Detecting building changes from multitemporal aerial stereopairs

摘要

Our goal is to detect changes in an aerial scene by comparing grey scale stereopairs taken several years apart in order to update a geographic database. A set of image locations that have a high likelihood to contain changes will be submitted to a human operator who will either reject the proposed change or validate it and update the database accordingly. We are mainly interested in changes in buildings. To isolate new construction and buildings, which disappear, we provide an algorithm that works in two steps. First, during a focusing phase, we eliminate a large part of the scene without losing any actual changes by comparing a Digital Elevation Model (DEM) for the two dates. Second, we classify the resulting regions of interest (ROI) based on four images-stereopairs of the area at the two dates. To decide whether or not the ROI contains a change, we classify each of the four, images as "building" or "no-building". This classifier is a combination of several decision trees induced from training data. Each node of each decision tree is identified with a graph of features which is more likely to occur on buildings than background. Finally, the classification results at the two different dates are compared. The final set of locations submitted to an operator omits less than 10% of the true changes. The false positive rate represents less than 5% of the scene surface.