Novel approach of combining temporal segmentation results to the region-binding process for separating moving objects from still background

Abstract

To automatically segment moving objects in video sequences, FUB and ETRI have proposed several approaches to combine results provided by temporal segmentation methods (by FUB and UH) and spatial segmentation methods (by ETRI). In this paper, the authors present a novel approach that fuses temporal and spatial information during the process of segmentation rather than combine temporal and spatial segmentation results. The proposed approach is based on a region binding process, during which temporal segmentation results are integrated. That regions are distributedly represented and characterized distinguishes the region binding from the region merging and region growing. By fusing both temporal and spatial information, primitively segmented regions are bound to form the Binding-Cores (BC), whose role is similar to that of seeds in the region growing. Then the rest regions are bound to their neighboring BCs under strong or weak rules. The approach is composed of four stages. The experiment results show the performance of the approach.