Real-Time Motion Detection For A Mobile Observer Using Multiple Kernel Tracking And Belief Propagation

We propose a novel statistical method for motion detection and background maintenance for a mobile observer. Our method is based on global motion estimation and statistical background modeling. In order to estimate the global motion, we use a Multiple Kernel Tracking combined with an adaptable model, formed by weighted histograms. This method is very light in terms of computation time and also in memory requirements, enabling the use of other methods more expensive, like belief propagation, to improve the final result.