An application of MAP to change detection in moving video

Change detection in the presence of noise is one of the most important problems in video data processing. The traditional statistical models are based on Gaussian tests of interframe variations. The crucial thresholds, based on which the decision rules are made, are inevitably experimental because the Gaussian assumption is often unrealistic. We present a new approach to change detection by applying the maximum a posteriori (MAP) criterion which does not require selection of empirical thresholds. As an alternative to the previous pixel-based thresholding methods, a new change detection method formulated as an optimization problem is derived by modeling the video frame as a Markov random field (MRF) and applying the mean field theory (MFT). The change labeling of the pixels is translated into seeking the optimal configuration of the change map. Under the MRF assumption, the solution to this problem is obtained by minimizing the energy function associated with the MRF. By applying prior knowledge of the noise and smoothness constraints on the MRF's, we choose the first-order neighborhood system and design the potential functions that reflect the prior beliefs. The algorithm that computes the potentials is constructed by applying MFT, which greatly reduces the computational complexity. The experiments on several medical video sequences have shown promising results