Probabilistic Methods for Background Subtraction

Background subtraction belongs to core methods used in visual sensor network research, and has had a long history in computer vision. Its basic task is to identify locations in camera’s field of view that are changed compared to some template background. There are two fundamental issues in background subtraction: the computation of a template background and the comparison current video frame with the template background. In terms of the template background, it is usually computed by means of deterministic operators such as temporal mean or median over a number of video frames or  by means of statistical modeling such as non-parametric kernel density estimation (also known as Parzen window), mixture of Gaussians, etc. As far as the comparison method, it can vary from a simple luminance/color thresholding to binary hypothesis test with priors.

We have explored two approaches to probabilistic background subtraction:

• Foreground-Adaptive Background Subtraction: A novel approach explicitly modeling changed areas within a binary hypothesis test criterion.
• Background Subtraction with FDR Control: Adoption of False Discovery Rate (FDR) control algorithm to background subtraction.