The Study On Multisensor Fusion

Multisensor systems provide a purposeful description of the environment that single sensor can not offer. Fusing several types information enhances the capability of intelligent system and yields more meaningful information otherwise unavailable or difficult to acquire by a single sensor modality. Multisensor fusion is an emerging technology related to information science, computer science and automatic science.In this dissertation, the composition, the existing problems and the future research directions of fusion systems are stated. We emphasize on the study of fusion architectures and fusion methods. The models of function, architecture and mathematics for multisnesor systems are provided, which compose the core problem of fusion systems. The general approach to the design of fusion system is also discussed.There has not been a standardized classification for the sensor fusion architecture. So, based on many references, we divide the fusion architecture into four basic structures, which are centralized architecture, distributed architecture, hierarchical architecture without feedback and hierarchical architecture with feedback. By using Kalman filter, the specific fusion algorithms for the four architectures are respectively deduced in detail. Their advantages and disadvantages are also analyzed. So, we should select a fusion architecture for the fusion system according to the specific problem.We propose a method of fusing uncertain information and provide the standards of evaluating fusion methods. Based on fuzzy integral, an intelligent fusion system is given, which can solve the FEI-DEO and DEI-DEO fusion problems. The problem of object identification for industrial robots is also solved effectively and quickly by applying the intelligent fusion system to combine several sensor data. It enhances the capability of handling uncertain information.Based on neural networks and evidence theory, a spatial-temporal two-layer architecture for sensor fusion is presented. On the other hand, beginning with M.sageno inference model and combining the advantages of neural networks and fuzzy sets, we provide FMLPNN and FBFNN fuzzy neural networks which are used to fuse multisensory information. The trained fuzzy neural networks not only fuse exact information but also imprecise or fuzzy data, which can overcome the difficulty of getting fuzzy rules and membership function.The approach to multisensor fusion with rough sets is discussed. By analyzing uncertain, incomplete, and imprecise sensor observations, the fastest fusion algorithm is extracted, which may overcome the problem of fusing over-loaded or incomplete information in multisensor systems.