Stereo visual odometry failure recovery using monocular techniques

Abstract

Stereo visual odometry is one of the most accurate dead-reckoning methods for estimating the motion of a moving vehicle but it strongly depends on a robust matching of the image features in the stereo frame. If a stereo camera is observing the environment from a critically small distance the two field of view can be subjected to poor or absent overlapping. That leads to failure of the computation pipeline because no stereo observations can be made. In this paper, we present a solution to this problem by taking advantage of monocular visual odometry techniques to propagate the pose estimations when the number of feature matches in the stereo frame is too low to produce accurate results. The proposed algorithm is tested on a challenging scenario for a stereo setup and a ground truth is given by mounting the stereo camera on a linear slide. Experimental results show that our algorithm is able to successfully recover failures of the stereo pipeline, obtaining a final position error of 1.2% of the total travelled path length in our dataset.