Evaluating Pose Estimation Methods for Stereo Visual Odometry on Robots

Abstract

Researchers have made significant progress in solving the stereo visual odometry problem, where a mobile robot uses stereo video imagery to estimate its pose, and optionally the world structure. In this paper, we focus on Structure- From-Motion methods that first develop an initial pose estimate and use it to reject outliers, and then refine that estimate in a non-linear optimization framework. We consider two classes of techniques to develop the initial pose estimate: Absolute Orientation methods, and Perspective-n-Point methods. To date, there has not been a comparative study of their performance on robot visual odometry tasks. We un- dertake such a study to measure the accuracy, repeatability, and robustness of these techniques for vehicles moving in indoor environments and in outdoor suburban roadways. Our results show that Perspective-n-Points methods out perform Abso- lute Orientation methods, with P3P being the best performing algorithm. This is particularly true when triangulation uncertainty is high due to wide Field of View lens and small stereo-rig baseline.