Comparison of Correlation Filters for Vehicles Detection in High Resolution Airborne Imagery

The use of electro-optical imagery acquired over urban areas is becoming an important way to detect and recognize different kind of vehicles for various applications: highway monitoring, traffic management, air pollution assessment, and vehicle fleet management. Indeed, for such applications, high resolution imagery gathered from airborne platforms provides the ability to present a synoptic view of traffic and an enhanced capability to monitor traffic trajectories. Nonetheless, high resolution imagery requires advanced image analysis algorithms to extract each object of interest as a whole, instead of detecting each part of the object at the pixel level. This paper provides a comparison of correlation filters for the classification of vehicles using high resolution airborne imagery, in order to assess how correlation filters could be helpful in the detection of Sport Utility Vehicles (SUVs) in large images or image databases (image mining). The filters considered were the standard matched filter (MF), the phase only filter (POF), the tandem component filter (TCF), the locally adaptive contrast-invariant filter (LACIF), and the locally nonlinear matched filter (LNMF). As the LACIF and LNMF filters contain the others as special cases, theses two correlation filters were the ones specifically evaluated in this research. ROC curves were calculated for the LACIF and LNMF filters on airborne images that include parking lots containing up to 1000 vehicles. The experiments showed that the LACIF filter yields the best detection results (over 80% correct classification) over the LNMF filter.