Technical Report Number
Aerial video provides strong cues for automatic road extraction that are not available in static aerial images. Using stabilized (or geo-referenced) video data, capturing the distribution of spatio-temporal image derivatives gives a powerful, local representation of scene variation and motion typical at each pixel. This allows a functional attribution of the scene; a “road” is defined as paths of consistent motion --- a definition which is valid in a large and diverse set of environments. Using a classical relationship between image motion and spatio-temporal image derivatives, road features can be extracted as image regions that have significant image variation and a motion consistent with its neighbors. The video pre-processing to generate image derivative distributions over arbitrarily long sequences is implemented in real time on standard laptops, and the flow field computation and interpretation involves a small number of 3 by 3 matrix operations at each pixel location. Example results are shown for an urban scene with both well-traveled and infrequently traveled roads, indicating that both can be discovered simultaneously. This method works robustly in scene with significant traffic motion and is thus ideal for urban traffic scenes, which often are difficult to analyze using static imagery.
Pless, Robert and Jurgens, David, "Road extraction from motion cues in aerial video" Report Number: WUCSE-2004-30 (2004). All Computer Science and Engineering Research.