<p>In unsupervised clustering of non-stationary data, adaptive Gaussian Mixture Models (GMMs) are commonly used. This is especially true for real-time video segmentation methods. Nevertheless, the problem of choosing the right model complexity in terms of the number of mixture components and the learning stability speed tradeoff raises some unresolved issues. In this paper, we propose an algorithm that segments the first frame of a video sequence by means of an adaptive GMM that online adapts in order to estimate the number of components and the parameters, and then uses the found description mixture as a starting point for modelling non-stationary temporal distributions of pixels in video images. Then, we developed a statistical framework to perform background/foreground segmentation, interpreted at each single pixel level being a binary classification problem, where a probability function is used to determine how much that pixel belongs to the background or foreground. We performed experiments demonstrating that the method is insensitive to initialization (deterministic initialization strategy). Finally, we compared our work against other methods that employ GMMs, showing that it leads to improved segmentation performance compared to state-of-the-art alternatives. The algorithm showed a good segmentation performance for non-stationary data. However, even thought it did not bring about the best results in the foreground/background classifier task, it requires less computational burden compared no the other approaches, makes it suitable for non-stationary data segmentation.</p>