Development of a classification model for UAVs and birds based on the YOLOv9 neural network to improve anti-drone systems

The article presents the materials of the development of a model for classification and recognition of UAVs and birds based on the neural network of the YOLOv9 architecture in the optoelectronic channels of Anti-drone systems. To train the neural network, a dataset was prepared in the form of annotated images of UAVs and birds. The total number, taking into account augmentation, was 5265 images. The authors implemented training, verification and testing of neural networks in the Windows 11 operating system, in the Python 3.10.8 runtime environment and the Pycharm 2024 development environment. The training process was carried out on the basis of the AD103 graphics processor of the NVIDIA GeForce RTX 4080 video card with support for CUDA Toolkit 12.1. As a result of training the neural network, the following metrics were obtained: mAP50-95: 0.59; mAP50: 0.95; Recall: 0.89; Precision: 0.95. According to these indicators, the trained model outperforms the UAV and bird recognition and classification models trained on the basis of YOLOv2, YOLOv4, YOLOv5, YOLOv7 and YOLOX. The inference results on two videos with DJI Inspire 2 and DJI Mini 3 UAV flights showed FPS values of 131 and 119, respectively. It was found that, due to the obtained accuracy and FPS metrics, the trained YOLOv9 model can be used as a module for recognizing and classifying UAVs and birds in real time in the optoelectronic surveillance channels of Anti-drone systems.