Narrowband Tracking Using a Markov Random Field Algorithm
Abstract
We present an algorithm for formulating the narrowband (or target) tracking problem as a Markov Random Field (MRF) with discrete and continous-valued hidden state variables. We then derive a simplied algorithm to estimate the model state variables. An MRF exists whenever there is a collection of sites that statistically interact with their neighbors. This requires defining a neighborhood system for determining which sites are the neighboring sites - such as a distance metric. If the Markov property is met, then if we are trying to make statistical inference about a given site, knowledge of all the sites is no better than knowledge of the state variables of only the neighboring sites. In the narrowband tracking problem, we assume we have detected a number of "interesting sites" in a spectrogram. These "interesting sites" are regions where there appears to be straight-line motion of a narrowband signal, perhaps detected by the application of a radon transform. Each site n has a measurement of center frequency, f(n), frequency rate (r)n, and amplitude a(n). We seek to make "soft" connections between the interesting sites so as to make longer tracks. We apply a linear dynamical model, to explain the behaviour of the target. This is identical in formulation to a Kalman filter, with the exception that the state transition matrix is many-to-one (many neighbor sites to a single site). The method is generalizable to any tracking problem.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 2006
- Accession Number
- ADA498815
Entities
People
- Paul Baggenstoss
Organizations
- Naval Undersea Warfare Center