Camera Calibration and Precision Analysis Based on BP Neural Network

A new approach based on BP neural network for seeking-solution projection matrix of camera is presented in this thesis. The network structure is designed according to camera mathematical model, where there are 4 inputs composed of the sampled points´ coordinates in world coordinate system and constant 1, and 2 outputs are obtained, expected values of which are elements of constant vector [0, 0]^T. The sum of square of errors between the network outputs and constant vector is taken as performance index. The weight matrix between the input layer and hidden layer is tuned in the light of gradient descend method and can be achieved stable value while the performance index is reached to expected value, while two elements of weight matrix between hidden and output layer are coordinates of projected points in image plane and keep constant during the training. Thus projection matrix of cameras can be obtained from the weights between input layer and hidden layer of the neural network, and calibration of system is finished. Finally, the precision analysis is carried out for machine vision system.