Developing an accurate computational model of human visual attention has been a long-standing challenge. Such a model may allow any system to select only relevant information from a complex and cluttered visual input in numerous artificial vision applications, such as robotics, surveillance, driving assistance, multimedia recognition and retrieval.

The first biologically plausible model for explaining the human visual attention system was proposed by Koch and Ullman (1985), and later implemented by Itti et al (1998). This model analyzes still images to produce primary visual features, such as intensity, color and orientation, which are combined to form a saliency map that represents the relevance of visual attention. Although several attempts have been made to improve the Koch-Ullman model, they all suffer from a crucial problem in that the saliency responses are assumed to be deterministic.

Based on the above considaration, we propose a new stochastic model of visual attention. The proposed model is composed of a dynamic Bayesian network with four layers that combines several fundamental statistical models. The proposed model enable us to

1. automatically estimate eye focusing positions an their densities only from video frames,
2. automatically derive model parameters of the network with the EM algorithm when eye tracking sequences of human subjects are available,
3. execute the estimation procedure in near real-time (70 msec/frame @ 640x480 pixels) with stream processing through such as GPUs, by introducing MCMC-based particle filter.