Because of the loss of depth information, distortion of lens and so on, all make the camera model nonlinear. RBF neural networks which can approximate any non-linear function are adopted to describe relations between 3-D feature points and corresponding image point in left and right cameras. While the network is trained, sum of squared differences between outputs of network and actual coordinates of corresponding feature point in world coordinate system is taken as performance index. Weights, basis width and central vectors of gauss function are tuned and achieve stable value by iteration until the performance index is less, all of which and activation function are equivalent to projection matrix of cameras, thus calibration of system is accomplished. Finally, precision analysis is carried out for system.