Classifying Vehicles in Wide-Angle Radar Using Pyramid Match Hashing

We present a fast, scalable method to simultaneously register and classify vehicles in circular synthetic aperture radar imagery. The method is robust to occlusions and partial matches. Images are represented as a set of attributed scattering centers that are mapped to local sets, which are invariant to rigid transformations. Similarity between local sets is measured using a method called pyramid match hashing, which applies a pyramid match kernel to compare sets and a Hamming distance to compare hash codes generated from those sets. By preprocessing a database into binary hash codes, we are able to quickly find the nearest neighbor of a query among a large number of records. To demonstrate the algorithm, we simulated X-band scattering from ten civilian vehicles placed throughout a large scene, varying elevation angles in the 35°-59° range. We achieved better than 98% classification performance. Similar performance is demonstrated for a seven class task using airborne radar measurements.