EGENIOUSS: An End-To-End Multi-Sensor Navigation Framework for Complex Urban Environments
Satellite-based positioning, navigation and timing (PNT) underpins the EU’s critical infrastructure. Safeguarding and enhancing it is strategically important because many sectors including communications, energy, transport, defence and finance rely on it. GNSS signals, however, can be jammed or spoofed, are sensitive to atmospheric effects, and suffer from multipath and signal blockage in dense natural or urban environments. These vulnerabilities create worldwide risks and slow the adoption of GNSS for mission- or safety-critical operations such as drone delivery, urban mobility and autonomous driving. Consequently, civilian and military bodies are advancing complementary or backup capabilities, known as “augmentations,” ranging from local aids such as inertial sensor support to Galileo-specific features, such as Open Service Navigation Message Authentication (OSNMA) or High Accuracy Service (HAS). Yet, delivering robust, precise, drift-free augmentations at any scale remains challenging and costly.
To address the inherent issues of GNSS and provide an additional, complementary source of absolute positioning in complex urban environments, we propose an end-to-end framework for multi-sensor fusion. By tightly integrating GNSS measurements with visual odometry (VO) and a cloud-based module for visual localisation (VL), we demonstrate decimetre-grade accuracy on low-cost hardware. The backbone of our approach is a sensor fusion system called NEXA, which combines raw GNSS signals and augmentations, IMU data, velocity estimates coming from the VO, and absolute pose estimates obtained via the VL.
Given an initial, coarse position estimate, the framework supports both offline and online operating modes for VL. Reference descriptors can be pre-computed and deployed to the edge device prior to operation, or retrieved on demand from a cloud service, depending on connectivity and application constraints. In both cases, visual localisation is performed on the edge device, ensuring low latency and bandwidth usage while preserving user privacy, as query images never leave the device. The visual localisation module provides absolute metric 6-DoF pose estimates, which are injected into the multi-sensor fusion as external updates alongside GNSS, visual odometry and inertial measurements. The framework is inherently scalable: new geographic areas can be incorporated incrementally on a city-by-city basis, and the architecture flexibly accommodates a broad range of application requirements and deployment scenarios.
This contribution presents (1) an end-to-end architectural overview of the EGENIOUSS multi-sensor navigation framework, detailing the integration of GNSS, IMU, VO and VL; (2) a description of the cloud–edge design that enables scalable deployment, low-latency operation, and privacy-preserving VL; and (3) a unified system perspective that serves as the conceptual foundation for subsequent EGENIOUSS contributions (see list below) addressing individual modules and application-specific evaluations.

ENC 2026 EGENIOUSS contributions:
1. EGENIOUSS: An End-To-End Multi-Sensor Navigation Framework for Complex Urban Environments
2. Visual Localisation as a Building Block for Resilient Multi-Sensor Navigation Using Aerial, Terrestrial and Hybrid Meshes
3. Aiding HAS with inertial and visual measurements for navigation on low-cost devices
4. Smartphone-Based Decimetre-Level Positioning for Urban Navigation and Surveying: Use-Case-Driven Evaluation within EGENIOUSS
5. Pose Verification Based on Visual Localisation Using CityGML Models within the EGENIOUSS Framework

Reliable absolute positioning remains challenging whenever GNSS is degraded, unavailable, or untrustworthy (e.g., urban canyons, indoor–outdoor transitions, or adversarial conditions). Within the EGENIOUSS framework, we present a scalable, mesh-based visual localisation module designed to complement and strengthen multi-sensor navigation pipelines by providing metric 6-DoF pose estimates from commodity images, e.g. smartphone or UAV-acquired images. The module follows a two-stage hierarchical process that explicitly exploits weak priors from GNSS (or other coarse sources) while remaining robust to their errors. First, a coarse localisation stage performs hierarchical retrieval using learned global descriptors to identify candidate regions consistent with the visual appearance and the prior distribution. Second, a fine localisation stage refines the pose through learned feature matching between the query image and selected rendered views to establish robust correspondences. Depth information associated with the rendered views enables direct 2D–3D correspondence formation and subsequent geometric pose estimation through camera resectioning, yielding absolute camera poses that can be delivered as standalone fixes or fused within broader multi-sensor estimators.
We investigate two complementary operating modes of the localisation module. In a pre-rendered mode, reference views are generated offline from the mesh at previously sampled locations, enabling fast online localisation with predictable computational cost. In an on-demand mode, reference views are rendered at run time based on available pose priors, allowing the system to adaptively trade computational effort for accuracy and robustness. This flexibility enables deployment across a wide range of navigation scenarios and resource constraints.
A central contribution is an analysis of how localisation quality and runtime depend on both the operating mode and the provenance of the underlying 3D reference model. We compare three reference map classes: (i) aerial meshes reconstructed from nadir and oblique imagery, (ii) terrestrial meshes derived from high-accuracy mobile mapping, and (iii) “supermeshes” that combine airborne and ground-based data to unify façade and street-level detail with wide-area coverage. We report performance in terms of accuracy, robustness (successful localisation rate), and computational cost, highlighting trade-offs between scalability and precision. These results position mesh-based visual localisation as a practical building block for resilient multi-sensor navigation, enabling map-driven absolute positioning across heterogeneous environments, data sources, and operating modes.

