Angle Robustness Unmanned Aerial Vehicle Navigation in GNSS-Denied Scenarios

TL;DR

This work tackles navigation for UAVs in GNSS-denied conditions by reframing the problem as end-to-end angle prediction rather than position estimation. It introduces an angle robustness navigation model built from an Adaptive Feature Enhance Module, a Cross-knowledge Attention-guided Module, and a Robust Task-oriented Head, leveraging a transformer to fuse history frames into a global embedding and predict the direction $(\sin{\theta}, \cos{\theta})$. The authors also contribute UAV_AR368, a large point-to-point navigation dataset, and SFTI, a Google Earth–based simulator to evaluate vision-based navigation under diverse disturbances. Experiments show substantial improvements over baselines in both ideal and disturbed conditions, with high arrival success rates, reduced end-point and route errors, and a lightweight, fast model suitable for deployment in GNSS-denied environments.

Abstract

Due to the inability to receive signals from the Global Navigation Satellite System (GNSS) in extreme conditions, achieving accurate and robust navigation for Unmanned Aerial Vehicles (UAVs) is a challenging task. Recently emerged, vision-based navigation has been a promising and feasible alternative to GNSS-based navigation. However, existing vision-based techniques are inadequate in addressing flight deviation caused by environmental disturbances and inaccurate position predictions in practical settings. In this paper, we present a novel angle robustness navigation paradigm to deal with flight deviation in point-to-point navigation tasks. Additionally, we propose a model that includes the Adaptive Feature Enhance Module, Cross-knowledge Attention-guided Module and Robust Task-oriented Head Module to accurately predict direction angles for high-precision navigation. To evaluate the vision-based navigation methods, we collect a new dataset termed as UAV_AR368. Furthermore, we design the Simulation Flight Testing Instrument (SFTI) using Google Earth to simulate different flight environments, thereby reducing the expenses associated with real flight testing. Experiment results demonstrate that the proposed model outperforms the state-of-the-art by achieving improvements of 26.0% and 45.6% in the success rate of arrival under ideal and disturbed circumstances, respectively.

Figures (4)

• Figure 1: Comparison between the existing methods (a) Classification-based, (b) Matching-based methods and (c) Ours. (a) and (b) involve predicting the position and then calculating the direction angle. The inaccurate position prediction will lead to cumulative flight errors. In contrast, our method (c) directly predicts the direction angle and effectively solves flight deviation.
• Figure 2: Architecture overview of the proposed method. Firstly, the Adaptive Feature Enhance Module extracts both sequential feature-level and position-level features $R^{I}$, $R^{P}$. Then, the Cross-knowledge Attention-guided Module integrates these semantic features to a global perspective embedding $Z'_{K+1}$. Finally, the Robust Task-oriented Head Module utilizes the embedding to adaptively predict the direction angle $\theta$.
• Figure 3: Samples of simulated flight testing environments.
• Figure 4: Visualization results w/o and w/ disturbances on UAV_AR368. The subgraphs in the coordinate systems above indicate the MRE along the testing routes. The gray dash-line boxes below illustrate the flight routes generated by different methods, where the successful and ideal routes are denoted by green / blue / red and black lines.