| Abstract [eng] |
The rapidly increasing use of unmanned aerial vehicles in civil and military fields highlights the need to develop systems for their detection, identification and tracking. One possible component of such a system is computer vision methods, which enable drones to be detected and localized in images and video recordings. The aim of this project is to investigate drone image detection and tracking algorithms by evaluating their accuracy, speed and suitability for real-time systems. Six tasks were formulated to achieve this aim. The project consists of an introduction, three chapters, conclusions, lists of references and information sources, and appendices. The theoretical part of the project examines modern object detection and tracking methods based on convolutional neural networks, detection transformers and tracking-by-detection principles. In the experimental part, a system was developed for processing images and video recordings, detecting and tracking objects, saving and visualizing results, and evaluating performance. The system was implemented using the Ubuntu 24.04.3 LTS operating system, Python, PyTorch, OpenCV, Ultralytics YOLO, FFmpeg, TrackEval and GPU monitoring tools. The Anti-UAV, drone- datasets and UAVSwarm datasets were used for the research. In the research part, the YOLO26-X, RF-DETR-XL, DEIMv2-X, DEIM-D-FINE-X and DEIM-RT- DETRv2-X detectors were compared. It was found that when the unmanned aerial vehicle occupied 2% of the frame area, the RF-DETR-XL detector achieved the best mAP50-95 result of 55.7%. In the case of very small objects, when the drone occupied 0.2% of the frame area, the most accurate detector was DEIM-D-FINE-X, achieving 32.3%. The fastest detector was YOLO26-X, with an average frame processing time of approximately 72 ms. The study of images affected by noise showed that motion blur had the greatest negative impact on accuracy, while in general the DEIMv2-X detector was the most robust. In the detection and tracking pipeline study, the best result was achieved by the combination of RF-DETR-XL and BoostTrack++, reaching 67.2% MOTA, 54.8% HOTA and 70.9% IDF1. The project also presents a control concept for an unmanned aerial vehicle tracking and neutralization system based on detector and tracker outputs. |
