Solving Generalized Pose Problem of Central and Non-central Cameras

Bin Li, Yang Shang, Banglei Guan, Shunkun Liang, Qifeng Yu

Abstract: Recent applications in robotics, augmented reality, autonomous navigation, and self-driving involve cameras beyond the pinhole model, like fisheye cameras, multi-camera rigs, and other non-central cameras. We propose a unified method of solving the generalized pose problem of the central and non-central cameras. We first show the possibility of solving the generalized pose problem with the classical PnP solver initially designed for the central camera. We derive the closed formula of the translation in the generalized pose problem and transfer the classical PnP solver to the generalized absolute pose solver by computing a new coefficient matrix. In such a way, a category of classical PnP solvers can work with the non-central camera. Our generalized solvers inherit properties of the original PnP solvers, such as accuracy, robustness, numerical stability, and computational efficiency. Experiments on both synthetic data and publicly available real datasets show that our generalized solvers offer state-of-the-art performance.