A Research Of Information Fusion Based On Radar Sensor Networks

Information fusion is a type of technology of signal processing which uses multisource information of target obtained by sensor networks to conduct correlation,filtering and fusion.It builds a structure of information processing which can be used for target recognition,target tracking,information interpretation and final decision.The application of information fusion can effectively improve the spatial coverage of radar sensor networks,the credibility of decision-making and the viability of the system.In order to improve the target recognition rate and tracking accuracy,some key technologies of information fusion in airspace detection radar sensor networks and vehicleborne radar sensor networks are studied in this dissertation.1.In order to solve the problem of feature fusion in multi-source radar sensor networks for spatial search,a feature fusion scheme based on stacked autoencoder is proposed.The target recognition rates of three feature fusion methods,including serial feature fusion,genetic algorithm feature fusion and stacked autoencoder feature fusion,are compared through simulation experiments.The results show that the recognition rate of feature fusion based on stacked autoencoder is higher than that of other feature fusion methods because of the anti-noise processing and optimization.Otherwise,the performance of support vector machine is almost the same with multi-layer perceptron.2.In order to solve the problem of evidence conflict in D-S evidence theory for decision fusion,a combination rules based on information-entropy is proposed.Four rules including the original D-S rules,Yager combination rules,the evidence distance-based weighted combination rules and the information-entropy combination rules proposed in this dissertation are compared in simulation experiment.The results show that: the original D-S rules tend to deviate from the true decision when evidences conflicting seriously;Yager combination rules will greatly reduce the belief value of the combination results;the evidence distance-based weighted combination rules can reduce the deviation caused by the extreme value of belief function in evidence combination;the proposed combination rules can retain the collectivization effect and weaken the evidences with greater uncertainty by information-entropy weighting.3.In the research of vehicle-borne radar sensors,sensor networks based on lidar,millimeter-wave radar and camera are constructed.Meanwhile,a framework based on fusion information for target recognition and target tracking is designed and a data association algorithm based on weighted Intersection-over-Union is proposed to solve the problem of data association in detection fusion and target tracking.In the research,the proposed fusion framework is compared with YOLO,a target detection method based on deep learning.Compared with the fusion method proposed in this dissertation,YOLO is simpler and more direct.However,when the visual information is poor(especially on rainy or foggy days),the fusion method proposed in this dissertation can perform better than YOLO with radar's data.