UAV Inertial Navigation Correction System Based on Deep Learning for Matching Camera Images and Satellite Imagery

Abstract

Reliable navigation of unmanned aerial vehicles (UAVs) is a critical function that relies heavily on GNSS signals. In electronic warfare (EW) environments, these signals are suppressed, leading to the loss of vehicles due to the uncontrolled accumulation of errors in onboard inertial navigation systems (INS). Existing visual localization methods are often computationally expensive and suboptimal for rural landscapes.

The aim of this work is the development and experimental validation of a computationally efficient hybrid autonomous geolocation system capable of correcting INS errors by matching images from an onboard camera with reference satellite imagery, with a focus on viability for low-resource onboard platforms.

A hybrid method combining INS data with discrete correction from a deep learning module is proposed. A custom dataset of flights in rural areas was collected and labeled. A comparative analysis of CNN architectures (VGG11, MobileNetV2, ResNet18) and loss functions for the cross-modal matching task was performed. The accuracy and performance of the final model were validated on a Raspberry Pi single-board computer.

It was experimentally demonstrated that Triplet Loss, thanks to its mechanism of selecting "negative" examples, forces the model to learn discriminative features, enabling it to distinguish visually similar areas of the landscape. In comparative testing of architectures, the VGG11 model trained with Triplet Loss showed the best accuracy, outperforming MobileNetV2 and ResNet18. When testing the correction mechanism with a simulated initial 10-meter INS drift, the VGG11-based system achieved an average error of 5.4 meters, successfully calculating the correction vector. On a real test route of 200 m, where the maximum error of the unmanaged INS reached 15.2 m, the proposed system reduced the maximum deviation to 8.5 m. Performance validation on a Raspberry Pi 4 confirmed computational suitability: a full correction cycle for a 224x224 image takes 2.5 seconds, and 10 seconds for 448x448, which is acceptable for periodic (non-continuous) coordinate refinement.