一种基于神经网络的立体校正方法  

A New Stereo Rectification Algorithm Based on Neural Network

作　　者：胡海峰[1] 熊银根[1] 张萍[1] 

机构地区：[1]中山大学电子与通信工程系,广东广州510275

出　　处：《电路与系统学报》2003年第6期66-71,共6页Journal of Circuits and Systems

基　　金：广东省自然科学基金资助项目(001172)

摘　　要：提出了一种简便的立体校正算法。该方法巧妙地运用投影成像原理,构造出一组立体图对在校正前后的对应点,将其作为样本来训练BP神经网络,从而将整个校正关系存储到网络中。利用训练好的神经网络,可以快速有效地实现立体校正。与传统方法相比,该算法的特点是:(1)校正过程中没有利用到摄像机的内外参数,因此无需进行复杂的摄像机定标运算;(2)校正后的图像大小与极点所在位置无关,不会出现校正图像无界的情况。通过实际应用,验证了该方法的正确性和有效性。A simple stereo rectification algorithm based on neural network is presented. From the principle of projective geometry, a group of corresponding image points between original and rectified stereo pairs are determined, These correspondent points then used as samples to train a BP network. Finally the stereo rectification can be realized quickly through using the trained network. This algorithm has two traits compared with traditional approaches. Firstly, the camera parameters are not neccessry to be determined and the complicated camera calibration is avoided. Secondly, the size of rectified image is independent of the position of epipole and will not be unbounded. It is proved through application that the proposed algorithm is correct and reliable.

关 键 词：立体校正 BP神经网络 极线约束 对应点问题 

分 类 号：TP242.61[自动化与计算机技术—检测技术与自动化装置] TP183[自动化与计算机技术—控制科学与工程]