运用投影反馈的神经网络摄像机标定  

Camera calibration using feedback regulation of neural network

作　　者：郭政业 罗延[2] 胡雯蔷[2] 朱李瑾[3] 

机构地区：[1]中国舰船研究设计中心,武汉430064 [2]华中科技大学软件学院,武汉430074 [3]中国地质大学图书馆,武汉430074

出　　处：《计算机应用研究》2015年第10期3179-3182,3195,共5页Application Research of Computers

基　　金：国家"973"计划资助项目(2013CB035805);国家青年自然科学基金资助项目(51205145)

摘　　要：针对计算机视觉跟踪处理中传统摄像机标定的径向与切向的非线性畸变带来了极大的误差,降低了实验数据的准确性,以及传统神经网络标定法的高冗余这两个问题,研究提出了运用投影反馈的神经网络摄像机标定算法。所运用的原理是首先对输入图像进行锐化边缘提取以及二值化,然后对图像的角点提取和投影变换参数求解,将所求值代入神经网络计算模型,并进行双重反馈调节,消除径向与切向的非线性畸变。最后将消除误差参数代入单应性矩阵,得到摄像机参数。实验结果表明,径向与切向的非线性畸变基本消除,同时运行程序的冗余性也明显减少。对于结果的分析更好地证明该算法解决了摄像机标定的精度问题,并且在高精度标定的同时保证了算法的实时性,为今后高精度实时计算机视觉系统的标定提供了一种新的方法。The purpose of the paper is to solve the two problems for computer vision tracking. One is radial distortion and tangential distortion problem of Zhang Zhengyou calibration which makes great error and reduces the accuracy of the experiment.Another one is high redundancy problem for traditional neural network calibration. From the research,this paper put forward the algorithm of camera calibration using feedback regulation of neural network. There were three steps for the procedures of the research. First,it sharpened and binary-converted primary image. Then,it detected the corner of the image and calculated the parameter of the image projection transformation,sent the result to the neural network model with the methods of double feedback adjustment to eliminate the distortion. Last,it sent the parameter without error to the homography matrix,came out the parameter of the camera. The result of the experiment shows that the radial distortion and tangential distortion are eliminated and the redundancy of the program is reduced a lot. In conclusion,the algorithm gives better solution for the accuracy of the camera calibration and keeps the real-time response of the system all at once. It is a new method for real-time response and high accuracy computer vision system in the future.

关 键 词：神经网络 投影变换 双重反馈调节 实时性 高精度 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]