An Adaptive Learning Rate GMM for Background Extraction

摘要

The rapidness and stability of background extraction from image sequences are incompatible, when a conventional Gaussian Mixture Models is used to rebuild background. If the background region of the scene is changed, the extracted background becomes bad until the transition is over. A novelty adaptive method is presented to adjust learning rate of GMM in Hilbert Space. Background extraction is treated as the process of approaching to certain point in Hilbert Space, so the real-time learning rate can be obtained by calculating the distance between the two adjacent extracted background images, and the judgment method of stability of background is got. Comparing with conventional GMM, the method has both high rapidness and good stability at one time, and it can adjust the learning rate online. The experiment shows that it is better than conventional GMM, especially in transition process of background extraction.