Research On Multi-sensor Fusion Method Based On Neural Network

Target tracking refers to processing the measurement from the sensor in order to maintain the estimation of the actual state of the target.More diversified detection methods are needed to produce better target tracking accuracy in the increasingly complex detection environment.The single sensor system will have a negative impact on its results,while the multi-sensor system is more advantageous than the single one.The multi-sensor fusion system contains many similar or even dissimilar sensors.In order to improve tracking accuracy,it is necessary to reasonably schedule and distribute each sensor,and calculate the accurate description of the target in motion state from the redundant measurement information.The uncertainty of measurement information and the complexity of target motion render high complexity and integration characterized by the multi-sensor fusion system.This thesis focuses on how to design the fusion algorithm to improve the tracking ability of the multi-sensor fusion system.Firstly,this thesis put forwards a multi-sensor fusion algorithm based on UKF and Elman neural network in response to the poor performance of traditional fusion filtering methods in the fusion tracking of maneuvering targets.Elman neural network can be introduced to compensate the fusion results,thereby remarkably improving the accuracy of fusion tracking.In addition,a hierarchical sequential covariance intersection fusion algorithm is used to transform a high-dimensional nonlinear optimization problem into several one-dimensional nonlinear function optimization problems.Secondly,this thesis provides a more general multi-sensor asynchronous fusion algorithm based on IMM and RBF neural network,against the backdrop of real scene where the most asynchronous measurement data attained by the fusion center of the system source from every sensor.Interactive multi-model as the prediction model of motion target is utilized to arrange asynchronous measurement obtained in each fusion cycle of the fusion center in chronological order for sequential filtering,and the results are sent to the RBF neural network for online learning to obtain more accurate fusion filtering results.Finally,a multi-radar fusion tracking simulation platform is designed to facilitate the study of fusion algorithm and verify the effectiveness of the fusion algorithm proposed in this thesis.Users can configure the observation accuracy of each radar according to their needs and select the fusion tracking algorithm,and the fusion results and error information will be displayed visually through the interface.