Research Of Moving Target Tracking Algorithm And Its Application Under Complex Background

As one of the key problems in signal processing, target tracking has a very broad marketapplication value and become a hot research of scholars. Traditional multi-target trackingalgorithms, for example nearest neighbor and joint probabilistic data association, designsingle target filters, choose corresponding observation for every filters, and then receivetracking results. Their tracking performance is mainly influenced by data association.However, with the rapid development of information technology and the increasinglycomplicate environment, more and more requirements have been put forward by variousfields. The tracking algorithms on the basis of the data association can’t face the newchallenges. In recent years, the tracking methods based on random finite set has become animportant research subject and a variety of filtering algorithms emerge endlessly. Mahler’sprobability hypothesis density filter algorithm which laid the foundation for the future ofmulti-target tracking has a profound significance. In this algorithm, the collection ofindividual targets is treated as a set-valued state, the collection of observations received bysensors is treated as a set-valued observation, and the posterior intensity function ispropagated by Bayesian filter, so the algorithm is able to effectively estimate the number oftargets and their states. Depending on this tracking approach, we successfully avoid thecomplex data association and break the limitation of traditional algorithm. Meanwhile, itattracts more attention as a result of improving computational effort greatly. This algorithm isstudied deeply in this dissertation. Based on the in-depth study of the PHD filter, we give theparticle PHD filter, Gaussian mixture PHD filter and its improved algorithm. The specificresearch contents are as follows:1. We begin with a review of single-target Bayesian filtering, and introduce the definitionof random set theory and the concept about set of integral and set derivative. Then usingmulti-target Bayesian tracking models based on random finite set, the PHD filter algorithm isformulated. There is also a detailed introduction about the basic recursive equations and stepsof PHD. The above theory has laid a theoretical foundation for target tracking technology onthe basis of random finite set.2. Although, the PHD filter has transformed the set of integral of multi-target state intosingle-target condition, it still involves multiple integrals that have no closed form solutions ingeneral. Aiming at this, two different improvement method based PHD filter are proposed.The particle PHD filter, combined with sequential Monte Carlo approach, can be used innonlinear non-Gaussian environment. On the other side, the Gaussian mixture PHD filter,using Gaussian mixture target model and give a closed-form solution to the recursion, is onlyapplied under linear Gaussian assumptions. At last, analysis and summarizes of the abovemethods are performed through simulation experiments.3. An improved algorithm is proposed to resolve the missed detection problem andenhance the accuracy of the filter while tracking close proximity targets. Under Gaussianmixture assumptions, the predication and update equations of the PHD filter are modified,which effectively solve the information loss problem of missed true targets. And then depending on the weights of Gaussian components which decide whether the components canbe utilized to extract states, the proposed algorithm avoids the components which have higherweights are merged and improves the tracking performance when the targets move closely.Simulation results show that the new algorithm has advantages over the ordinary one in boththe aspects of filter precision and multi-target number estimation.