A COMBINATION OF FASTER R-CNN AND YOLOv2 FOR DRONE DETECTION IN IMAGES

Abstract

Today, drones are widely used for different purposes since they are not too expensive. Drones employed as explosive material, camera and illegal thing carriers can cause security threats. Computer vision can be applied to detect illegally acting drones effectively in a variety of conditions. A computer-based system using modern cameras is possible to discover small distant drones. The system can also become aware of low-speed and non-ground controlled drones. Furthermore, the system can display true drones. This makes the system friendly to users. This paper proposes a hybrid approach combining two emerging convolutional neural networks: Faster R-CNN and YOLOv2 to detect drones in images. Experimental results show that the approach can add up to almost 5% and more than 11% to precision and recall for Faster R-CNN and add up to 3% and more than 6% to these two metrics for YOLOv2. This better detection is resulted from the combination of the two networks. If a network is failed to detect drones in an image, the other network can help.