A fast screening method for detecting cars in UAV images over urban areas

The paper presents a fast screening method to isolate asphalted areas in urban images acquired with unmanned aerial vehicles (UAV). The screening is a key stage of a standard car detection and counting approach allowing to improve the computational time and reduce the number of false alarms. The proposed screening method subdivides the original UAV image into tiles which are then considered separately. From each tile a signature which represents the color information of the scene is extracted and compared with a training library to find the most similar tile. In the context of this work, two matching strategies have been considered. Promising experimental results are conducted on real UAV images acquired over urban areas. In particular, we show the accuracy of the screening approach compared with two reference techniques. In addition, in the last part of the work, we analyze the influence of the different masking methods on a car detection and counting approach.