An accurate camera calibration method is required to achieve precise visual measurements,and traditional binocular calibration methods involve complicated mathematical models. As neural networks are effective for dealing with non-linear mapping, in the paper a BP neural network is proposed and implemented,which could satisfactorily describes the non-linear relations between 3-D characteristic points and their corresponding stereo images in the binocular vision. In order to improve the learning ability of the network,a dynamic gene is also introduced to the BP algorithm. Compared with traditional calibration methods, experimental results show that the proposed binocular calibration method based on neural network could obtain high accuracy.