Research On The Multi-target Tracking Techniques Based On Random Finite Set Theory

With the fast development of the sensor technologies, the applications andrequirements of multi-target tracking expand rapidly, especially in military field. Newdemands are proposed for tracking, such as Space-based Tracking and SurveillanceSystem (STSS), Antimissile and Precision Guidance and hit. The traditional multi-targettracking methods are limited by the complicated data association. Tracking targets insituations where the number of targets is unknown or with dense clutters andmiss-detection etc. brings combination explosion and N-P hard problem. Fortunately,the development of multi-source multi-target information fusion theory based onRandom Finite Set (RFS) provides a scientific and uniform Bayesian framework. Thistheoretical optimal approach to multiple targets tracking can be used to jointly estimatethe number of targets and their states from measurement sets directly while avoidingdata association, and the computational complexity of which is much lower comparedwith traditional ones. Thus, it has wide application prospects. This dissertation focus onthe problems of multi-target tracking by both single and multi-sensor based on RFStheory, and the major contributions are as follows:The Chapter2is the fundamental of the subsequent chapters introducing the basictheory of RFS for multi-target tracking. An overview of RFS framework is firstlypresented, such as the definition of RFS and mathematical operators of the set calculusand so on. Then the first and second moment of multi-target Bayesian filters, known asprobability hypothesis density (PHD) and Cardinalized-PHD (CPHD) filters areintroduced in both single and multi-sensor forms respectively. At last, several metricsfor multi-target tracking performance evaluation are analyzed.In Chapter3, algorithms of PHD and CPHD filters with amplitude features oftargets are proposed to solve the problem of tracking multiple targets on the focal planeof optical sensor systems. The models of the amplitude likelihood of targets and clutterswith known and unknown SNR cases are established firstly, and then the amplitudelikelihood ratio of target to noise is cast into standard PHD and CPHD filters’recursions, called AI-PHD and AI-CPHD filter respectively. Gaussian Mixture (GM)implementations of both AI-PHD and AI-CPHD filter recursions are given. Simulationresults demonstrate that the proposed methods has well efficiency of target initiationand continuity, and outperform the standard ones in tracking performance, and havelower computational complexity especially in scenarios with dense clutters.The linear-complexity CPHD (LC-CPHD) filter is studied in Chapter4. For theoriginal derivation of LC-CPHD is extraordinarily complicated and hard to understandintuitionally, the formulations are derived based on the physical principals which obtainthe exactly the same results but much easier to understand. And then the GM implementation is presented to propagate the posterior intensity of the LC-CPHD filter.For the problems of the standard LC-CPHD filter such as having no identifications foreach target leading to no ability of forming target tracks and having no model forspawned targets resulting in poor tracking performance while there are targets spawned,improved methods are proposed to solve these problems. And the simulation results arevalidated their capabilities.In Chapter5, a multi-sensor single target tracking algorithm is studied based on theRFS framework to solve the problems of targets detection and tracking in clutters.Firstly, the RFS dynamic model and RFS observation model are established, modelingthe detection and tracking of multi-sensor single target as the optimal Bayesianestimation of target state sets. And then the optimal Bayesian filter formulations arederived rigorously based on Finite Set Statistics (FISST). However, the recursioninvolves complicated integrals that have no closed-form solutions in general, theSequential Monte Carlo (SMC) implementation is used for the proposed algorithm.Simulation results show that the proposed algorithm outperforms the traditional singletarget tracking methods both in linear and nonlinear scenarios.In Chapter6, multi-sensor multi-target tracking algorithm is studied calledProduct-Multisensor CPHD (PM-CPHD). Firstly, the SMC implementation of thePM-CPHD filter is derived. However, in dense clutters situation the computational loadof the SMC-PM-CPHD filter is insufferable. Thus, an algorithm based on the adaptivegating technique is proposed to improve its computational efficiency. Also, ameasurement-driven multi-target state estimation algorithm is proposed to improve themulti-target state estimation performance of the SMC-PM-CPHD filter. Simulationresults show the efficiencies and superiorities of both algorithms compared withexisting ones.