Application of Computer Vision Algorithms Onboard Flying Drones for Earth Observation

Flying drones are one of the modern and effective means of earth observation. Most often, drones transmit streaming video to ground observation points, but in certain conditions this may not be acceptable. The paper proposes to detect objects on board the drone based on CV algorithms and transmit limited data. Research on determining the most effective computer software algorithms and energy-efficient platforms for their implementation is given. The architecture of convolutional neural networks was analyzed. The study was carried out on the basis of neural network training on VEDAI dataset with vehicle objects. The peculiarity of this dataset is that the area of objects is several percent of the image area, which corresponds to the conditions of earth observation. The method of weights was used to analyze and select the most effective platform of single-board computers for the implementation of CV algorithms. The results of the research can be used in standard tasks for EO, such as observing climate change, monitoring environmental problems, studying flora and fauna, etc. Moreover, proposed solutions can significantly improve the military domain by reducing the risk of interception and destruction of flying drones.