A Robust Visual Localization Method With Unknown Focal Length Camera

Abstract

PnP problem is well researched in many fields, such as computer vision. It is considered the fundamental method to solve the key problems of robot SLAM. However, in pedestrian visual localization, uncalibrated PnP (UPnP), specifically PnP with unknown focal length (PnPf) is more suitable for solving the problem. Recently, a few researchers proposed some methods to alleviate this problem. However, the localization accuracy of the existing methods is not satisfied when image pixel noise is larger. In other words, RANSAC should be running before solving the PnPf problem to get less noisy input, which means the localization delay increases inevitably. In this paper, we propose a more robust method for solving the PnPf problem without the help of RANSAC based on Gröbner basis and convex optimization. We build a Gröbner basis solver on the offline stage with one instance in the prime field. Then, we substitute the coefficients with real value and find multiple solutions on the online stage. Finally, we construct a convex optimization program to seek the final robust solution to PnPf problem. The other purpose of this paper is to provide a second-level localization experience for the end-user. The simulation result shows that our method can give a localization solution, which is more reliable than benchmark methods by both synthetic and real data verified.

Monocular Vision Guidance for Unmanned Surface Vehicle Recovery

“…In contrast, monocular vision only requires a single camera, enabling implementation through the pinhole imaging principle, resulting in lower costs and convenient operation [18,19]. The commonly used methods for monocular vision positioning include the Perspective-n-Point (PNP) method [20], the 2 of 14 imaging model method [21], the data regression modeling method [22], and the geometric relationship method [23]. The imaging model method achieves positioning based on a similar triangle relationship and is typically utilized for forward vehicle positioning in the transportation field but is not suitable for directional shifts of water surface targets.…”

Fast and Accurate Pose Estimation with Unknown Focal Length Using Line Correspondences

“…When all the intrinsic parameters are known and all the extrinsic parameters are unknown, at least three 2D–3D point correspondences are needed to estimate the camera pose, which is called a P3P solver [ 28 ], with up to four solutions, and an additional constraint (e.g., one more 2D–3D point correspondence) is needed to determine the unique solution. When there are four 2D–3D point correspondences, the camera pose and an intrinsic parameter (e.g., the focal length) can be simultaneously estimated, which is called P4Pf [ 29 , 30 ]. When five 2D–3D point correspondences are present, the camera pose and three intrinsic parameters (e.g., the focal length and radial distortion) can be simultaneously estimated, which is called P5Pfr [ 31 ].…”

An Accurate and Robust Method for Absolute Pose Estimation with UAV Using RANSAC

“…In addition, some methods can estimate the pose and partial intrinsic camera parameters when more than three 2D–3D point correspondences can be given. Some methods can work well with cases where the focal length is unknown, and the size of the minimal subset is four, which are called P4Pf solvers [ 8 , 20 , 21 , 22 ]. Actually, four 2D–3D point correspondences give eight constraints; hence, some methods can work well with cases where the focal length and radial distortion are unknown, which are called P4Pfr solvers [ 23 , 24 ].…”