Dynamical Pose Estimation with Graduated Non-Convexity for Outlier Robustness

Abstract

In this paper we develop a method for relative pose estimation for two sets of corresponding geometric primitives in 3D with a significant outlier fraction. This is done by using dynamical pose estimation as a solver in registration problems formulated with graduated non-convexity for truncated least squares (GNC-TLS). Dynamical pose estimation provides a unifying solver that can be used for point cloud registration, primitive registration, and absolute pose estimation. The solver is straightforward to implement, and it does not require specialized software for optimization. The main contribution of this paper is to show how the dynamical pose estimation method can be extended to fit into the GNC-TLS framework so that high outlier fractions can be handled. The proposed method is validated for point cloud registration, primitive registration, and absolute pose estimation. The accuracy and robustness to outliers is shown to be on the level of existing GNC-TLS methods.