Research On Camera Calibration Method With Neural Network

As the existence of the many unwanted factors, such as radial distortion and decentering distortion, the model of camera imaging is actually a nonlinear one. In order to successfully realize this kind of nonlinear mapping relationship between the 3D object points and their corresponding 2D image points, neural networks were and are still used. This dissertation laid emphasis on camera calibration methods based on neural network technique.Current researches on camera calibration using neural network can only keep their precision within a small space, and there has not been such a method that can solve the conflict between calibration precision and speed as the space extends. This dissertation first introduced the characteristics of each fields, camera calibration and neural network, respectively, found another way of combining the two and therefore a new calibration method, and made it possible for the neural network to focus on nonlinear mapping.Using the pinhole model, this dissertation proved that there are different imaging rules along the three axes (XW axis, YW axis and ZW axis) in the world coordinate, and a brand new calibration method was created to calibrate the camera along these three directions respectively, namely parallel calibration. Besides, a new normalization algorithm was proposed and better calibration precisions were gained along both XW axis and YW axis.Experiments and analysis were done to testify calibration methods proposed in this dissertation. After the analysis, we offered ways that may improve calibration precision. Experiment results show that parallel calibration method belongs with large scale calibration since it has extended the calibration scale while keeping both calibration precision and speed. Also, there comes the conclusion that calibration along ZW axis is critical to the whole large scale calibration since the standard deviation of reconstruction along ZW axis is much higher than those along XW axis and YW axis. Therefore, a calibration method that can improve calibration precision along ZW will be our next destination.