Please use this identifier to cite or link to this item:
|Title:||Deep learning based tools for drone surveillance and detection in aerodromes (dynamic images)||Authors:||Leong, Kai Feng||Keywords:||Engineering::Aeronautical engineering||Issue Date:||2022||Publisher:||Nanyang Technological University||Source:||Leong, K. F. (2022). Deep learning based tools for drone surveillance and detection in aerodromes (dynamic images). Final Year Project (FYP), Nanyang Technological University, Singapore. https://hdl.handle.net/10356/158036||Project:||C030||Abstract:||With a surge in unauthorized drone intrusion incidents, counter-unmanned aerial vehicle (UAV) systems have become a key area of focus for civil airport authorities. Most commercial C-UAV systems use a combination of radar and sensors to detect and jam the radio signal between the operator and the drone. Information gathered by the system is fed back to a control center at the airport and disseminated to relevant parties. However, by the time the information is received by the air traffic controllers, the drone might have already crashed into an oncoming aircraft. Hence, there is a need for a program which is able to detect and track illegal drone intrusions and display it directly for the controllers to take immediate action e.g. cease all runway operations. In this report, a program was developed based on two deep learning based algorithms: (1) YOLOv4 and (2) DeepSORT. The first algorithm, YOLOv4 serves as the object detection model and outputs a series of detections per frame together with the confidence levels. The second algorithm, DeepSORT, serves as the object tracker which issues a numbered identity to each drone and tracks their historical trajectory path. The development involved extensive data collection and preparation through the use of the Tower Simulator in ATMRI to simulate drones in the vicinity of Singapore Changi Airport. Both of the algorithms were trained and evaluated against a set of videos of drones flying within Singapore Changi Airport. The YOLOv4 model achieved a mean Average Precision (mAP) score of 86.39% and offered good performance for detecting drones up to 4 kilometers. The same could be said for the object tracker, DeepSORT. However, there were some instances of identity switching which occurred as a result of intersecting paths between a ground-truth and a false positive. Both of the algorithms were observed to offer better performance when used in a multi-object scenario as opposed to a single object detection/tracking task. Generally, the performance of both algorithms suffer as the distance of the drone gets further away from the Changi Control Tower. Hence, it was suggested that a human air traffic controller is still required to verify the information sent by the program and determine the next course of action.||URI:||https://hdl.handle.net/10356/158036||Schools:||School of Mechanical and Aerospace Engineering||Fulltext Permission:||restricted||Fulltext Availability:||With Fulltext|
|Appears in Collections:||MAE Student Reports (FYP/IA/PA/PI)|
Updated on Sep 25, 2023
Updated on Sep 25, 2023
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.