Direct model-based image motion segmentation for dynamic scene analysis

Abstract

Analysing the dynamic content of a scene observed by a mobile camera often requires the segmentation of each image of the sequence into region entities of apparent homogeneous motion. To each region is associated a 2D polynomial model (e.g., an affine one) able to describe at each location the underlying 2D “true” motion with a predefined precision η. Thanks to the use of a multiresolution robust estimator [1] to compute the motion models, the determination of the boundaries between the different regions, which is stated as a statistical regularization based on multiscale Markov Random Field (MRF) models, can be achieved in one pass only. This avoids the time consuming iterations between motion estimation and boundary identification that are encountered in almost all other motion-segmentation schemes (for instance [2, 3, 4]). We explicitly detect areas where the error between the underlying motion and the one given by the estimated models is not whithin the precision η. This allows us to handle the appearance of new objects in the scene. We have performed numerous experiments with real indoor and outdoor image sequences which demonstrate the efficiency of the method.

This study was supported in part by the French Ministry of Research in the context of the GDR-PRC “Man-Machine Communications” (Vision research program, MRT contract 91S269), and by “Région Bretagne” (Brittany Council) through a contribution to student grant.