Depth-Based Efficient PnP: A Rapid and Accurate Method for Camera Pose Estimation

Abstract

This letter presents a novel approach, DEPnP (Depth-based Efficient PnP), addressing the Perspective-n-Point (PnP) problem crucial in vision-based navigation and SLAM (Simultaneous Localization and Mapping) in robotics and automation, which estimates the pose of a calibrated camera by observing the 2D projections of known 3D points onto the camera image plane. The method employs eight variables to control the depth of control points and orientation of camera, formulating camera pose estimation as an optimization task. By optimizing these variables utilizing mean-subtracted rotation equations, rapid and accurate camera pose estimation is achieved. Notably, the careful selection of variables and objective function simplifies the computation of the Jacobian matrix, ensuring computational efficiency. DEPnP demonstrates robustness against noise and inlier disturbances, consistently delivering accurate camera pose estimation. Experimental evaluations validate the effectiveness and accuracy of DEPnP, positioning it as a competitive solution for real-time applications requiring precise camera pose estimation in robotics and automation. Our code has been open-sourced on GitHub.