Multiresolutional Particle Filters With Finite Mixtures for Ground Target Tracking and Identification

MRLets Technologies, Inc. proposes an innovative technology for high performance and computationally efficient ground target tracking and identification using powerful multiresolutional particle filters with finite mixtures and GMTI/HRR measurements. This innovative technology consists of four major functional components: particle filters for handling nonlinearity and non-Gaussianity, multiresolutional processing for computational efficiency, finite mixture models with an expectation maximization (EM) algorithm for effective information exploitation, invariant-based joint tracking and identification for technology application. In addition to these four components, a set of high performance and innovative nonlinear and non-Gaussian filters will be developed to provide the theoretical foundations. A tight integration of these four functional components will provide us with unprecedented power in information exploitation, target modeling, data association, state filtering, and target identification. Nonlinear and non-Gaussian phenomena have long existed in almost every area of target tracking and identification and have become the bottleneck for developing high performance algorithms. Particle filters have been proven to be very effective for nonlinear/non-Gaussian estimation. However, the most notorious disadvantage of a particle filter is its formidable computational complexity, since thousands particles are usually needed, which is equivalent to running thousands of filters simultaneously. The heart of this research is to drastically reduce the computational complexity of particle filters while maintaining their full power. It has been proven that the only technique of truly solving an algorithmic computation problem is multiresolutional processing, both in temporal and spatial domains. Therefore, we propose a multiresolutional particle filter to reduce significantly the number of particles and update them at a lower rate, while maintaining the full strength of a particle filter. In addition, when applying a particle filter to a density function with multi-modality due to multiple targets, weak targets could be dropped or merged because of the poor localization property of a particle filter. We will improve localization accuracy by using a finite mixture with particles. Information extraction and target modeling for rigid-body targets could be greatly benefited from the invariant principle. In this project, we will extend our invariant-based tracking and identification algorithms to a multiresolutional particle filter implementation. We will also introduce a mixture-EM technique for effective HRR information exploitation. During the course of this research, we will develop a set of new techniques for nonlinear and non-Gaussian estimation. This proposal is packed with many important innovations. When fully developed, the proposed research should have a significant impact on the target tracking and recognition communities. Raytheon Company, a subcontractor, will investigate the possibility of transferring this innovative technology into their product lines.