Adaptive Non-parametric Foreground Subtraction Using Rotation-invariant Local Frequency Pattern

Foreground subtraction in video surveillance is a very challenging task, especially in complex scenes. This paper presents a novel method for foreground object detection based on adaptive non-parametric model using rotation invariant local frequency pixel patterns. The aim of this method is to provide an efficient and robust foreground objects from videos which may contain illumination variations, camera gain and exposure variations, noises and dynamic background motions, etc. First, we propose an efficient scale-invariant local frequency pattern, which describes relations between adjacent pixels in frequency domain with illumination changes and shadows suppressed. Then, non-parametric model is used to estimate the probability distribution of pixel features in the process and an adaptive multi-modal framework is also introduced for learning. Extensive experimental evaluations on complex scenes of surveillance videos show that the method has obtained satisfactory results.