Improved Adaptive Mixture Learning for Robust Video Background Modeling

2 Related Works Gaussian mixtures are often used for data modeling in many real-time applications such as video background modeling and speaker direction tracking. The real-time and dynamic nature of these systems prevents the use of a batch EM algorithm. Currently, online learning of mixture models on dynamic data is achieved using an adaptive filter coupled with reassignment rules. However, convergence is very slow with a fixed learning rate typically employed in existing systems. In this report, we utilize an adaptive learning rate schedule to achieve fast convergence while maintaining adaptability of the model after convergence. Experimental results show a dramatic improvement in modeling accuracy using an adaptive learning schedule. Application of the proposed learning algorithm for video background modeling directly leads to improved approximation and robustness.