An Adaptive Thresholding Method for Background Subtraction Based on Model Variation

Background subtraction is an important task in computer vision. Pixel-based methods have a high processing speed and low complexity. But when the video frame with camouflage problem is processed, this kind of methods usually output incomplete foreground. In addition, the parameters of many algorithms are invariable. These methods cannot tackle non-static background. In this paper, we present an adaptive background subtraction algorithm derived from ViBe. Gaussian Kernel template is used to model initialization and update. Standard deviation is used to measure background dynamics. We test our algorithm on a public dataset, named changedetection.net. The results show that we can handle most of scenarios. Compared to ViBe, we achieve better result generally, especially in dynamic background and camera jitter categories.