We propose an effective online background subtraction method, which can berobustly applied to practical videos that have variations in both foregroundand background. Different from previous methods which often model theforeground as Gaussian or Laplacian distributions, we model the foreground foreach frame with a specific mixture of Gaussians (MoG) distribution, which isupdated online frame by frame. Particularly, our MoG model in each frame isregularized by the learned foreground/background knowledge in previous frames.This makes our online MoG model highly robust, stable and adaptive to practicalforeground and background variations. The proposed model can be formulated as aconcise probabilistic MAP model, which can be readily solved by EM algorithm.We further embed an affine transformation operator into the proposed model,which can be automatically adjusted to fit a wide range of video backgroundtransformations and make the method more robust to camera movements. With usingthe sub-sampling technique, the proposed method can be accelerated to executemore than 250 frames per second on average, meeting the requirement ofreal-time background subtraction for practical video processing tasks. Thesuperiority of the proposed method is substantiated by extensive experimentsimplemented on synthetic and real videos, as compared with state-of-the-artonline and offline background subtraction methods.
展开▼
