Overview of University of Michigan MURI

SEQUENTIAL ADAPTIVE MULTI-MODALITY TARGET DETEC-

TION AND CLASSIFICATION USING PHYSICS-BASED MODELS

CO-PI	Year #1: Sept. 2002-Sept. 2003	Year #2: Sept. 2003-Sept. 2004	Year #3: Sept. 2004-Sept. 2005
OVERALL PROJECT GOALS	Physics-based models for vehicles under foliage Physics-based models for GPR imaging of landmines Myopic statistical algorithms for sensor management	Statistical models for vehicles under foliage Physics-based models for EMI detection of landmines Non-myopic statistical sensor management algorithms	Initial multimodal algorithms for vehicles Statistical models for GPR and EMI for landmines Statistical algorithms for vehicle detection
KAMAL Radar for: Foliage; Trunks; Vehicles	Phenomenological studies of radar scattered fields Multiple scattering from pine needle clusters Time reversal algorithm for radar imaging in foliage	Iterative physical optics for hidden objects Estimation of lossy propagation constants Frequency correlation for channel identification	Model validation using UM anechoic chamber Data generation for specific scenarios Improvement of model computational speed
HERO Stats for: Sensor schedule; Detection algorithms	Markov random field model for scattered fields Myopic one-step ahead sensor management Statistical test for global maximum in algorithms	Monte Carlo simulations of radar scattered fields Non-myopic multi-step sensor management MATILDA: Bounds on time reversal performance	Distributed algorithms for sensor management Reinforcement learning for sensor management Time reversal imaging sensor management
YAGLE Land mines; Blind deconv.	Ground-penetrating radar range-migration imaging 2D blind deconvolution for even Green's functions Basis-function-based Born inverse scattering	Induction time decay for metal classification Hyperbola-flattening transform pre-stage 3D blind deconvolution for even Green's functions	Monte Carlo simula- tions of GPR and EMI Statistical detection and classification algorithms Blind deconvolution applied mine imaging

Performed phenomenological studies of physics-based clutter and target models.
Significance: Basic understanding of these effects is vital for interpreting results.

Developed algorithm for modelling multiple scattering from pine needle clusters.
Significance: An example of how detailed our physics-based models are.

Developed time-reversal method for foliage-camouflaged target detection.
Significance: This is one of the physics-based vehicle detection algorithms.

Developed non-parametric Markov-Random-Field (MRF) approach for statistically characterizing
the scattered fields from clutter and from a target in clutter.
Significance: This permits scattered fields to be modeled from few observations.
Statistical models for entire scattered fields can be extrapolated from little data.

Developed myopic distributed multi-sensor multi-look detection and tracking
sensor management algorithms using Renyi information divergence measures.
Significance: Theoretically optimal sensor scheduling is infeasible
when the numbers of targets, their locations, and available sensors are all large.
Using particle filtering and Renyi-divergence-based scheduling, complexity was reduced
and simultaneous tracking of dozens of actual target trajectories was demonstrated.

Developed a hypothesis test for whether a known local minimum is also a global minimum.
Significance: Algorithms for objection function converge only to local minima.
Can peformance after convergence be improved? This test provides a statistical answer.

Range-migration algorithm for landmine detection using ground-penetrating radar.
Significance: Mine detection and tanks-under-trees are the two problems which the
Michigan MURI is addressing. Since Sarabandi has extensive experience with foliage
and vehicle modelling, Yagle is covering mine detection (with less previous work).

2-D and 3-D blind deconvolution algorithms for even point-spread functions.
Significance: These will be applied to deconvolution of radar propagation channel
effects for mines, and perhaps foliage as well, depending on how well the frequency
decorrelation approach currently being developed by Sarabandi works out.

Developed a linear-algebra-based method for inverse scattering when the Green's function
and object are both modelled as a linear combination of known basis functions.
Significance: This was part of the optimal basis function portion of the project.
This effort was discontinued when Marcin Bownik departed Michigan.

Developed an iterative physical optics approach to account for the effects of shadowing
on a hard target under foliage.
Significance: Previous approaches required a tremendous amount of computation to
account for part of a target or foliage blocking another part. The new approach
greatly reduces the computation required, making implementation more practical.

Developed a new approach to estimate signal attenuation through dense foliage.
Significance: This is another effect of propagation that we can now recover.

Development of frequency correlation approach to radar channel identification.
Significance: Despite our good modelling of foliage and tree trunk radar scattering,
some propagation effects are not amenable to modelling. This new approach seems
to be promising for deconvolution of propagation effects, though not physics-based.

Monte Carlo simulations of scattered fields from radar probing of targets in foliage.
Significance: This is the first set of statistical models based on physics-based
models. These will be inserted into the statistical sensor management algorithms.

Development of reduced-complexity non-myopic sensor management methods,
and quantification of their performance-vs.-complexity tradeoff.
Significance: The complexity of theoretically-optimal non-myopic multi-modal
scheduling increases exponentially with numbers of target states and modalities.
Sub-optimal approximations will enable practical target detection and tracking.

Development of MATILDA, an algorithm for computing bounds on performance of
time-reversal imaging algorithms (in any physical setting) in noise.
Significance: This is a first step toward the goal of algorithm evaluation,
although it is also important for evaluating performance of sensor configurations.

Development of model for decay time for metal-detector-induced electromagnetic field.
Significance: This is a second modality (after GPR) for mine detection.

Application of prestack hyperbola-detection algorithm (reciprocal+Radon transform).
Significance: This is a promising preliminary step in sequential mine detection, since
much less computation is required. If a likely mine is detected, the RMA is applied.
Statistical priors would be helpful here, but are not as vital as in vehicle detection.

Development of 2-D and 3-D blind deconvolution algorithms for even point-spread functions.
Signficance: These should help in improving imaging for classification of land mines.

Development of statistical models for scattered fields for vehicles in clutter.
Significance: The Michigan MURI has the unique ability to develop detailed
statistical models, due to our excellent foliage and vehicle models. These models
will be central to any statistics-based vehicle detection algorithm. However, each
Monte Carlo simulation requires two weeks of computation on 20 processors!