Foreground and background reconstruction in poisson video

Image foreground and background separation is an essential step in a variety of image processing, video analysis, and computer vision tasks. Typically, these methods accept streaming video data, compute an estimate of the background, and subtract this from the observed frames to generate a foreground scene. While such methods are very effective in high SNR regimes, they face serious limitations in low-light settings occurring in night vision surveillance and astronomy. Existing methods cannot be easily modified to yield good results. Therefore, new methods must be created to deal with the low light setting. This paper specifically addresses the problem of foreground and background separation and reconstruction in the case of Poisson distributed observations. The proposed approach builds upon recent advances in both the online learning community and sparse reconstruction methods for Poisson images. To aid in the reconstruction and separation tasks, the method learns and incorporates the dynamics of objects in both the background and foreground in real time.