Mixed-state Markov models in image motion analysis

TOMÁS CRIVELLI; PATRICK BOUTHEMY; BRUNO CERNUSCHI FRÍAS; JIAN-FENG YAO

Machine Learning for Vision-Based Motion Analysis

Springer-Verlag London

Lugar: Londres; Año: 2011; p. 77 - 115

When analyzing motion observations extracted from image sequences onenotes that the histogram of the velocity magnitude at each pixel shows a large probabilitymass at zero velocity, while the rest of the motion values may be appropriatelymodeled with a continuous distribution. This suggests the introduction ofmixed-state random variables that have probability mass concentrated in discretestates, while they have a probability density over a continuous range of values. Inthe first part of the chapter,we give a comprehensive description of the theory behindmixed-state statistical models, in particular the development of mixed-state Markovmodels that permits to take into account spatial and temporal interaction. The presentationgeneralizes the case of simultaneous modeling of continuous values andany type of discrete symbolic states. For the second part, we present the applicationof mixed-state models to motion texture analysis.Motion textures correspond to theinstantaneous apparent motion maps extracted from dynamic textures. They depictmixed-state motion values with a discrete state at zero and a Gaussian distributionfor the rest. Mixed-state Markov random fields and mixed-state Markov chains aredefined and applied to motion texture recognition and tracking.