-
利用光场相机Lytro Illum对该方法进行验证。进行标定的实验系统如图9所示,包括光场相机和标定板。从图像传感器上记录的原始
$2 {\rm{D}}$ 图像中恢复$4 {\rm{D}}$ 光场$ L(s,t,u,v) $ ,并使用MATLAB toolbox LFToolbox V0.4[9]进行子孔径图像的提取。使用了相机提供的白图像来定位微透镜图像中心和矫正镜头的渐晕。提取到的四维光场有$ 15 \times 15 $ 个子孔径图像,每个子孔径图像有434 pixel×625 pixel。实验中,使用Lytro光场相机拍摄16个不同视角的棋盘格图片,该棋盘格有$ 12 \times 9 $ 个网格,标定板相邻特征点之间的距离为30.0 mm×30.0 mm。光场相机标定结果详见表1。该模型在非线性优化和畸变校正前后的重投影误差如图10所示。图10(a)中,主透镜边缘的重投影误差相比主透镜中间的重投影误差有较大浮动,而优化校正后的重投影误差大致相同,如图10(b)所示。文中改进的标定方法与Dansereau[9]等人的方法相比,均方根误差由0.363 mm降低到0.332 mm,精度提升
$ 8 \text{%} $ ,如图10(b)、(c)所示。另外,对于Dansereau[9]等人方法中采用的径向畸变模型,文中验证了采用多项式模型中双参数、单参数径向畸变模型以及除法模型,其标定精度不如三参数的径向畸变模型,也说明了原本径向模型的准确性。Parameter Initial value Optimized value H1,1 0.0009 0.0005 H1,3 0 −0.0004 H1,5 −0.0044 0.0812 H2,2 0.0009 0.0005 H2,4 0 −0.0005 H2,5 −0.0044 0.084 H3,1 0 −0.0011 H3,3 0.0018 0.0018 H3,5 −0.3443 −0.3436 H4,2 0 −0.0011 H4,4 0.0018 0.0018 H4,5 −0.3425 −0.3454 k1 - 0.1199 k2 - −0.0426 k3 - 1.4977 p1 - −0.0066 p2 - −0.0094 Table 1. Before and after optimization of light field camera parameters
Light field camera modeling and distortion correction improvement method
doi: 10.3788/IRLA20220326
- Received Date: 2022-05-12
- Rev Recd Date: 2022-08-27
- Publish Date: 2023-01-18
-
Key words:
- machine vision /
- light field camera /
- reprojection error /
- camera calibration /
- lens distortion
Abstract: As a new type of imaging system, the light field camera can directly obtain 3D information from a single exposure of the image. In order to make more sufficient and effective use of the angle and position information contained in the light field data, complete more accurate scene depth calculation, and thus improve the accuracy of the 3D reconstruction of the light field camera, it is necessary to establish accurate geometric modeling and precisely calibrate its model parameters. This method starts from the thin lens model and pinhole imaging model, the main lens is modeled as the thin lens model, the micro modeling for pinhole imaging model, combined with the two-parallel-plane model of the light field camera, each extracted feature point is associated with its ray in three-dimensional space, the physical meaning of each parameter in the internal reference matrix is explained in detail, as well as the process of determining the initial value in the process of calibration. Furthermore, based on the radial lens distortion model, the tangential lens distortion model and the nonlinear optimization method based on ray reprojection error are further applied to improve the calibration method of light field camera. The experimental results show that the RMS ray reprojection error of this method is 0.332 mm. Compared with the classical Dansereau calibration method, the ray reprojection error accuracy of the proposed method is improved by 8% after nonlinear optimization. The derivation process of scene points and specific pixels analyzed in detail in this method has important research significance for the calibration of optical field cameras, which lays the foundation for establishment of optical model and the initial calibration of light field cameras.