Abstract from Univ. of Michigan DARPA-ARO Proposal


Sequential Adaptive Multi-Modality Target Detection

and Classification Using Physics-Based Models

The goal is to develop adaptive algorithms for detection and classification of the following: (1) vehicles under forest canopies; and (2) land mines. The data will be obtained from a variety of sensing modalities, including radar, optical and thermal emissions.

The approach will be to start with detailed physics-based radar propagation and scattering models for canopies, vehicles, and land mines already developed here in the Radiation Laboratory of the University of Michigan. These coherent models are capable of preserving detailed structure of tree canopies (species, density, vegetation moisture content, etc.) and also account for near- and far-field interactions among vegetation particles, as well as their interactions with embedded targets.

We will derive simpler models for these physics-based models which capture the relevant physics of each modality while preserving features essential to adaptive target detection algorithms. Hierarchical representations of the forward model will be the first avenue taken to accomplish model reduction. For this we will start by applying a recently developed theory of multi-dimensional anisotropic wavelets whose footprints can be matched to spatio-temporal resolution ellipsoids specified by the modality. Other low-dimensional models, obtained using functional approximation theory, will also be investigated.

The resulting lower-dimensional model will be used in a sequential adaptive detection procedure in which information from one sensor modality is used as a first stage to adaptively alter source waveforms and arrays for another modality. How to select and sequence modalities is a research question to be addressed. Modalities and sensors to be considered include: imaging radar at various frequencies, polarizations, and resolutions; thermal imagers at various frequency bands; and lidar.

The expected result of this research is a new approach to solving the above inverse problems by combining different sensor modalities. Specific algorithms for each of the above problems will be developed, along with a set of performance measurements obtained from both Monte Carlo simulations on existing detailed forward problem models, and from real data.