Models for images and video foreground segmentation using finite mixtures of generalized Gaussians

Résumé

In this thesis, we deal with the foreground segmentation (FS) problem that is a key issue to numerous computer vision applications such as document analysis, object recognition, and video surveillance. Two problems can arise from the foreground segmentation depending on the input data: image foreground segmentation vs. video foreground segmentation. Many approaches have been proposed to address both of these problems, among of them is the histogram-based approach. However, most of the histogram-based approaches assume a unimodal Gaussian histogram shape for data classes and make use of parameters to estimate their distributions. Consequently, the efficiency these techniques could be affected by the fact that the data distribution is multi-modal and/or non-Gaussian. Moreover, addressing the problem of video foreground segmentation is not a trivial task especially in some challenging situations such as illumination changes, cast shadows, dynamic backgrounds, and pan-tilt-zoom (PTZ). These challenges have been widely studied to be able to make the current approaches more robust to such scenarios. To address the FS in images and videos, we use the mixture of generalized Gaussians (MoGG’s) for modeling the histogram of data (image or video). The merits of using the MoGG include: (i) An additional degree of freedom that controls its kurtosis. (ii) Histogram modes, ranging from sharply peaked to flat ones, can be accurately represented using this model. (iii) Skewed and multi-modal classes are explicitly represented using mixtures of GGDs. Furthermore, an online mixture of generalized Gaussian (MoGG) model is used to model the temporal information represented by the pixel history in image sequences. The former model is enriched by integrating temporal co-occurrence of background/foreground classes to deal with complex background dynamics. Besides, spatial analysis is introduced to deal with shadows, stopping objects, and PTZ camera effects. The proposed algorithms have been developed to run in real-world environmentswith near real-time performance. Experiments on the available datasets show that the proposed algorithms significantly enhance results (both qualitatively and quantitatively) compared to the other state-of-the-art techniques.

Citer ce document

Accès au document

Ce document est hébergé sur son dépôt institutionnel d'origine.

Auteur(s)

Statistiques

Consultations : 5

Téléchargements : 0