Research On The Multi-target Tracking Techniques Based On Probability Hypothesis Density Filter

Multi-target tracking technology, because of its great role to play in the field ofmilitary and civilian, has been a vital research area. Stable and efficient multi-targettracking algorithm is the core of the multi-target tracking technology, also the focus anddifficulty of this article. With the vigorous development of sensor technology in recentyears, there have been some new problems that the traditional multi-target trackingalgorithm difficult to solve, and there is an urgent need to explore new multi-targettracking algorithm to solve these problems. Probability hypothesis density (PHD) filterbased on the random finite set (RFS) theory is proposed in recent years, as a non-dataassociated multi-target tracking algorithm, can overcome a series of problems caused bytraditional data association algorithm, and has the capacity of tracking time-varyingnumber of targets. The PHD filter has been attracting attention of scholars from variouscountries.In Chapter2, the theoretical foundations of the multi-target tracking algorithmbased on PHD filter are introduced. First, a deeply study for a variety of multi-targetmeasurement likelihoods under different conditions is carried out. Then the particlefilter algorithm and main factors influencing its performance are discussed in detail.Finally, the basic contents of the multi-sensor information fusion theory are provided.This chapter provides theoretical support for subsequent chapters.In chapter3, as to the fact that it is difficult to obtain analytical form of optimalsampling density and tracking performance of standard particle probability hypothesisdensity (P-PHD) filter would decline when clustering algorithm is used to extract targetstates, a free clustering optimal P-PHD (FCO-P-PHD) filter is proposed, based onlinearization of measurement equation and the association between state particle andmeasurement set. This method could lead to obtainment of analytical form of optimalsampling density of P-PHD filter and realization of optimal P-PHD filter without use ofclustering algorithms to extraction target states, which avoids the negative effects ontracking performance of clustering algorithm. Besides, as sate extraction method inFCO-P-PHD filter is coupled with the process of obtaining analytical form for optimalsampling density, through decoupling process, a new single-sensor free clustering stateextraction method is proposed, which could be used independently as a state extractionmethod. Combining this method with standard P-PHD filter, FC-P-PHD filter could beobtained, which would significantly improve tracking performance of P-PHD filter. Theeffectiveness of proposed algorithms and their advantages over other algorithms arevalidated through several simulation experiments.In Chapter4, firstly, the advantages and disadvantages of the product multi-sensorPHD (PM-PHD) filter and sequential PHD(S-PHD) filter are discussed, and the relationship between them is analyzed. We find S-PHD filter is a special case of thePM-PHD filter. Then, for the poor performance of using the clustering algorithm toextract multi-target states when the distance between targets is very small, a freeclustering state extraction method for multi-sensor multi-target tracking issues isproposed. By combining this state extraction method with product multi-sensor particlePHD (PM-P-PHD) filter, we proposed free clustering product multi-sensor particle PHD(FC-PM-P-PHD) filter. The experiments showed that: the proposed multi-sensor freeclustering state extraction method is superior to clustering algorithm, especially whenthe distance of targets is small, the tracking performance of FC-PM-P-PHD filter issubstantially superior to PM-P-PHD filter.In Chapter5, a new multiple extended target tracking algorithm using theprobability hypothesis density (PHD) filter is proposed, to solve problems on trackingperformance degradation of the extended target PHD (ET-PHD) filter under thenonlinear conditions and its intolerable computational requirement. It is noted that theexisting Gaussian mixture implement of ET-PHD filter can only gain excellent trackingperformance under linear and Gaussian conditions. To extend the application ofET-PHD filter for the nonlinear models, a particle implement of ET-PHD (ET-P-PHD)filter is derived. Our study finds that the main factors influencing the computationalcomplexity of the ET-P-PHD filter are the partition number of measurement set and thecalculation of the non-negative coefficients of the cells in partitions. Through thepretreatment of the measurements and application of a new K-means clustering basedmeasurement set partition method, we have successfully decreased the partition number.In addition, a gating method for target state space based on likelihood relationshipbetween target state and measurement is proposed, to simplify the calculation of thenon-negative coefficients. Simulation results show that the algorithms proposed coulddeal well with multiple extended target tracking issues under nonlinear conditions, andsignificantly lower computational complexity with tiny effect on tracking performance.