1
$\begingroup$

I'm learning about Gaussian Mixture Model (GMM) for background subtraction.. And have some questions about it..

  1. In Chris Stauffer paper "Adaptive background mixture models for real-time tracking" Gaussian probability density function is $$\mathcal{N}(X_t,\mu,\Sigma)=\dfrac{1}{(2\pi)^\frac{n}{2} \left|{\Sigma}\right|^\frac{1}{2}}\exp\left(-\dfrac{1}{2}(X_t-\mu_t)^T\Sigma^{-1}(X_t-\mu_t)\right)$$
    where $n$ in multi dimension, is $n$ in 3-D which is pixel position in x,y, and intensity of pixel?
  2. In the same paper, they said: $$\Sigma_{k,t}=\sigma_k^2.I$$ Is $I$ is Indentity matrix?
  3. They also said red,green, and blue pixel values are independent and have the same variances, if it independent why R,G, and B have same variances? shouldn't it have different variances because the value of each channel is different..
  4. If I set the algorithm with number of distribution is 3, so it's only 1 distribution that chosen as background model and the rest is foreground model? Because they said the first B distribution are chosen as the background model, where $$B=\arg \min_b(\sum\limits_{k=1}^b w_k>T)$$
$\endgroup$
0
$\begingroup$
  1. Yes, the dimension of $X$ should be 3, one for each of the red, green, and blue channels. You can see this from their Figure 2, where they show plots of the distributions of red and green channels only (to motivate why they want to use a Gaussian Mixture Model in the first place --- the distribute of pixel values is multi-modal).

  2. Yes, $I$ is the identity matrix.

  3. Random variables can have the same variances and still be independent. The variance is part of the model. In this case what the model is saying is: the red, green, and blue pixel values vary without regard to each other (they are independent) but the amount that they vary is about the same as each other (they have the same variances).

  4. Whether only 1 distribution is chosen when $K=3$ will depend on the threshold $T$, surely? Yes, it is possible that only one distribution will be chosen.

$\endgroup$
  • $\begingroup$ 1. In their paper they're not telling what dimension to make Gaussian distribution they used, so I assuming if it's 3 which is x,y position and intensity of pixel or maybe 3 dimension of R,G,B pixel? $\endgroup$ – user3794938 Mar 9 '15 at 14:35
  • $\begingroup$ 3. Ahhh, I see.. I don't carefully read the paper, my mistake.. $\endgroup$ – user3794938 Mar 9 '15 at 14:52

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.