Work

Methods and Means of Precise Attribution of Uav Photo-video Data at the Line of Contact for Training Neural Networks Based on a Dual-loop Architecture

Year: 2026
Type: article
Abstract: The article develops methods and software-hardware tools for the precise attribution of photo-video data obtained from an unmanned aerial vehicle under challenging observation conditions at the line of contact. It is substantiated that the quality of training datasets for neural networks is determined not only by frame resolution but also by the accuracy of the temporal, spatial, and inertial referencing of each data fragment. A dual-loop architecture is proposed in which the lower loop, based on a microcontroller, 12ISSN 2788-5518 performs deterministic acquisition of inertial module data, navigation parameters, and telemetry, while the upper loop, based on a single-board computer, captures video streams, initiates a synchronous request, and forms a structured dataset. A distinctive feature of the approach is the installation of an independent IMU sensor directly on the camera body, which enables recording of the actual orientation of the optical axis and reduces the impact of suspension errors. SPI is used for communication between the loops, ensuring predictable latency and enabling each frame to be associated with an attribute packet within a single time cycle. The structure of the attribution packet, the metadata generation algorithm, the model of homographic frame normalization, and parametric filtering procedures are considered. The practical significance of the work lies in improving the geometric correctness of training data, reducing hidden desynchronization, and laying the groundwork for creating high-quality computer vision datasets
Language: English
Subfield: Geology
Open Access status: diamond
OpenAlex ID: W7220696147
FWCI:
Cites: 2
Cited by:
Related to: 10