Ross Deming

Neural Networks for Improved Tracking

In this letter, we have developed a neural network (NN) based upon modeling fields for improved object tracking. Models for ground moving target indicator (GMTI) tracks have been developed as well as neural architecture incorporating these models. The neural tracker overcomes combinatorial complexity of tracking in highly cluttered scenarios and results in about 20-dB (two orders of magnitude) improvement in signal-to-clutter ratio.

Concurrent multi-target localization, data association, and navigation for a swarm of flying sensors

We are developing a probabilistic technique for performing multiple target detection and localization based on data from a swarm of flying sensors, for example to be mounted on a group of micro-UAVs (unmanned aerial vehicles). Swarms of sensors can facilitate detecting and discriminating low signal-to-clutter targets by allowing correlation between different sensor types and/or different aspect angles. However, for deployment of swarms to be feasible, UAVs must operate more autonomously. The current approach is designed to reduce the load on humans controlling UAVs by providing computerized interpretation of a set of images from multiple sensors. We consider a complex case in which target detection and localization are performed concurrently with sensor fusion, multi-target signature association, and improved UAV navigation. This method yields the bonus feature of estimating precise tracks for UAVs, which may be applicable for automatic collision avoidance. We cast the problem in a probabilistic framework known as modeling field theory (MFT), in which the pdf of the data is composed of a mixture of components, each conditional upon parameters including target positions as well as sensor kinematics. The most likely set of parameters is found by maximizing the log-likelihood function using an iterative approach related to expectation-maximization. In terms of computational complexity, this approach scales linearly with number of targets and sensors, which represents an improvement over most existing methods. Also, since data association is treated probabilistically, this method is not prone to catastrophic failure if data association is incorrect. Results from computer simulations are described which quantitatively show the advantages of increasing the number of sensors in the swarm, both in terms of clutter suppression and more accurate target localization.

Multi-Target/Multi-Sensor Tracking using Only Range and Doppler Measurements

A new approach is described for combining range and Doppler data from multiple radar platforms to perform multi-target detection and tracking. In particular, azimuthal measurements are assumed to be either coarse or unavailable, so that multiple sensors are required to triangulate target tracks using range and Doppler measurements only. Increasing the number of sensors can cause data association by conventional means to become impractical due to combinatorial complexity, i.e., an exponential increase in the number of mappings between signatures and target models. When the azimuthal resolution is coarse, this problem will be exacerbated by the resulting overlap between signatures from multiple targets and clutter. In the new approach, the data association is performed probabilistically, using a variation of expectation-maximization (EM). Combinatorial complexity is avoided by performing an efficient optimization in the space of all target tracks and mappings between tracks and data. The full, multi-sensor, version of the algorithm is tested on simulated data. The results demonstrate that accurate tracks can be estimated by exploiting spatial diversity in the sensor locations. Also, as a proof-of-concept, a simplified, single-sensor range-only version of the algorithm is tested on experimental radar data acquired with a stretch radar receiver. These results are promising, and demonstrate robustness in the presence of nonhomogeneous clutter.

Three-channel processing for improved geo-location performance in SAR-based GMTI interferometry

This paper describes a method for accurately geo-locating moving targets using three-channel SAR-based GMTI interferometry. The main goals in GMTI processing are moving target detection and geo-location. In a 2011 SPIE paper we showed that reliable target detection is possible using two-channel interferometry, even in the presence of main-beam clutter. Unfortunately, accurate geo-location is problematic when using two-channel interferometry, since azimuth estimation is corrupted by interfering clutter. However, we show here that by performing three-channel processing in an appropriate sequence, clutter effects can be diminished and significant improvement can be obtained in geo-location accuracy. The method described here is similar to an existing technique known as Clutter Suppression Interferometry (CSI), although there are new aspects of our implementation. The main contribution of this paper is the mathematical discussion, which explains in a straightforward manner why three-channel CSI outperforms standard two-channel interferometry when target signatures are embedded in main-beam clutter. Also, to our knowledge this paper presents the first results of CSI applied to the Gotcha Challange data set, collected using an X-band circular SAR system in an urban environment.

