Abstract
The availability of unmanned aerial vehicles (UAVs) has led to a significant increase in the number of offenses involving their use. This makes the development of UAV detection systems relevant. Solutions based on deep neural networks show the best results in detecting UAVs on video. This article presents a study of various neural network detectors and focuses on identifying objects as small as possible, up to \(4 \times 4\) and even \(3 \times 3\) pixels. This paper studies SSD (VGG16) and YOLOv3 architectures and their modifications. Precision and recall metrics are calculated separately for different intervals of the object areas. The best result is obtained for the YOLOv3 model with the bbox parameter values selected through the clustering of object sizes. Small ( \(3 \times 3\) px) drones are successfully identified with 76% precision and a very small recall of 26%. For objects between 10 and 20 pixels in area, the recall is 64% with precision of 75%. For objects with an area more than 20px the recall is about 90%, the precision is 89%, and the F1-measure is 90%. These results show that it is possible to recognize even \(4 \times 4\) pixel drones, which can be used in video surveillance systems.
