Segmentations of spatio-temporal images by spatio-temporal Markov random field model

Abstract

There have been many successful researches on image segmentations that employ Markov Random Field model. However, most of them were interested in two-dimensional MRF, or spatial MRF, and very few researches are interested in three-dimensional MRF model. Generally, ‘three-dimensional’ have two meaning, that are spatially threedimensional and spatio-temporal. In this paper, we especially are interested in segmentations of spatio-temporal images which appears to be equivalent to tracking problem of moving objects such as vehicles etc. For that purpose, by extending usual two-dimensional MRF, we defined a dedicated three-dimensional MRF which we defined as Spatio-Temporal MRF model(S-T MRF). This S-T MRF models a tracking problem by determining labels of groups of pixels by referring to their texture and labeling correlations along the temporal axis as well as the x-y image axes. Although vehicles severely occlude each other in general traffic images, segmentation boundaries of vehicle regions will be determined precisely by this S-T MRF optimizing such boundaries through spatio-temporal images. Consequently, it was proved that the algorithm has performed 95% success of tracking in middle-angle image at an intersection and 91% success in low-angle and front-view images at a highway junction.