Background generation is a fundamental step for video surveillance. It is challenged by many effects such as illumination, intermittent motion and clutter. In this paper, we propose a robust background generation method which can estimate the background from frames which all contain foreground. Moreover, it has good robustness toward illumination, clutter and other effects. The value channel from the HSV color space is used to detect any illumination changes in order to initiate a procedure for retaining frames with close intensity. A Foreground Rate per Frame (FRF) measure is developed based on the motion pixels to reduce the effect of stationary foreground and dynamic background. To refine the motion pixels and reconstruct the background, a refinement algorithm is developed to select the frames which can lead to reliable background estimation. To confirm the performance of the proposed method, we hold experiments on the extensive benchmark SBMnet2016 which contains 79 videos and includes 8 different challenging effects. The proposed method ranks as the fourth , on average, compared to 30 state of art methods and it maintains competitive scores through all benchmark categories.
Read full abstract