SMALL UAV (DRONE) DETECTION AND TRACKING USING IMAGE PROCESSING AND MACHINE LEARNING

Abstract

Small UAVs are commonly available and becoming a favourite pastime nowadays. However, the lack of public awareness is a major concern for the security of sensitive areas such as airports and military installations. This project aims to tackle an increasing threat to aviation security by applying known technology in an entirely novel and innovative way. It will explore the advanced image processing and analysis techniques for the detection and tracking of drones using machine learning algorithms. The main objective is to develop a framework of machine learning based techniques and evaluate the capabilities of available algorithms. Detect drones and classify their type are the main technical aims of this project. For the detection task, Faster R-CNN and YOLO (version 3) algorithms have been applied to perform the task. Followed by the classification task which has classified the drones into three types: single rotor, multi-rotor and fixed wing drones. Two models have been performed to classify drones’ type. The first model was built from scratch and the second one uses Transfer Learning. As a result, the average loss of 0.235 for Faster R-CNN and 0.04 for the YOLOv3 model have been achieved after the training. In addition, Faster R-CNN achieved a mAP score of 0.935 with IoU set at 0.5 and 0.605 with IoU set at 0.75, while YOLOv3 achieved a mAP score of 59.64 with IoU value set at 0.5 and mAP score of 11.82 with IoU value set at 0.75. More than 95 percent was the accuracy of the classification task. The work has been done to achieve offline (media files) and real-time (live stream cameras) processes.