Image segmentation for improvised explosive devices

Abstract

This thesis creates algorithms to preprocess colored images in order to segment Improvised Explosive Devices (IEDs). IEDs are usually concealed and camouflaged and therefore more difficult to segment than other objects. We address the increased difficulty with three key contributions: 1) Our algorithm automatically divides a user-defined background area into smaller areas. We generate separate color models for each of these areas to ensure that a color model includes only colors that appear in the same area of the background. 2) We compress each of these complex color models into a statistical model. This increases the number of background models we can hold simultaneously in working memory, and allows us to generate a set of background models that describes a complete environment. 3) We estimate the initial object labels based on the color distance to the background. This approach allows us to generate color models for IEDs without user input that labels parts of the IED