Development and research of a hardware and software system based on a single-board computer using neural network models for the task of recognizing objects of interest on the Earth's surface

Problem. The problem of recognizing objects on the Earth's surface using hardware and software systems based on single-board computers and neural networks is of particular relevance in the context of the development of autonomous aviation systems based on small aircraft. Currently, there is a need to create compact energy-efficient hardware and software solutions capable of analyzing optical video frames in real time. However, the use of single-board computers on board of small aircraft with the realization of neural networks often faces the problem of limited computing power, requirements to their power consumption and mass-size characteristics for their work in real time. With the increase in the volume and detail of video frames of the Earth's surface, received from onboard sensors, there is an opportunity to improve the quality of object recognition through the use of modern neural networks. However, it is difficult to directly implement neural networks in systems based on single-board small-size computers because of the complexity of their architectures, which require more computational resource than is available in such computational on-board facilities. Therefore, it is necessary to adapt neural networks to the resources of used single-board small-size computers. Purpose. To analyze single-board computers and neural network models for the realization of aviation hardware-software vision system, controlling agricultural objects of the earth surface on the basis of small aircraft. Results. The research of efficiency of neural networks realization on the basis of small-sized single-board computer and capable to be based on a small aircraft has been carried out. It was found that the implementation of neural network “YOLOv8n” on the basis of single-board computer “Raspberry Pi 4 Model B” demonstrates the best efficiency over other neural network architectures when recognizing haystacks, providing the required speed of video frames processing and a high proportion of correctly recognized objects of interest. Practical relevance. The presented results provide automation of the processes of agricultural land control in real time. The proposed system based on the neural network “YOLOv8n” and single-board computer “Raspberry Pi 4 Model B” allows to quickly recognize haystacks, which is important for planning harvesting operations, controlling the quality of forage and reducing yield losses.