Multispectral Foreground Detection via Robust Cross-Modal Low-Rank Decomposition

In this paper, we propose a novel approach which pursues cross-modal low-rank decomposition for robust multi-spectral foreground detection. For each spectrum, we employ the idea of low-rank and sparse decomposition to detect sparse moving objects against background with low-rank structure for its robustness to noises. Unlike simply combining multi-modal detecting results or compulsively enforcing a shared foreground mask in existing methods, we propose to pursue the cross modality consistency among heterogeneous modalities by introducing a soft cross-modality consistent constraint to the multi-modal low-rank decomposition model. Extensive experiments on the benchmark dataset GTFD suggest that our approach achieves superior performance over the state-of-the-art algorithms.