Deep Learning Algorithms for Vehicle Detection on UAV Platforms: First Investigations on the Effects of Synthetic Training

Vehicle detection on aerial imagery taken with UAVs (unmanned aerial vehicles) plays an important role in many fields of application, such as traffic monitoring, surveillance or defense and rescue missions. Deep learning based object detectors are often used to overcome the resulting detection challenges. The generation of training data under different conditions and with the necessary variance is difficult and costly in real life. Therefore, virtual simulation environments are meanwhile often applied for this purpose. Our current research interests focus on the difference in performance, also called reality gap, of trainable vehicle detectors between both domains and the influence of differently designed training data. A general method for automatic image annotation with the required bounding boxes is described. In the first part of the investigations the training behavior of YOLOv3 on the natural UAVDT data set is analyzed and examined to what extent algorithms trained with natural images can be evaluated in the simulation. Finally, it is shown which performance can be achieved by exclusively synthetic training and how the performance can be improved by synthetic extension of the natural training set.