Simultaneous detection and tracking of multiple objects in noisy and cluttered environment using maximum likelihood estimation framework

We discuss a versatile framework for multiple target detection and tracking based on maximum likelihood estimation with expectation maximization and a cognitive theory called dynamic logic. In this contribution extend the framework to detection of moving objects in video sequences. The paper presents the theory and an example of detection and tracking using a real world video sequence.

Dynamic Logic Applied to SAR Data for Parameter Estimation Behind Walls

Identifying and localizing targets within buildings using exterior sensors will offer superior advantages to the military and law enforcement communities. Research on wall-penetrating radar has produced significant advances in recent years regarding this topic. However, wall parameter ambiguities, multiple reflections, clutter, and measurement noise pose significant challenges to developing robust detection and estimation methods. In the present work we demonstrate can be mitigated using dynamic logic (DL), an adaptive method for iterative maximum likelihood.

A MATHEMATICAL THEORY FOR LEARNING, AND ITS APPLICATION TO TIME-VARYING COMPUTED TOMOGRAPHY

The brain has evolved to enable organisms to survive in a complicated and dynamic world. Its operation is based upon a priori models of the environment which are adapted, during learning, in response to new and changing stimuli. The same qualities that make biological learning mechanisms ideal for organisms make their underlying mathematical algorithms ideal for certain technological applications, especially those concerned with understanding the physical processes giving rise to complicated data sets. In this paper, we offer a mathematical model for the underlying mechanisms of biological learning, and we show how this mathematical approach to learning can yield a solution to the problem of imaging time-varying objects from X-ray computed tomographic (CT) data. This problem relates to several practical aspects of CT imaging including the correction of motion artifacts caused by patient movement or breathing.

Tutorial on Fourier space coverage for scattering experiments, with application to SAR

The Fourier Diffraction Theorem relates the data measured during electromagnetic, optical, or acoustic scattering experiments to the spatial Fourier transform of the object under test. The theorem is well-known, but since it is based on integral equations and complicated mathematical expansions, the typical derivation may be difficult for the non-specialist. In this paper, the theorem is derived and presented using simple geometry, plus undergraduatelevel physics and mathematics. For practitioners of synthetic aperture radar (SAR) imaging, the theorem is important to understand because it leads to a simple geometric and graphical understanding of image resolution and sampling requirements, and how they are affected by radar system parameters and experimental geometry. Also, the theorem can be used as a starting point for imaging algorithms and motion compensation methods. Several examples are given in this paper for realistic scenarios.

Concurrent Tracking and Detection of Slowly Moving Targets using Dynamic Logic

We describe a new approach for combining range and Doppler radar data to perform multi-target detection and tracking. The algorithm framework is based upon dynamic logic, a biologically-inspired neural architecture, which yields advantages over conventional multi-target tracking algorithms by reducing the computational complexity during data association by several orders of magnitude. The algorithm is tested on experimental range-plus-Doppler radar data, and the results demonstrate a surprising degree of robustness in the presence of nonhomogeneous clutter and uncertainty in the number of targets

Track-Before-Detect of Multiple Slowly Moving Targets

We describe a new approach for combining range and Doppler data from multiple radar platforms to perform multi-target detection and tracking. Increasing the number of sensors can cause data association by conventional means to become impractical due combinatorial complexity, i.e., an exponential increase in the number of target to signature mappings. If the azimuthal resolution is coarse, this problem will be exacerbated by the resulting overlap between signatures from multiple targets and clutter. Our approach avoids combinatorial complexity during data association by performing an efficient optimization in the space of all target tracks and mappings between tracks and data. The reduced computational complexity of our approach scales only linearly with increasing numbers of targets and sensors. As a proof-of-concept, a simplified (single-sensor, range-only) version of the algorithm is tested on experimental radar data acquired with a stretch receiver. These results are promising, and demonstrate a surprising degree of robustness in the presence of nonhomogeneous clutter. Also the full, multi-sensor, version of the algorithm is tested on synthetic data. These results demonstrate that very accurate tracks can be estimated by exploiting spatial diversity in the sensor locations. The algorithm appears to be robust in the presence of clutter and uncertain knowledge regarding the number of targets present.