Self-Calibration Of A Vehicle-Mounted Fish-Eye Camera

As the fish-eye camera can capture the scene information of a very large field of view, it is widely used on the vehicle. But because of their special optical structure, the fish-eye camera has serious distortion, especially radial distortion. This problem arouses inconvenience for further applications on fish-eye images. Therefore, it is necessary to calibrate the fish-eye camera. Fish-eye images captured by a fish-eye camera on a vehicle have rich feature information of the scene. This paper presents self-calibration methods of a vehicle-mounted fish-eye camera based on the geometric constraints of feature points in fish-eye images.Most of generic cameras adhere to the pinhole camera model. However, fish-eye cameras have noticeable radial distortion. The pinhole camera model is not suitable for them. Several authors have developed models that can be used to describe the fish-eye cameras. This paper analyzes four kinds of fish-eye camera projection models which fish-eye cameras are usually designed to obey, and then introduces the generic polynomial model used in Open CV with equidistance projection assumption. Finally, we propose a novel fish-eye camera model which is suitable for self-calibration. Using this fish-eye camera model, it is easy to get undistorted points from distorted points in fish-eye images. This model is also a polynomial model using the equidistance projection assumption. Simulation results show the validity of the proposed model.We aim to find the fish-eye intrinsic parameters and distortion coefficients. The intrinsic parameters includ distortion center and focal length. The fish-eye camera parameters are computed through solving nonlinear constraint equations using the nonlinear optimization method. In order to prevent optimization process converging at wrong solution, we need to get good initial values of distortion center and focal length. This paper focus on the fish-eye camera with circular contour in the image. Firstly, we extract points in the contour. Then, we use the RANSAC algorithm to fit a circle, and let the center of the circle as the distortion center. The initial value of focal length is calculated based on a geometric constraint of collinear points under the equidistance projection model. The paper analyzes the algorithm for solving initial values, and experimental results illustrate that these methods are useful.The self-calibration algorithm of a fish-eye camera is proposed based on geometric constraints of the feature points in fish-eye images. We use SIFT algorithm to extract and match feature points in fish-eye images, and use LK optical flow to track these points. Then, we present three algorithms to estimate fish-eye camera parameters. One is based on fundamental matrix constraints. One is based on homography matrix constraints induced by the space plane between two fish-eye images. And the last is based on collinear feature points. We use the nonlinear optimization method to solve equations produced by these geometric constraints, so we can obtain accurate fish-eye parameters. Several experiments with real fish-eye cameras have been used to test the proposed techniques, and good results have been obtained. These results demonstrate that the self-calibration algorithms of the fish-eye camera proposed in this paper are simple, accurate, and practical.