Skip to Main Content
This paper describes an algorithm for visible/infrared fusion for video surveillance. Our technique combines signal-level and decision-level multispectral information fusion. We combine several techniques: we model observations in each spectral channel by a typical pixel-level mixture-of-Gaussian-based model; we model high level factors such as confidence of each modality, motion, object area, and lighting with a hierarchical probabilistic model. Our hierarchical model improves performance under global illumination changes, random parameter failures of background subtraction, occlusion and merge/split.