Development of deep learning based techniques for Unmanned Aerial Vehicle detection and classification

Nowadays, the interest in Unmanned Aerial Vehicles (UAVs) is increasing due to the increase in application areas and autonomous flight capabilities. The increase in the usage areas of UAVs also brings some security problems of these devices. Security issues can also be diversified in terms of its internal structure. These areas can be listed as the safety of living things such as humans, animals and nature, all kinds of military security and environmental security. It can be called public or national security in general. Therefore, it is important to take precautions for the early detection of threats originating from UAVs. This study is supported by experiments to examine the methods for detecting a UAVs in flight related to public safety and to develop new methods. In this context, two multidimensional methods, acoustic and image-based, are examined. You Look Only Once (YOLO) models are compared on the image base. In the literature review, it is seen that there is no study and no produced results related to You Look Only Once Representation (YOLOR) for UAVs detection. It is presented that the most effective result among the tested models is provided by YOLOR. YOLOR provides a 95.6% success rate on the run dataset. Mel Frequency Spectrum coefficients techniques are applied to extract the sound characteristics required to distinguish the sounds of UAVs depending on the acoustic base, which is another approach type. A light weighted convolutional neural network model is proposed for the classification of these clusters, with 6 classes of sounds whose properties are determined. With the proposed method, it is possible to detect UAVs and distinguish a flying İHAe from sounds such as bird, airplane, helicopter, background noise and storm. Experimental results show that the proposed method is 98.78% successful and better results are obtained than other methods suggested in the literature.