Research On Multi-target Filtering Algorithm Based On Random Finite Set

With the fast development of the sensor technologies,the application scope and requirements of the multi-target tracking technology have been expanded constantly.The traditional multi-target tracking methods face the difficulties in both techniques and implementation methods in the presence of clutter,detection uncertainty and unknown target number since the complicated data association is needed.In recent years,the multi-source multiple target information fusion theory based on finite set statistics(FISST)theory has been developed,which lays a scientific and rigorous theoretical foundation for multi-target tracking problem.With the utilization of the multi-target system modeled by random finite set(RFS)and the probability statistics established upon set,the optimal multi-target Bayesian filter was derived for solving the problem of multi-target tracking.The resulting multi-target Bayesian filter has laid the foundation for developing many innovative multi-target filters or multi-target filtering algorithms that can directly estimate the number of targets and the corresponding target states in the space.Besides,this kind of method has many advantages over the traditional methods when dealing with the complex multi-target filtering problem.In view of the fact that the RFS-based methods have complete theoretical support and broad application prospects,this dissertation focus on the multi-target filtering algorithm in this framework for the single sensor multi target tracking applications.Specifically,we proposed some novel solutions and techniques which aim at addressing some problems in the existing algorithms and laying the foundation for improving their performance and extending their applications.The major contributions are given as follows:Firstly,the study on improving the estimation performance of the probability hypothesis density(PHD)filter is carried out.To overcome the drawbacks that the standard PHD filter needs the prior target birth intensity information and is easily affected by clutter measurements,a novel measurement-driven PHD filtering algorithm is proposed.In this method,the PHD recursion equation is performed in a decomposed form and the measurements are classified so that the filtering processes for the existing targets and the newborn targets are run separately.Therefore,the influence of the redundant measurements on the existing targets can be avoided,and also,the target birth intensity can be established adaptively according to the measurements of newborn targets at each processing step.Besides,a new multi-target state extraction method is proposed to address the troublesome problem ofstate estimation in the sequential Monte Carlo PHD(SMC-PHD)filter,where the individual measurements and particle likelihood information are exploited to achieve the association validation between the existing particles and measurements.Subsequently,the particles of detected targets are divided into different clusters corresponding to the effective measurements.Then,the point estimates of different target locations can be extracted according to the estimated target number.Simulation results demonstrate the effectiveness of our methods in comparison with the existing methods.Secondly,to solve the missed detection update problem in the cardinalized PHD(CPHD)filter,an improved Gaussian mixture CPHD(GM-CPHD)filtering algorithm is presented.Based on the Gaussian mixture representation of the posterior PHD,a weight redistribution scheme is incorporated into the filtering recursion to modify the updated weights of the Gaussian components when missed detections occur.The proposed method can help to prevent the detected components from gaining of extra weights originating from the PHD of the undetected targets.Accordingly,the updated weights of the undetected targets are compensated effectively and can be concentrated in the predicted vicinity of the undetected targets.Thus the PHD mass interaction via missed detections in the CPHD filter is reduced effectively.In addition,the process for compensating the weights of the undetected target considers the information in multiple frames,which can adapt the situations of consecutive missed detections.Moreover,the proposed method has no effect on the cardinality estimation result,as well as the cardinality distribution.Simulation results demonstrate that our method shows a significant improvement in the estimation performance as compared with the original CPHD filter when there exists detection uncertain.Finally,to extend the delta generalized labeled multi-Bernoulli(?-GLMB)filter to non-linear multi-target systems,a novel implementation method based the Gaussian mixture technique is proposed.Since the GM form is used to model the probability density function of individual targets,the multidimensional integral related to the Gaussian density and nonlinear density function need to be computed during the ?-GLMB filtering iteration,which has no close-form solution in general.To address this issue,the third-degree spherical-radial cubature rule is introduced to obtain an approximate solution,which allows the recursive propagation of the parameters of the ?-GLMB density via the prediction and update steps.The proposed method can be used in high dimensional nonlinear systems,and has good numerical stability.In addition,considering that the detection and tracking of the new targets in the ?-GLMB filter requires the prior information about the target birth distribution model,a measurement-driven initialization method which allows for establishing target birthdistribution adaptively during the filtering process is also presented.Simulation results show that the proposed methods can provide better tracking performance and exhibit good ability for detecting new targets without the requirement of the prior birth information.