AUTOMATIC DETECTION OF HIDDEN OBJECTS USING UAVS: MODERN NEURAL NETWORK APPROACHES

Abstract

The article presents a comprehensive analysis of modern approaches to the automatic detection of camouflaged objects using unmanned aerial vehicles. The study focuses on reviewing deep learning architectures and computer vision methods that are actively used in the military to improve the accuracy and speed of aerial reconnaissance systems. Particular attention is paid to models of the YOLO family, in particular YOLOv8, as well as advanced architectural solutions — Cross Stage Partial Networks (CSPNet), Spatial Pyramid Pooling Fast (SPPF), and lightweight variants of LightCSPNet, which are focused on mobile platforms. The role of high-quality and balanced datasets, annotation tools, and image preprocessing methods in ensuring the high efficiency of recognition systems is analyzed. The results of experimental studies demonstrating the ability of modern models to achieve accuracy of over 95% even in difficult conditions, such as smoke, fog, or the use of camouflage nets, are summarized. Key problems related to the detection of small objects, the lack of diverse datasets, and the integration of algorithms into resource-constrained onboard systems are identified. Based on the review, it was concluded that hybrid architectures and multispectral approaches are promising for ensuring a new level of efficiency in aerial reconnaissance systems