Adaptive image analysis for aerial surveillance

Abstract

The authors describe a computer vision system that would automatically monitor its own performance and dynamically adapt to changing situations and requirements. Their approach is based on computational reflection and control system theory. An adaptive approach to aerial surveillance image analysis has these related benefits: the image analysis process can smoothly adapt to changes in image quality that would otherwise lead to abrupt changes in interpretation quality; appropriate resources can be selected for an image's needs in a way that allows the efficient use of resources; the system requires less tinkering to make it work; reflection allows the knowledge of the filters, task, and imaging environment to guide the automatic integration of filter results; the image bank provides a convenient mechanism for converting expert (photo interpreter) image analysis knowledge into a form usable for control and adaptation. However, this approach has two problems: filter implementation requires significant additional effort; and stability analysis is typically part of control system design, but generalizing stability analysis for image analysis might be difficult.