双目视觉摄像机神经网络标定方法

Neural Network Technique in Camera Calibration for Binocular Vision

摄像机标定是精密视觉测量的基础。为了描述双目视觉中三维空间物点坐标和两个摄像机像面像点坐标间的非线性关系,传统的标定方法需要建立复杂的数学模型。而神经网络可以有效地处理非线性映射问题,笔者介绍了一种BP(ErrorBackPropagation)神经网络,并且为了提高网络的学习能力引入了动态因子。用相同的参考数据,将神经网络标定方法与线性标定方法比较,实验结果表明基于神经网络的双目视觉标定方法能获得较高的标定精度。

Accurate camera calibration is required for achieving precise visual measurements. In order to describe the non-linear relations between 3D geometry and stereo image point in the binocular vision, traditional calibration method involves many complicated mathematical models. Neural networks are effective for dealing with non-linear mapping. A BP （Error Back Propagation） neural network is proposed and implemented. A dynamic gene is introduced to the BP algorithm to improve the learning ability of the network. In comparison with linear calibration, for the same reference data, experiment results show that the proposed binocular calibration based on neural network could obtain high accuracy.