Integrating Local Correlation and Interframe Continuity for Robust Foreground Object Segmentation

Per-pixel foreground segmentation methods are mostly based on the assumption that each pixel is independent (e.g. MOGs and HMMs), which makes the segmented foreground objects not integral. In this paper, a novel technique without this deficiency is proposed. Firstly, a binary map of initial foreground segmentation is achieved by performing Bayesian strategy according to spectral, spatial, and temporal features, where the foreground map is fragmented due to independence hypothesis among pixels. Secondly, pixels´ cross correlation in neighborhood of each foreground object patch is calculated considering spatial homogeneity. The pixels´ cross correlation between two frames regarding background interframe continuity and foreground discontinuity is also computed. Finally, pixels in each neighborhood are reclassified according to the above cross correlations in order to compensate small holes within foreground object. Experiments show that this method is robust in complicated background traffic scene video and can obtain more integral foreground objects