Improvement of 3-D position accuracy derived from weak perspective stereo vision using perceptron model

In this paper, we show you an approach for estimating a calibration matrix of an affine stereo rig using a perceptron model. A pinhole camera model is first established, then an affine stereo is used, which simplifies the calibration task, and leaves less parameters to be calculated than in a full perspective case. More reliable results can often be obtained by using an affine approximation in a lot of real industrial applications. As in most of estimation problems, the matrix of the stereo rig is ill-conditioned, the result of estimation is even worse when the number of input data are limited and noisy. To improve the precision of the elements in the calibration matrix, we use a perceptron model by applying iteratively the learning data. This approach can be extended to full perspective case, more general neural network model with back-propagation should be studied in the future.