Detection and tracking of drones in infrared images

Inês Cristina Catanas Forte Reis Pinto

Key information

Authors:

Inês Cristina Catanas Forte Reis Pinto (Inês Cristina Catanas Forte Reis Pinto)

Supervisors:

José Silvestre Serra da Silva; Alexandre José Malheiro Bernardino (Alexandre José Malheiro Bernardino)

Published in

December 5, 2024

Abstract

The rising use of small unmanned aerial vehicles (UAVs) poses significant safety and security challenges, necessitating effective detection and tracking systems. This thesis investigates models leveraging infrared (IR) imaging to address these challenges in varied environments. While RGB imaging provides detailed visuals, its performance declines in low-light conditions or with obstructions like fog or smoke. IR thermal imaging, however, relies on heat signatures for accurate detection in darkness, improving differentiation between drones and false positives. Two detection and tracking systems were developed using advanced machine learning algorithms. The first combines the Modified YOLO-DRONE detector with BoT-SORT, and the second pairs YOLO11 with ByteTrack. The Modified YOLO-DRONE incorporates a large detection head to handle drones of various sizes and adapts the YOLOv8-OBB architecture for datasets with oriented bounding boxes. Evaluations were conducted on three datasets: the New IR Multi-Drone Dataset (SWIR and LWIR), the DUT-AntiUAV dataset (RGB), and the Anti-UAV Challenge dataset (IR). Model 1 achieved a mAP@0.5 of 0.797 and a MOTA of 75.4%, while Model 2 scored 0.782 and 74.1%. On the New IR Multi-Drone Dataset, both models reached a mAP@0.5 of 0.993 for drone detection. On DUT-AntiUAV, Model 1 recorded a mAP@0.5 of 0.906 and a tracking precision of 0.940, versus 0.910 and 0.934 for Model 2. For the Anti-UAV Challenge dataset, Model 1 attained a mAP@0.5 of 0.901 and a MOTA of 69.4%, while Model 2 scored 0.905 and 56.9%. These datasets and tools will be publicly released, driving IR research for UAV-related security challenges.