Home / About Us
||A Tracker based on a modified CPHD filter
||Thu 27 May 2010
||The PHD has been introduced by Mahler ([Mah07], [Mah03]) to cope with the problem of Multi- Target filtering. Roughly, instead of propagating the density multi-object, the PHD filter propagates only the first moment order of this density, the PHD (Probability Hypothesis Density Filter). Nevertheless, the PHD is not tractable directly, so we need some approximations. Two approximations have been developed. The first considers that the PHD is a Gaussian Mixture (GM) [VM06] while the second uses particle approximations [VSD05] to propagate the PHD. Here we are interesting in the first case. Nevertheless, the output of the algorithms based on PHD (or CPHD, which propagates the cardinality probability distribution too) does not enable to distinguish tracks. This could be a problem in a combat scene for example, where we need to distinguish a friendly plane to an enemy plane. Some schemes have been proposed, and the main idea is to associate a label at each Gaussian [PVC06], which will enable to distinguish the different targets on the scene. Schemes proposed are not efficient in the case where severals targets are close or are crossing. Thatís why the main objective of this work is to introduce a scheme which addresses the issue of the multi-target conflict, when several targets are crossing for example. We first introduce a 2-D assignment method based on PHD or CPHD, then we introduce some classical techniques like clustering and gating to improve the performance of the tracking. Finally, we do some simulations to compare the classical CPHD 2-D assignment method with the Clustering CPHD 2-D assignment method proposed. We compare our algorithm with the NNJPDA, the current algorithm used in some defense embarked systems (Infra-red cameras).
[Mah03] Ronald Mahler : Multitarger bayes filtering via first-order multitarget moments. IEEE Transac- tions on Aerospace and Electronic Systems, vol 39(no.4), October 2003.
[Mah07] Ronald Mahler : Statistical Multisource-Multitarget Information Fusion. Artech House, 2007.
[PVC06] Panta.K, Vo.B et Clark.D.E : An efficient track management scheme for the gaussian-mixture probability hypothesis density (gm-phd) tracker. In Proc. Fourth International Conference on Intelligent Sensing and Information Processing, 2006.
[VM06] B. Vo et W.K. Ma : The gaussian mixture probability hypothesis density filter. In IEE transactions Signal Processing, volume vol. 54, 2006.
[VSD05] B. Vo, S. Singh et A. Doucet : Sequential monte carlo methods for multi-target filtering with random finite sets. In IEE Trans. Aerospace and Electronic Systems, volume vol 41, 2005.
<< Back to list of seminars