NICOP - A new sequential Monte Carlo framework for tracking of non-linear complex dynamical systems
Abstract
Many problems related to environmental sensing, situation awareness and information fusion boil down to the ability of e ciently tracking complex nonlinear, high-dimensional stochastic dynamical systems. Examples abound, ranging from classical multi-target tracking in battle eld scenarios to weather/environmental forecasting for tactical planning. The algorithms for prediction and tracking of random dynamical systems are collectively termed stochastic lters. Most of these techniques seek numerical approximations, since closed form solutions do not exist for general nonlinear systems. This is the case of particle lters (PFs), which are recursive (online) methods based on statistical Monte Carlo principles. While PFs can be applied to any dynamical system, they are often criticized as computationally heavy and ine cient in high-dimensional models, precisely because of their reliance on Monte Carlo integration. However, although a number of deterministic methods have been recently proposed (e.g., cubature, optimal transportation or deterministic ow lters) as potential replacements of PFs, none of them has e ectively overcome the dimensionality/complexity problem yet. In this proposal, we advocate the development of a new particle ltering framework (including an extended methodological setting and the theoretical tools for its analysis) that still has sequential Monte Carlo integration at its core but is endowed with a number of features that address directly the key issues of dimensionality and complexity. Such features include the partitioning of high-dimensional state spaces (a divide and conquer approach), the prevention of the degeneracy phenomenon in importance samplers and the `automatic stabilization of the tracker. We aim at developing both the methodological and the theoretical aspects of the new framework, and to apply the resulting algorithms to selected problems related to the tracking of multiple and/or complex targets. The design of new and e cient nonlinear trackers for multiple and/or complex targets is relevant to several focus areas of the US Naval Science & Technology Strategic Plan, including, at least, Assure Access to the Maritime Battlespace (focus area #1), Autonomy and Unmanned Systems (f. a. #2) and Expeditionary and Irregular Warfare (f. a. #3). We will speci cally address the application of the new methodology to two problems: the joint tracking of a large number of targets and the forecasting of complex meteorological phenomena for tactical planning. This research will be partially carried out in collaboration with Prof. Petar M. Djuric, from the Department of Electrical and Computer Engineering of Stony Brook University (NY).
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2016
- Source ID
- N629091512011
Entities
People
- Joaquín Míguez
Organizations
- Office of Naval Research
- United States Navy
- Universidad Carlos III de Madrid