The Galileo open-access High Accuracy Service (HAS) for both Galileo and GPS signals constitutes a major advancement in accurate positioning, surveying, and navigation applications, with a steadily expanding user base and range of operational implementations. With its [validated in practice] specification of nominal [static open-sky] performance of 10 cm and 20 cm 1-sigma level errors in the horizontal and vertical components it can meet the requirements of applications like medium-accuracy surveying, or GIS mapping. Smartphones equipped with dual-frequency receivers are obvious candidates to benefit from HAS.

The HAS nominal accuracy levels, though deteriorate in kinematic positioning and are overshadowed by errors introduced by multipath interference (MP-I) and non-line-of-sight reflections (NLOS-R). For the time being, the HAS corrections to satellite orbits and clocks are not provided for the MP-I resistant E5 AltBOC band or the more advanced GPS L5 band and additional motion measurements from other sensors must be used to obtain acceptable results.

The EGENIOUSS project investigates the use of smartphones’ and low-cost (in mobile platforms) inertial measurement units (IMUs) and cameras – for monocular visual odometry (MVO) and for visual localisation (VL) against existing geographic information. In the considered setup, visual aiding relies on monocular visual odometry and absolute visual localisation against pre-rendered reference views generated from georeferenced aerial mesh data.

EGENIOUSS stands for “EGNSS-based Visual Localisation to enable AAA-PNT in small devices & applications,” and is a Horizon Europe project funded by the EC/EUSPA and coordinated by the Austrian Institute of Technology (AIT) with the participation of German, Spanish, Swiss universities, SMEs, and a large cloud company.

In the article, we will analyse the performance of the two dominant low-cost motion sensors, the IMU and the GNSS receiver, in urban environments, focusing on the impact of GNSS differentiators (new signals: Galileo E5Altboc and L6, GPS L5, and services: Galileo HAS, Galileo Open Service Navigation Message Authentication (OSNMA)). As the preliminary research confirms that the results in nominal conditions are not what can be expected in dynamic mode or urban areas, the paper also examines the performance of aiding other available low-cost sensors, such as cameras, with MVO and VL models.

Drone operations have expanded rapidly in recent years, particularly in urban environments, where applications such as delivery, infrastructure inspection, and façade cleaning demand precise positioning. Accurate and reliable navigation is critical for mission success and operational safety, especially in dense city areas where complex urban orography affects signal reception.

One of the main challenges is GNSS availability, an externa, an external service whose continuous access cannot be guaranteed. Buildings, bridges, and other structures create non-line-of-sight conditions, occlusions, and signal reflections, which degrade GNSS accuracy and reliability. Such limitations can compromise drone localization, posing risks to safety and performance in operations requiring precise manoeuvres.

To address these challenges, EGENIOUSS provides a multi-sensor localization and navigation framework designed for both terrestrial and aerial platforms. The system fuses GNSS (Galileo and GPS) with High Accuracy Service (HAS) corrections, inertial measurement unit (IMU) data, visual odometry (VO) from onboard RGB cameras, and absolute visual localization (VL): object-based localization using CityGML models and feature/mesh-based localization using georeferenced imagery.

Under normal conditions, GNSS, IMU and VO act as primary navigation sensors, with positioning augmentation provided by HAS. In GNSS-degraded environments, visual localization supplements or replaces GNSS inputs. In fully GNSS-denied scenarios, EGENIOUSS can function as an alternative primary navigation system within mapped coverage, maintaining safe UAS operation until GNSS availability returns.

EGENIUOSS is thought from the start to work in a fully GNSS-denied scenario for the UAV use case. An example of the EGENIOUSS functionality integrated into a delivery mission would look as follows:

Mission preparation: For feature/mesh-based VL, the operator defines the intended trajectory, triggering the cloud service to render reference views with feature descriptors along the route. These descriptors are downloaded pre-flight and stored onboard, ensuring functionality without continuous internet access. For object-based VL, the required object data is pre-downloaded. The operator verifies that all necessary data packages are locally available before launch.

In-flight operation: Self-contained navigation using pre-downloaded data when offline, or cloud-assisted rendering of reference views when connectivity is available, with continuous monitoring of fusion integrity and confidence.

Fall-back and contingency: If neither pre-downloaded nor online reference data are available, VO/IMU provide a safe navigation baseline. The operator remains responsible for switching to fall-back procedures, such as GNSS-only return-to-home or contingency landing, if EGENIOUSS confidence falls below operational thresholds.

By combining multi-sensor fusion, pre-downloaded and cloud-assisted visual localization, and robust fall-back strategies, EGENIOUSS ensures precise, reliable, and safe navigation for urban drone operations, supporting complex missions such as delivery, inspection, and façade cleaning even in environments where GNSS alone cannot be trusted.

This contribution presents:

A multi-sensor navigation architecture for UAVs that integrates GNSS(+HAS), IMU, VO, and absolute visual localization using both object-based CityGML models and feature/mesh-based approaches.

A mission-oriented operational concept in which visual localisation relies on pre-rendered reference data generated along a predefined trajectory prior to flight, enabling deterministic and fully GNSS-denied operation.

An integrity-aware navigation strategy with confidence monitoring and explicit fall-back procedures to ensure safe urban UAV operations.