Abstract:

We introduce a frequency decomposition based background generation and subtraction method that explicitly harnesses the scene dynamics to improve segmentation. This allows us to correctly interpret scenes that would confound appearance-based algorithms by having high-variance background in the presence of low-contrast targets, specifically when the background pixels are well modeled as cyclostationary random processes. In other words, we can distinguish near-periodic temporal patterns induced by real-world physics: the motion of plants driven by wind, the action of waves on a beach, and the appearance of rotating objects. To capture the cyclostionary behavior of each pixel, we compute the frequency coefficients of the temporal variation of pixel intensity in moving windows. We maintain a background model that is composed of frequency coefficients, and we compare the background model with the current set of coefficients to obtain a distance map. To eliminate trail effect, we fuse the distance maps.