ANALYSIS OF EXISTING OBJECT DETECTION MODELS FOR UAV DETECTION ON THERMAL IMAGERY

Nowadays, unmanned aerial vehicles (UAVs) are becoming widely used. They have proven their effectiveness, reliability, and feasibility. However, improper use or abuse of this technology can lead to significant human rights violations and threaten public safety. In connection with this, appropriate methods of countering UAVs should be made. One of the areas that is actively developing nowadays is the detection of UAVs based on optoelectronic radiation. Since most of such systems are designed for the visible range, the IR range has not been so widely investigated. UAV detection based on thermal imaging can be performed using artificial intelligence. Currently, there are ready-made approaches and models that perform object detection tasks, but they are more general purpose. To review the problem of UAV detection based on thermal imaging and AI, it is necessary to analyze the most effective object detection models for the problem of UAV detection in the IR range. In this work, an analysis and comparison of such object detection models as YOLOv5 and YOLOv8, Faster RCNN, DETR was carried out. Datasets were used for model training, both in the IR range and in the visible range. Preliminary digital processing of the dataset of UAV images in the visible range was carried out to convert them into pseudo-thermal images. This was done in order to increase the amount of training data for the models and thereby improve their accuracy. The results showed that models such as YOLOv8 and DETR are the most effective for UAV detection tasks in thermal imaging, but their accuracy is still insufficient to be effectively used for real-time UAV defense systems.