Sequential Adaptive MultiModality Target Detection and Classification using

Sequential Adaptive Multi-Modality Target Detection and Classification using Physics-Based Models • Professor Andrew E. Yagle (PI) (EECS) Signal and image processing, inverse scattering • Professor Alfred O. Hero III (EECS) Statistics, signal and image processing • Professor Kamal Sarabandi (Director, Rad Lab) Scattering, inverse scattering, remote sensing • Assistant Professor Marcin Bownik (Mathematics) Wavelets, functional analysis and approximation

Sequential Adaptive Multi-Modality Target Detection and Classification using Physics-Based Models PROJECT SUPERVISION: • Dr. Douglas Cochran (DARPA) • Dr. Russell Harmon (ARO) INDUSTRY COLLABORATION: • Veridian (formerly ERIM) of Ann Arbor

Research Project Objectives • Develop overall algorithm for sequential detection, sensor management & selection • Develop physics-based models • Simplify physics-based models using functional-analysis-based approximation • Evaluate the resulting procedure on realistic models (statistical simulations) and real data

Issues: Overall Algorithm • How to select sensing modalities? • What is value-added for combining other modalities? Is it worth additional cost? • How do we implement data-adaptive configurations, e. g. , selection of sources/receivers, based on scattering of targets and propagation in medium? • What are the figures of merit? • How to select decision thresholds?

Practical Applications Develop a set of signal processing/statistics algorithms to solve multi-modal sensing problems Examples: • Detection of “tanks under trees”: vehicles under canopy of tree foliage • Detection of buried objects (land mines)

Issues: Physics-Based Models • Scattering models – Hard targets of different types (vehicle, mines, etc. ) – Clutter of different types (trees, rough surfaces, etc. ) • Propagation models, e. g. , tree canopies (attenuation, phase deformation, dispersion, etc. ) • Sensor models – – Radar, SAR, IR, etc. Frequency, polarization, incidence angle, etc. • Model order reduction – Using function approximation, e. g. , wavelets

Issues: Evaluation of Results • Behavior of algorithm on realistic models (UM Radiation Lab). • Benchmarking against physics-based models. • Figures-of-merit for evaluation of algorithm • Behavior of algorithm on real data

Overall Algorithm: Overview • Sequential feedback structure: Detectibility for given sensor waveform/source/receiver used to guide future sensor selection • Possible targets organized into tree structure leading to sequence of binary classifications • Inverse filter based on physics-based model used to improve performance of the above

Overall Algorithm: Overview Target detector/ classifier

Target Detector/Classifier Hybrid hypothesis tests (majority rules): • Bayes optimal test: Optimal, but requires Bayesian priors for unknown parameters • GLRT: Use maximum likelihood estimates (MLE) for unknown parameters • Maximal Invariant: Project data onto subspace on which density functions are independent of unknown parameters

Target Detector/Classifier • EXAMPLE: ATR • 1 of 9 images~below hidden in image above (forest-grassy plain) • Location: column 305 at the A/B boundary • Clutter: #clutter-only reference chips used A B

Target Detector/Classifier • RESULTS: ATR • “Structured Kelly”: standard ATR GLRT • Figure-of-merit: min. detectable target amp. • MI best with 200 chips • GLR best with 250 • Hence need hybrid test TEST TYPE 200 250 CHIPS maxim. 0. 0609 0. 0145 invar. GLR #1 0. 104 0. 0146 Struct. 0. 105 Kelly 0. 0141

Target Detector/Classifier • • • SAR imaging: none of these 3 better than rest We propose to use all 3 and let majority rule Apply log(P) times to distinguish P possible target types (different models of tanks, mines) organized into a tree structure (known models) • Conditional pdf unknown parameters: Orientation, location, reflectivity of target Propagation characteristics of the medium

Target Detector/Classifier Vehicle type Tank T 72 Orientation HMMWV

Physics based models

Physics-Based Models • Need models to develop conditional pdfs • Fast; include target, medium, sensor characteristics, with unknown parameters • Evaluation of integrals in Bayes optimal test: Monte Carlo marginalization too slow for us. • Sequential Importance Sampling: Denumerable-source blind deconvolution; Digital multi-user communications (real-time)

Physics-Based Models • Models needed for propagation and targets • Models include: unknown random parameters (e. g. , wavelet coefficients-see the following) • Use Monte-Carlo-type simulations to obtain non-parametric estimates of pdfs for these • Validate these with real data for targets/clutter • Result: statistical models of targets and clutter

Physics-Based Models Fractal generated tree stand SAR image of tree stand using VV polarization

