Gaussian Mixure Model in Object Tracking

Summary on this Object Tracking lesson: https://www.youtube.com/watch?v=0nz8JMyFF14

The overall idea is to find large supporting evidence / small standard deviation: if this value is large, it is background. Else, foreground.

The evidence is probably the value for the point in the Gaussian model. Or it can be simply the distance from the mean of the Gaussian model. The lesson didn't mention clearly.

For each pixel there will be a Gaussian Mixture Model. Compute a pixel color histogram for the first N frames of a video. Normalize the histogram, and model it as a mixture of 3 to 5 Gaussians. 

During detection, for a value X,  | X - gaussian_center | < 2.5 standard_deviation, it is considered part of the Gaussian distribution.

To make the model adapt to new data, update Histogram H using the new pixel intensity. If new histogram differs from the old histogram a lot, refit the GMM.

GMM can be calculated using Expectation Maximization - Start by randomly picking means and standard deviations, cluster data points using these Gaussian distributions; then refine means and standard deviations. Repeat this process until the Gaussian distributions don't change any more (converged). Or, just use Python sklearn.mixture GaussianMixture.