基于RBF神经网络的双目立体视觉系统的研发

在摄像机标定的过程中,由于深度信息的丢失,摄像机镜头的畸变以及图像处理时的误差等因素影响,最终影响系统标定的准确度.本论文采用RBF神经网络的训练方法,开发出一套双目视觉系统,以匹配点在左右图像的坐标为网络的四路输入,通过网络得到三路输出,性能指标为网络的三路输出和该对应点在世界坐标系坐标差值的均方值;并根据梯度下降法来调整隐含神经元与输出神经元之间的连接权值,节点基宽参数与中心矢量,直到性能指标达到预定值,使其和激发函数共同组成具有投影矩阵功能的系统,完成双目视觉系统的标定.最后对系统进行精度分析.

Development of Binocular Stereovision System Based on RBF Neural Network

The loss of depth information,distortion of camera lens,errors caused by image processing and so on,influence accurate degree of systems calibration.In this paper radial basis function neural network training means was adopted to develop binocular vision systems.The operator is the coordinate image of a match point in left and right camera images used as 4 inputs,and the output through network are 3,and the mean square value between the outputs and actual coordinate of it in world coordinate system was taken as performance index.The weights,basis width parameters of joint and central vectors are tune in the light of gradient descent mean until the performance index is less,all of which and the activation function make up calibration systems of stereovision and is equivalent to projection matrix,thus calibration system is accomplished.Finally,the precision analysing is carried out for system.