The Multiple Target Tracking Algorithm Based On Probability Hypothesis Density Study

An intrinsic problem in multi-target tracking is the unknown association of measurements with appropriate targets. Most traditional multi-target tracking formulations involve explicit associations between measurements and targets. Due to its combinatorial nature, the data association problem makes up the bulk of the computational load in multi-target tracking algorithms.Mahler’s Finite set statistics (FISST) provides a general systematic foundation for multi-target filtering based on the theory of random finite set (RFS). The basic idea is to do filtering on set-valued observations and set-valued states. One such technique, known as the Probability Hypothesis Density (PHD) filter, propagates the posterior intensity function, or first-order statistical moment, of the RFS of targets in time. This approach has led to efficient multiple target tracking algorithms for estimating the number of targets and their states. The PHD filter operate on the single-target state space and avoids the combinatorial problem that arises from data association. The random finite set (RFS) approach to multi-target tracking is an emerging and promising alternative to the traditional association based methods.Based on the theoretical analyzing and computer simulation, this dissertation takes research works as follow:First of all, we begin with a review of single-target Bayesian filtering. Then using random finite set models, the multi-target tracking problem is then formulated as a Bayesian filtering problem. This provides sufficient background leading to the PHD filter.Secondly, the PHD recursion involves multiple integrals that have no closed form solutions in general. The first such algorithm, known as the particle PHD filter, used a sequential Monte Carlo approach for approximating the posterior intensity function. The particle PHD filter reduces to the standard particle filter in the case where there is only one target with no birth, no death, no clutter and unity probability of detection. Through simulation, illustrates the performance of particle PHD filter under different clutter and the probability of detection. Thirdly, by analogy the Kalman filter as a solution to the single-target Bayes filter, under linear Gaussian assumptions, given a closed-form solution to the PHD recursion, called GM-PHD, and gives the simulation analysis.Finally, introduced a generalized form of the PHD filter algorithm:GMP-PHD. Through simulation, illustrates the effect of different sampling number of particles on filter performance. For the actual situation, PHD particle algorithm and its application in passive location is discussed.By the work above can be seen, the PHD filter is bound to occupy an increasingly important position in the field of multi-target tracking.