Abstract: We combine the traditional photogrammetry calibration with neural network in this paper to overcome instability and computation complexity due to nonlinearity factors such as aberration. Calibration parameters for computer vision are obtained based on neural networks, and the unknown point world coordinates calculated with a five-step method. In traditional photogrammetry calibration, a calibration block is needed, with some control points that require accurate determination of three dimensional coordinates. To achieve high accuracy, these points must be evenly distributed in a three dimensional space, and the coordinates measured with high precision. It is difficult to make the calibration block, especially a big one. We propose to move the calibration plate along its normal direction instead of using a calibration block. The moving plate is a three dimensional block. While achieving a three-dimensional effect, it significantly increases the number of control points. Furthermore, the plate is much easier to make than a three dimensional block. Experimental results show that the proposed neural network based method with moving calibration plate can provide high accuracy and stability

Key words: camera calibration, computer vision, neural network, calibration board