Object Detection in Low Resolution Overhead Imagery

The proliferation of overhead image sensors has yielded benefits to many applications, including civil transportation, military reconnaissance, and environmental monitoring. Central to these and other applications is the ability to reliably detect and localize rigid man-made objects within large field-of-view images. In this paper we present an automated object detection approach that performs reliably on low-resolution imagery given limited annotated training data. Our two-stage system couples a fast sliding-window object detector with a more computationally-intensive, high-accuracy stage using ensembles of 2D templates. A thorough development process has led to contributions including: (1) multiple feature extraction during template matching; (2) formulation of template selection as a feature selection problem; and (3) the use of background models to normalize each template and image pair. For validation, we demonstrate successful detection of a specific aircraft model over a range of lighting conditions and operating environments.