PR - TN 2004 / 01133 Issued : 12 / 2004 Statistics of the visual normal flow

We describe statistical properties of the visual normal flow. These are used to model space-time texture, such as waterfalls, an open problem in visual motion analysis. The normal flow is perpendicular to edges. It is the only component of the image velocity that can be locally computed, so that it is simple to implement. We analysed the statistical representation of the normal flow magnitude for two regions, foreground and background. The log of the normal flow magnitude was discovered to exhibit a Gaussian fit and strong local correlations describing directional motion preferences. We build a statistical model used for region segmentation. This model is given by a prior representing a segmentation map plus a likelihood term of the log normal flow magnitude given this segmention map. Conclusions: Experiments using different videos with space-time texture resulted in good agreement with visual characteristics of these textured regions, such as, motion orientation.