GMM: Learning patterns of activity using real-time tracking
Gaussian mixture model (GMM) is one of the most popular motion detection models. As a parametric model, GMM generates multiple Gaussian models for each pixel in the video. When a new frame arrives, GMM compares each pixel in the frame with every Gaussian model in the mixture until a matching Gaussian model is found. The matched Gaussian will be updated, and the pixel will be considered to be a background pixel. If no Gaussian model matched, the least probable distribution will be replaced with a new distribution with its mean equal to the current pixel value, and the pixel will be considered to be a foreground pixel.
GMM model significantly improves the motion detection performance comparing with the Single Gaussian model (SGM), especially for the videos with background motion and illumination changes.