Unmanned aerial vehicles (UAVs) on patrol that recognize scene changes are a powerful instrument in a lot of surveillance and reconnaissance applications, e.g., protection of convoys from improvised explosive devices (IEDs) by early detecting deployment traces, intruder or manipulation detection for real estates and military camps, or danger recognition and assessment in disaster monitoring. For that purpose, a visual-optical camera records video streams in which an automatic video change detection computes relevant or suspicious changes between two patrols. The main challenge is to identify true scene changes, i.e. to filter out changes caused by camera brightness variations, illumination changes due to cloudiness or time of day, or varying shadows cast by clouds or scene objects. Different partly overlapping or orthogonal solutions for all that, like chromatic adaptation, local mean adjustment, retinex, intensity equalization, and CIE Lab shadow removal, are evaluated and advantages/disadvantages are compared. Finally, the optimal method selection and combination is determined in order to maximize the change detection performance. Depicted results document the reached progress.
|