Active and dynamic information fusion with Bayesian networks

This dissertation addresses the problem of active and dynamic information fusion in multisensor systems. An information fusion framework is formalized based on dynamic Bayesian networks (DBNs) coupling with an active sensor controller. The DBNs provide a coherent and unified hierarchical probabilistic framework to represent, integrate and infer dynamic and uncertain sensory information. The active sensor controller serves as sensor cueing that allows the fusion system to actively select a subset of sensors to produce the most decision-relevant information and to arrive a reliable conclusion with reasonable time and limited resources. The proposed framework provides active and dynamic information fusion well suited to applications where the decision must be made efficiently and timely from dynamically available information of diverse and disparate sources.; The problem of determining an optimal subset of sensors from N available sensors arises in active information fusion. There is currently no efficient approach to this problem when N is large and the sensors are conditionally dependent. A theory is developed for finding the sensor subset with greatest expected contribution to the reduction of uncertainty of hypotheses when data fusion is performed with regard to Bayesian Networks. The heart of the strategy is to utilize the sensor pairwise information to infer the synergy among multiple sensors through exploiting the properties of mutual information. An approximate algorithm is constructed. The algorithm can reduce the time complexity exponentially in searching for an optimal sensor subset, yet guarantee a high-quality of the solution without assuming that sensor observations are conditionally independent.; Two illustrative applications are presented to demonstrate the feasibility of the proposed framework of active and dynamic information fusion and the theory of sensor selection. These applications include facial expression understanding from dynamic image sequences and multistage battlefield situation assessment.