Drone detection using YOLOv3 with transfer learning on NVIDIA Jetson TX2
Daniel Tan Wei Xun, Yoke Lin Lim, Sutthiphong Srigrarom
Abstract
The rise of drones in the recent years largely due to the advancements of drone technology which provide drones the ability to perform many more complex tasks autonomously with the incorporation of technologies such as computer vision, object avoidance and artificial intelligence. However, the misuse of drones such as the Gatwick Airport drone incident resulted in major disruptions which affected approximately 140,000 passengers. To deter this from happening in the future, drone surveillance are extremely crucial. With this, it will be achieved firstly by detection and followed by tracking of drones. This paper presents and investigates the use of a deep learning object detector, YOLOv3 with pretrained weights and transfer learning to train YOLOv3 to specifically detect drones. We demonstrated that the detection results from YOLOv3 after machine learning had an average accuracy of 88.9% at input image size of 416×416. Finally, we integrated into NVIDIA Jetson TX2 for real-time drone detection.