Exploiting the dependencies in information fusion

This paper presents a novel approach for multisensory information fusion in the Bayesian inference framework. Specifically, under the maximum entropy principle, a formula is derived for estimating the joint probabilities of multisensory signals. The formula uses appropriate mapping functions to reflect the dependencies among multisensory signals. Selection of the mappings is guided by the maximum mutual information criterion. In addition, an algorithm is proposed for linear mappings of Gaussian random variables. Experiments on simulated Gaussian data and video/audio signals have been carried out. Preliminary results demonstrate that the proposed method can significantly improve the recognition accuracy for this type of tasks.