With the rise in popularity of drones and due to various security reasons that come along with them, there has also been a demand for systems that can detect them. In this diploma thesis, we address the problem of drone detection using computer vision methods and propose a new long-term tracker specialized for drone detection and tracking. The proposed tracker consists of a short-term tracker, a detector, and a failure detection module. The system works in such a way that, during tracking, the failure detection module constantly checks whether the short-term tracker has failed and, if so, reinitializes it using the detector. It detects the failure using two subsystems, the first relying on periodically triggering the detector and checking that detections match the tracker, and the second relying on the tracker's confidence. When creating the architecture of the method, we also paid special attention to the speed of execution. We propose three variations of the long-term tracker, which differ in speed and accuracy. The most accurate of them achieves F1, which is 31.8 % higher than the YOLOv5 detector, and the fastest method is 17.6 % faster than the aforementioned detector and processes images at a speed of 173 frames per second. The third version of the detector represents a good compromise between speed (4.1 % faster than YOLOv5) and accuracy (29.5 % higher F1 measure than YOLOv5).
|
