unable to fit Gaussian mixture model

Hi, I am trying to recognise the moving objects in binary image. In each frame i have identified the blobs and also have their centroids. I need to fit a GMM so that they gets identified when the blobs overlap each other. could anyone help me in doing this. I tried the gmdistribution.fit function but i am confused with the fact that whether input parameter needs to be the distribution of the coordinates of the blobs in each frame? If this is the case then how will it identify the bob in each frame? Please provide me some help.

[kittu]

You do not have to give the coordinates of blobs,rather you need to provide the coordinates of each pixels in the blobs.This will give you 2xN array of x,y values.On this you preform GMM with the number of clusters you have/predict "clusters".For each cluster you will get a centroid and covariance matrix. This is how you perform the GMM.I hope it helps.

[Niraj]

Thanks kittu..it worked .

[Algorithms Analyst]

Can you share the source code..Becasue I am working on the same problem..

[kittu]

this could be done something like:

    %
pixelList = [struct.PixelList]; struct is a structure which contains the pixels of interest
obj = gmdistribution.fit(pixelList,k); %k is the number of clusters you want.
sig=obj.Sigma;
mu=obj.mu;
for cluster=1:k
[v,d]=eig(sig(:,:,cluster));
dd=sqrt(d);
ra=2*dd(1,1);
rb=2*dd(2,2);
x0=mu(cluster,1);
y0=mu(cluster,2);
ang = atan(v(2,1)/v(1,1))*(180/pi);
area=pi*ra*rb;
end

you can pass these parameters top draw the ellipse.

Kittu

[Niraj]

Thanks kittu, it helped..

Answer 1

[Tom Lane]

I don't know much about analyzing images. But gmdistribution expects its input data to be samples drawn from a Gaussian mixture distribution. If your image is being modeled so that the image values are something like the density of a gmm, then that's more like surface fitting than distribution fitting. Is that the case?