Neural Network-Driven UAV Course Correction Using Camera Images

Abstract. This paper presents a UAV course correction algorithm leveraging a neural network and the analysis of a video stream from the onboard camera. The algorithm is designed to identify key landmarks efficiently and requires only a limited set of training images. Significantly, it demonstrates operational capabilities at viewing angles and altitudes that differ from those used during training. Experimental results indicate that the algorithm achieves satisfactory landmark recognition accuracy even with substantial perspective deviations, thus enhancing the robustness and effectiveness of UAV operations.