Mantula O. Methods and models of multidimensional visual information forecasting

The thesis is devoted to the development of models and methods of multi-dimensional time series forecasting, which are induced by video streams. Besides the traditional problems of time series forecasting, their segmentation is considered, which provides 'semantic' structuring of video to improve the efficiency of information retrieval technology. The specifics of the proposed models and forecasting methods is orientation to on-line processing. A method of adaptive forecasting combining based on adaptive random search modification, adaptive forecasting based on nonlinear ANARX-additive model are proposed. Modified GMDH-neural network, which allows to improve extrapolating and approximating properties have been investigated. An adaptive method of nonlinear time series extrapolation with unevenly spaced observations is proposed and analyzed. Matrix models which represent fragments of images for extrapolation by the spread of adaptive identification procedures to the matrix case are considered. The results of experimental researches of video sequences forecasting directly in the image space and in feature spaces, which are produced by the spatial segmentation, and their using in the ecological monitoring tasks are shown.