Motion objects segmentation based on structural similarity background modelling

Abstract

It is important to efficiently segment motion objects from video in computer vision applications. A novel foreground segmentation approach has been developed based on structural similarity background modelling, which responds quickly to sudden illumination changes and dynamic background. Both structural similarity map and environmental variation parameters are taken as a dynamic feedback controller to update the background. A multi-modal features fusion strategy has been proposed to segment foregrounds in a dynamic cluttered scene without any hypothesis for the scenario content in advance. Experiments for videos with some challenging content have been performed. Comparative study with state-of-the-art methods has indicated the superior performance of the proposed method.