Probability Hypothesis Density Filter Algorithm And Its Application In Multi-Targets Tracking

With the multi-target tracking theory getting mature increasingly, the application of themulti-target tracking technique has been developed from military to civilian area. and themethods for solving the multi-target tracking problem emerge endlessly. The classical dataassociation algorithm which is efficient in the simple civilian area have become unsuitable formodern complex military combat system, which need to real-time track and localize everytarget accuratly in the monitor area to achieve the best fighting effectiveness. As a result, theresearch of multi-target tracking algorithm still has some of importance and urgency. Theprobability hypothesis density (PHD) filtering multi-target tracking algorithm based onrandom set theory is put forward by Mahler, which has strict mathematical theory foundationand high estimation precision, reduces the computation burden, can be realized easily, breaksthe traditional data association method and exploit the area of multi-target tracking research.On the basis of above theories, the study contents of this paper are expressed as follows:(1)Some classical multi-target tracking methods, such as Bayes filtering, Kalmanfiltering and particle filtering, are introduced at first. Bayes recursive multi-target trackingmodel has been set up based on the foundation of random set theory. The evaluation indexcurrently applied to the PHD filtering performance is analyzed and the evaluation methodsuitable to the filtering in this paper is identified. The theoretical foundation is laid for thefollowing research.(2)The feasibility of the random set theory application in multi-target tracking isanalyzed and the reasons and advantages of choosing random set theory for the multi-target isstated. Then PHD filtering algorithm, CPHD filtering algorithm and the realization of particlefiltering and Gaussian mixture filtering are given out briefly. At last, the filtering algorithmswhich based on random set theory for multi-target tracking are researched and analyzed;meanwhile the advantages, disadvantages, the applicable environment and the scope of eachalgorithm for these algorithms are discussed and some improvements are made.(3)Because of the presence of strong clutter, Gaussian mixture probability hypothesisDensity filtering algorithm (GMPHDF) will miss targets or tracks wrong targets, and thecalculation burden will increases while the strength of the cluster get bigger. For this reason,the GMPHDF based on kernel density estimation is proposed. After pruning and merging inGaussian mixture PHD filtering algorithm, the Mean-shift algorithm is introduced to estimatekernel density of Gaussian mixture PHD distribution density function, which replaces the state estimation method of Gaussian mixture PHD filtering algorithm. By choosing theestimated peak value as the state estimation of targets, the purpose of ruling out strong clutterinterference is achieved well.(4)Aiming at the low filtering estimation accuracy, filtering divergence of particlePHD and the undetected problem of local filter number, the SMC-CPHD filtering algorithmis proposed. By using random samples to approach the PHD distribution and the cardinaldistribution simultaneously, the problem that there is no closed solution in the filtering processcan be solved. The PHDF is close to Bayes optimal estimation with the increasing of theparticle sample number, which avoids the weight transfer problem of PHD filter when sometargets miss detection. Comparing with particle PHD and CPHD filtering, it has more reliabletracking ability.(5)The single sensor can only get local and one-side information and the changing intarget number and the gusty maneuvering of the target caused by target derivation andmission will make single sensor lay off. Aiming at this problem, the CPHD filter algorithmwith multi-sensor for multi-target tracking based on the adaptive “current” statistic model isproposed. The adaptive “current” statistic model has a very good adaptive ability, but thetracking performance would reduce when the targets are in the weak maneuvering state.Therefore, three sensors (radar) set at different places in the observation area are used to takesequential fusion to realize target tracking. Compared to the single sensor method, thismethod increased the tracking accuracy, which reflects the advantage of the multi-sensorinformation fusion.(6)The Variable Structure Multiple Model algorithm(VSMM) is combined withGaussian Mixture Cardinalized Probability Hypothesis Density (GMCPHD) filter algorithmand applied to the multiple marneuver targets tracking system, which leads to the propositionof filter algorithm based on VSMM. the real-time transform model is used to delete theunnecessary models. Because of excessive competition among the models leads to thereduction of filtering performance, the tracking effectiveness of the algorithm is achieved wellfor both strong maneuver and weak maneuver.