Skip to Main Content
Foreground segmentation is a very difficult task in dynamic background, which is common in real-world environments, caused by for example waving foliage, rippling water or illumination changes owing to light switching etc. A large number of methods for foreground segmentation in dynamic background have been proposed in the past decades, but most of them can only handle repetitive movements or gradual changes in background, and fail when sudden illumination changes occur. This study proposes a new method, which can deal with both repetitive movements and gradual/sudden illumination changes. The authors use a two-layer Gaussian mixture model to represent the background under different lighting conditions and formulate a joint posterior function of background state and segmentation based on the learned model. Given a new image, the background state and foreground segmentation are simultaneously optimised in a Bayesian perspective using a nested two-layer optimisation. The authors test their method on several image sequences and compare the results qualitatively and quantitatively with some state-of-the-art methods to demonstrate the effectiveness of the method.