Abstract:  Acquisition of accurate positioning information is a core technology in mobile robots. This paper proposes a stereo visual odometry method based on robust features for robot’s autonomous highprecision positioning. First, a robust feature algorithm AcceleratedKAZE (AKAZE) is adopted to extract the interest points after comparation with other local invariant feature algorithms in three aspects: repeatability, accuracy and efficiency. Then, we present a robust feature matching strategy and the improved Random Sample Consensus(RANSAC) algorithm to remove the outliers which are mismatched features or on dynamic objects. Thus the proposed method can be applied to dynamic environment. Geometry constraint based fractionalstep egomotion estimation algorithm provides the accurate motion of camera rig. Last, we test the presented egomotion scheme on the benchmark datasets of Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI) and data captured on campus in a considerably cluttered environment, and compared with stateoftheart approaches. The proposed approach can restrain the error accumulation and satisfy the requirement of realtime and high positioning system.

Key words:  visual odometry, local invariant feature, random sample consensus, egomotion estimation