Adaptive learning of Bayesian Networks for the qualification of traffic data by Contaminated Dirichlet Density Functions

Abstract

The concept of Bayesian networks (BNs) is an established method to model data fusion in sensor networks of several equal or different sensors. Although the method is powerful, there is a particular need for accurate sensors, the consideration of the affecting external, e.g. environmental conditions, and internal influences, e.g. the physical life of the sensor, in the sensor model and an accurate a-priori knowledge about the underlying process. In this paper an adaptive algorithm for learning BNs is introduced, which is applied to update the time-variant a-priori probabilities in sensor networks. This algorithm makes use of contaminated Dirichlet density functions (CDDFs). The effectiveness of adaptive learning is demonstrated for vehicle classification in traffic surveillance.