A method for parameters estimation of motion and planar surfaces based on pixel intensity is proposed. Conventional approaches for above parameters estimation are based on detected edge or feature points correspondence. The proposed method is constructed by probability model using whole pixels from pixel intensity of input images. There is no need to detect feature points. In this paper, it is assumed that the labels of each pixels, which gives plane components, is regarded as MRF (Markov random field) to the labels of neighborhood pixels. These labels can be estimated based on MAP (maximum a posteriori) estimate, and simultaneously model parameters, in which motion parameters, planar surface parameters and hyper parameters are included, can be estimated by ML (maximum likelihood) estimation. For the purpose of pixel correspon4ence between the different input images. hierarchy probabilistic model, the upper level consists of the label belonging to the reference images, and the ower level consists of the label belonging to the input images, is proposed. To solve it, EM (Expectation Maximization) algorithm is used for model parameters estimation, and mean field approximation is used for calculating the expectation of the labels.
展开▼