Surveillance Image Processing on Embedded Systems

Abstract

Imaging through turbulence is deeply rooted in astronomy and over the centuries as- tronomers have conceived various means to try and improve their view of the heavens. Surveillance imaging through turbulence is a comparative new-comer to the scene and although superficially similar there are many important differences. The primary astronom- ical techniques of site selection and patience are rarely applicable to surveillance and its covert nature places limits on the available on-site equipment as well as the use of aides such as laser reference sources. Beyond these logistical issues however is a more fundamental problem. Light from astronomical sources is normally highly spatially coherent and often isoplanatic. Surveil- lance imaging, by contrast, is normally severely anisoplanatic and images are acquired under incoherent conditions. Because of the inherent size, power and weight constraints associated with surveillance this thesis explores potential algorithms with a view to even- tual application on embedded devices. Unfortunately some widely used techniques bor- rowed from astronomy such as speckle-interferometry are shown to be no more effective than considerably simpler algorithms when applied to images acquired under surveillance conditions. Other algorithms are modified both to improve their overall accuracy and their computational efficiency. A new algorithm is also presented, that relies on assumptions regarding image content, and is shown to be useful in some surveillance situations. Because reconfigurable hardware devices such as FPGAs offer unmatched embedded performance the new algorithms presented are further explored in terms of their efficiency on reconfigurable devices. Further, in order to lower the considerable development time associated with working with FPGAs the original work of this thesis concludes by introduc- ing an open protocol design to facilitate rapid development of image processing algorithms on FPGAs. In the last chapter I conclude with a discussion of where I see the future of surveillance imaging and an analysis of factors that I believe are limiting the advancement of the field.