Visual positioning of UAV over terrain using computer vision methods

Abstract

The paper proposes a method for visual positioning of unmanned aerial vehicles (UAVs) based on descriptor-based computer vision algorithms. The use of optical navigation is investigated as an auxiliary mechanism under conditions of degradation or total absence of global navigation satellite system (GNSS) signals, where standard positioning is unavailable or unreliable for autonomous flight. The developed approach is based on the preliminary formation of a reference aerial image database and its subsequent matching with the current video stream from the UAV’s onboard camera. To simulate real operational conditions, the reference fragments were subjected to synthetic transformations: affine distortions, scaling, rotation, shearing, brightness variations, Gaussian noise, and partial occlusion. This allowed modeling destructive factors caused by shooting angles, illumination, flight dynamics, and sensor noise. A comparative analysis of the prepared samples with the onboard camera frames was performed using SIFT, BRISK, and ORB methods. The efficiency of the algorithms was evaluated based on the criteria of computational processing complexity, matching accuracy, and the number of detected keypoints. It is experimentally proven that the ORB algorithm demonstrates an optimal balance between identification accuracy and computational complexity, making it promising for integration into real-time onboard UAV orientation systems.