the vehicle trajectory using stereo vision technology. Our method takes advantage of structure and motion-based approaches, so the pose sequence can be computed without prior information on the structure of the relevant scene. In order to improve the efficiency and robustness, a robust feature detection procedure is conducted between image triples. Estimating vehicle ego-motion based on its epipolar geometric constraint makes the presented method not require the time-consuming reconstruction of 3-dimensional scene structure. An improved random sample consensus (RANSAC) algorithm based on geometrical constraints is also employed, which can effectively remove those outliers that are mismatched features or belong to moving objects. For nonlinear problems, an innovative Extended Kalman Filter (EKF) is then adopted to refine the estimated position. Overall, the improvements enable the algorithm to robustly estimate motion in a dynamic environment. Our experiments show that the improved visual odometry (VO) approach performs better than other state-of-the-art positioning methods in terms of computational complexity and accuracy.