Robust Line-Based Radial Distortion Estimation From a Single Image

Luwei Zhang,Jingjiang Xie,Hongbo Shang,Rui Wang,Fanlu Wu,Tao Sun

doi:10.1109/access.2019.2959204

Abstract

The pinhole model utilized in most computer vision algorithms becomes unfeasible because of lens distortion. Thus it is a must to compensate lens distortion to make the pinhole model available. In this paper, we propose a new robust line-based distortion estimation method to correct radial distortion. Our method works from a single image and is able to estimate the distortion center rather than assuming it is at the image center. Distortion parameters are estimated from parameters of circulars arcs, on the basis that straight lines are imaged as circular arcs under one-parameter division model. A new feature selection scheme by refining circular arcs is introduced to make the process of distortion estimation fully automatic and more robust. Moreover, a linear optimization algorithm is applied to calculating parameters in each selection run, making our feature selection scheme more efficient. Experiments on synthetic images and real images show that our method performs well in radial distortion estimation even with severely distorted images.

Highlights

The pinhole model that straight lines in 3D world must project to straight lines in the image plane is the basis for most computer vision algorithms [1]
DISTORTION MODELS AND DISTORTION OF A STRAIGHT LINE In general, radial distortion is expressed in the following format [21]
Benligiray and Topal [31] proposed a sequential backward selection scheme to make feature selection more robust. They utilized line orientation as the error function, and conducted nonlinear iterative optimization to minimize the error function. Inspired by their method but differently, we propose a feature selection scheme based on distance error to refine circular arcs sequentially, considering that distance error is less sensitive to step points existed in lines extracted from low resolution images

Summary

Introduction

The pinhole model that straight lines in 3D world must project to straight lines in the image plane is the basis for most computer vision algorithms [1]. The pinhole model is no longer available owing to mapping errors introduced by lens distortion. There are two main categories as to radial distortion estimation [7], that are multiple views method [5], [8]–[12], and single view method [13]–[17]. Multiple views method relies on known calibration objects [5], [8], sequence of images under camera motion or at different frames [9], [11], [12]. It is able to obtain distortion parameters and other camera parameters simultaneously

Methods

Results

Conclusion