Application of neural network with embedded orthogonal weight to calibration of camera's intrinsic and extrinsic parameters

作者：Ge, Dong-Yuan; Yao, Xi-Fan^*; Xiang, Wen-Jiang

来源：Optics and Precision Engineering, 2011, 19(11): 2782-2790.

DOI：10.3788/OPE.20111911.2782

• Calibration precision
• Camera calibration
• Genetic-annealing algorithms
• Homogeneous coordinates
• Intrinsic and extrinsic parameters
• Orthogonal neural networks
• Simulation annealing algorithm
• Structured neural networks

According to the intrinsic and extrinsic parameters of a camera to be needed in some applications for machine vision, a method based on the neural network with embedded orthogonal weight matrix is proposed for camera calibration. Firstly, the neural network with the embedded orthogonal weight matrix is structured, whose weights are corresponding to the camera extrinsic parameters and intrinsic parameters. Thus, the structured neural network coincides with the model of camera. The generation of orthogonal weight matrix is served as the last mutation operator in the iteration of genetic algorithm, and the performance index is the square of 2-norm of the difference between vector consisting of network's outputs and homogeneous coordinate of corresponding feature point projected in image plane. Meanwhile, a hybrid genetic-simulation annealing algorithm is introduced into the solving-programming. When the system comes to the equilibrium, the intrinsic and extrinsic parameters of the camera can be obtained in the light of network's weights. The simulated experiments illustrate that the proposed algorithm is robust, and has the advantages of high calibration precision, simplicity and clarity according to real experiments. It provides an effective solution scheme for camera calibration of intrinsic and extrinsic parameters.

• 出版日期2011-11
• 单位华南理工大学