Stationary covariance realization with a specified distribution of amplitudes

The signals one encounters in examining image intensity data seldom appear to be even approximately Gaussian and as a consequence Gauss-Markov filtering theory, which vision researchers have found to be so useful in tracking and road following, has not been of much value in understanding the basic science involved in developing low level vision algorithms. We propose a methodology for stochastic modeling which allows one to explore a class of models better fitted to the distribution of values taken on by the data while maintaining the ability to fit the autocorrelation function