Physics-Based Models • Problem: Both the propagation and vehicle models are very complicated mathematically • Too complicated to be used as is in algorithm • Need: To simplify models so they can be used • How: Expand Green’s functions in efficient basis functions. Wavelets have proven to be very useful in electromagnetic modeling

Physics-Based Models • Solution: Need data-adaptive basis functions (precludes multipole expansions) • Adaptive anisotropic wavelet basis functions which are non-separable are more general and allow direction-dependent resolutions • Precomputed basis functions for different physical situations (e. g. , forest types, season) • Try: “Best Basis” algorithm, “Basis Pursuit”

Detectibility Computer

Detectibility Computer • Issue : Should we use/deploy another sensor? • “Sensor”: Radar, acoustic, infrared and different frequencies & polarizations of radar (Both type and waveform of various sources) • Need: To compute value-added for another sensor: Improvement in E[detectibility - cost]. Choose the “sensor” which maximizes this. • Formulation: Dynamic stochastic scheduling

Detectibility Computer • “Cost”: Penalty for deploying another sensor: • Dollar cost of a UAV or other sensor times Pr[interception and destruction of new sensor] • Time cost in switching antennae types • Power cost in operating power and weight

Detectibility Computer • • • “Detectibility”: What does this mean? Optimal: Use min Pr[detection] as criterion. But: Too difficult to compute in real-time: Unknown target, unknown medium, etc. • Hence: Use easier-to-compute detectibility as a surrogate function for Pr[detection]. • Then: Choose “sensor” that maximizes E[detectibility-cost] based on previous data.

Detectibility Computer • • • “Detectibility”: What do we use for this? Renyi information divergence: This is: Much easier to compute (see next slide); Related to Pr[detection] by error exponent; Equals Kullback-Liebler distance between null model and most-likely target model for the special case a ~ 1. Choose 0 < a < 1.

Detectibility Computer • Renyi Information Divergence: RID • • • Log Pr[error] < (1 -a)RID; so figure-of-merit. Y = past data; N = Nth model; 0 = null model. Conditional densities computed quickly with sequential importance sampling (see previous) Also may use particle filtering.

Mine Detection: Acoustic+Radar • • • One possible scenario of combining modalities Acoustic source vibrates buried objects Vibration significant at resonant frequencies Radar source images vibrating objects Doppler radar spectrum exhibits sharp peaks at resonant acoustic frequencies of objects • Use to identify shape and material of objects • Use this information in turn to detect mines

Mine Detection: Acoustic+Radar

Mine Detection: Multiple Sensors • Problem: False alarms time-consuming: Must treat each as if it is a real mine • Problem: Failure-to-detects disastrous! • Single-sensor technology is insufficient • Hybrid sensor modalities seem necessary to attain both very low Pr[F] and high Pr[D] • Multi-modal approach seems promising

Tanks Under Trees: Radar Sensor • • Present work with ARL: Tanks Under Trees 3 -D MMW (millimeter wave) radar image: Ø Ka-band image of HMMWV on platform; Ø HMMWV parked under deciduous canopy • • (UM/ARL field experiment in July 2000) We are equipped for realistic work on this

Tanks Under Trees: Radar Sensor

Tanks Under Trees: Radar Sensor • Problems: Multiple scattering off of the: ground, trunk, leaves, branches, vehicle • Unknown presence and type of vehicles • Unknown orientation, location, reflectivity of vehicles (unknown parameters in models) • Unknown radar propagation characteristics through the atmosphere

Tanks Under Trees: Radar Sensor

Tanks Under Trees: Radar Sensor • Solutions: UM Radiation Lab has basic models for radar scattering off of vehicles • Parametrized by a few unknown parameters (position, orientation, reflectivity, etc. ) • UM Radiation Lab also has good models for radar scattering off of tree canopies • Parametrized by a few unknown parameters (average leaf area, branch and trunk size)

Tanks Under Trees: Radar Sensor • Concept: Use radar at different frequencies and polarizations as multiple modalities • For each modality, already have good parametrized models for vehicles and canopies • Combine these using previous sequential detection and sensor management algorithm to be developed as part of this research

Evaluation of Resulting Algorithms • UM Radiation Lab has a number of scattering models for vehicles and canopies • Permits realistic testing of algorithms • Compute ROC curves for various choices of: #sensors, sensor type, model dimension, noise • Figures-of-merit: power=Pr[detection] at fixed level of significance; area under ROC curve

Summary • Sequential detection and classification • Sensor scheduling and management • Physics-based models with dimensionality reduced using functional analysis • Vehicle and canopy scattering models already at UM permit test evaluations