An Improved Car Detection using Street Layer Extraction

Automatic 3D modeling of urban environments has recently become a hot research topics since location aware applications (e.g. Virtual Earth, Google Earth) on the Internet aim for detailed models of the earth. In large scale 3D city models, moving objects, such as cars usually bear errors and distract the automatic reconstruction process. Therefore it is desirable to detect and remove these objects. This work introduces a strategy for extracting a street layer by using 2.5D height data and color information and demonstrates how this layer can be applied to improve the car detection process on high resolution aerial images. We detect street related objects, such as zebra crossings, to extract control points of a street layer, on which an occurrence of cars is accurately defined. These control points help to generate a Digital Terrain Model (DTM) and a color model for the street network. By introducing this information in the car detection process, the false positive rate can be decreased considerably. The extracted street layer is evaluated on available classification results and hand labeled ground truths. Furthermore, we compare the achieved performances of our approach to state-of-the-art car detection results on aerial images.