Research On Multisensor Fusion Algorithm Based On PHD Filtering

The study of multi-target tracking has come into a new century with the proposalof the PHD (Probability Hypothesis Density) filtering. Before that, researchers had toapply some filtering algorithms, which had been proved successful in the field ofsingle-target tracking, to study the problem of multi-target tracking with the help of dataassociation. Because of the avoidance of the explicit data association, the PHD filteringcan improve the situation. The PHD filtering is a filtering method based on randomfinite set statistics and the concept of random finite set can be used to exactlycharacterize the multi-target tracking problem. So the PHD filtering has a naturaladvantage in tackling the multi-target tracking problem compared with other methods.However, some external means, such as clustering, are necessary to extract peaks of thePHD and estimate the target states after estimating the number of targets.The STPHD (Single-Target Probability Hypothesis Density) filtering, a newmethod for peak extraction based on PHD filter, has been proposed. The STPHD is notan external method for peak extraction like others but can be considered as a naturalextension and improvement of the PHD filter and is the property of PHD in essence. Infact, STPHD is a kind of internal means for peak extraction from the PHD itself.However, the STPHD is proposed based on the scenario with one single sensor but notfor multi-sensor. To solve this problem, this paper extends the method to themulti-sensor scenario combined with a multi-sensor fusion framework and we get amulti-sensor fusion algorithm based on STPHD which, on one hand, has a wider rangeof applications, and on the other hand, can provide more accurate estimates and a bettertracking performance with the use of fusion.CPHD (Cardinalized PHD) filter is another important method based on randomfinite set statistics. Compared with the PHD filter, this method can provide moreinformation, but, in return, requires a larger amount of computation. In this paper, aconcept of the corresponding function has been abstracted from the certification processof STPHD and the STPHD is generalized to the CPHD filtering with the help of thecorresponding function to get a QSTCPHD (Quasi Single-Target CPHD) filtering, aCPHD version of this method. Furthermore, combined with the multi-sensor fusionframework, the QSTCPHD filtering can be extended to the multi-sensor scenario in thesame way to get the multi-sensor fusion algorithm based on QSTCPHD filtering. Thus,the theory system, which includes the PHD filtering, the STPHD filtering and themulti-sensor fusion algorithm based on STPHD filtering, is applied to the CPHDfiltering to get another theory system which includes the CPHD filtering, the QSTCPHD filtering and the multi-sensor fusion algorithm based on the QSTCPHDfiltering. In this way, we extend the range of application of the STPHD filter and furtherimprove the theory system.