UAV detection and identification using a convolutional neural network

UAVs may pose intentional or unintentional risks to airspace safety or violate non-flying zones in the vicinity of vulnerable ground objects. Detecting and identifying UAV threats is a challenging task due to the availability of commercial drones with extended operating ranges and payload capacity. This paper presents a proposal for a two-stage computer vision-based system for detection of the flying objects and the identification of drone types if the drone is detected in the previous step. A dataset containing 8142 labeled images was collected, or synthetically generated and used for detection model training while another dataset containing 2679, mainly synthetic images is used for classification (identification) model training. The detection system based on YOLOv9 showed superior performance, with 0.978 mAP50 and 0.767 mAP50-95, while the classification model using ResNet-152 performed best in the challenging task of classifying the drone model with 88% accuracy.