Pose estimation of a moving humanoid using Gauss-Newton optimization on a manifold

Abstract

Extracting the motion parameters of a moving humanoid is an important issue in machine vision. This is due to the need of emerging applications such as telepresence or robot navigation. In the context of telepresence, we consider a robot to be equipped with a stereo camera the pose of which has to be determined in a fast and robust way. The key issue is to compute the robust estimates of the (3times3) compatible essential matrices of the stereo cameras since the motion can then be uniquely estimated. In this paper, a robust technique that computes these entities is proposed under the assumption that the images are calibrated. The algorithm is based on the five-point relative pose problem using an optimization technique on a manifold. It determines these matrices by intersecting three essential manifolds. Test results show that the proposed method has a quadratic convergence rate and delivers accurate results. In addition, the algorithm has shown to exhibit better performance when compared to standard algorithms.