RBF网络在立体视觉系统中的研究

摄像机标定、立体校正以及三维表面重建是立体视觉研究的重要内容。论文充分利用RBF网络的泛函逼近以及插值能力,将其应用于以上三个方面。在摄像机标定过程,通过将标定平面放置在有效视场内的多个位置,得到一组完备的样本,经过RBF网训练后,将立体视觉的几何成像关系存储于网络中;在立体校正过程,利用极线性质,由RBF网络确定图中的一组极线,然后通过求解极值问题来确定极点位置,最后用优化方法解出校正变换矩阵;在三维重建过程,利用摄像机标定中建立的视觉模型,重建出与图像信息相一致的三维表面。与传统方法相比,该算法具有重建速度快,运算精度高,过程简易明了等优点。通过对实际的视觉系统进行实验,证明了该算法的正确性和有效性。

The Research of RBF Network in Stereovision System

Camera calibration ,s tereo rectification and three-dimensional reconstruction are the important con tents of stereovision research.In this paper,RBF network(RBFN)is adopted to provide effective methodologies for the above three aspects through fully using its universal approximation and generalization properties.In the process of cam era calibration,a complete sample set is acquired through placing a calibration plane in several positions.Then the image geometry relationship of stereovisio n system is memorized in the network after training the nets.In the process of stereo rectification,RBFN is introduced to obtain a group of epipoles in the im age,then the epipolar is acquired through solving the minimum extremum problem .Finally the homographies are obtained by the optimal method.In the process of three-dimensional reconstruction,the object's figuration consistent with ima ge information is reconstructed based on the vision model acquired in the camera calibration stage.The algorithm has advantage in terms of accuracy,speed and versatility over the traditional approaches.It is proved through the experimen ts on the real vision system that the method of the thesis is correct and effici ent.