Improved Adaptive Gaussian Mixture Model for Background Subtraction

Posted on 15/04/2019, in Paper.
  • GMM: The Mixture Gaussian model is for the posterior probs of the background– this is solid if we assume a uniform distribution for the foreground object appearance in eq(1).
  • Updating rule: eq (4-6) provides an online updating rule for the GMM which is explained as a MAP optimal solution.In order to adapt to changes, they update the training set by adding new samples and discarding the old ones, which becomes a rolling window.
  • Robust background model: One interesting trick is it only uses B largest components in inference, see eq (7-8). A more sophisticated version (including prior was discussed in the second half of section 4).
  • Result: This paper cut down the computation cost by 10-50% depending on the complexity of the video while keeping the performance roughly the same.

Related: Adaptive background mixture models for real-time tracking