Reliable positioning, navigation, and timing (PNT) are critical for autonomous vehicles operating in dynamic and safety-critical environments, where individual sensors may experience degradation, interference, or intermittent outages. Global Navigation Satellite Systems (GNSS), inertial measurement units (IMUs), and vehicle-based sensors provide complementary information; however, their effective integration requires estimation techniques capable of handling strong nonlinearities and time-varying uncertainty. Conventional sensor fusion methods, including Extended Kalman Filter (EKF)–based approaches, often rely on fixed noise assumptions and linearisation, limiting performance under degraded sensing conditions.
This paper introduces an adaptive sensor fusion approach based on the Unscented Kalman Filter (UKF) for resilient autonomous vehicle navigation. The proposed method incorporates innovation-based statistical monitoring to dynamically adjust measurement covariance and sensor weighting in response to changing signal quality, including GNSS degradation and interference. By exploiting the UKF’s sigma-point formulation, the approach enables accurate propagation of uncertainty through nonlinear system and measurement models while maintaining numerical stability during aggressive vehicle manoeuvres and sensor dropouts.
The fusion algorithm integrates GNSS position and velocity measurements with inertial sensor data and vehicle motion constraints to estimate vehicle state in real-time. Performance is evaluated using experimental datasets collected in representative autonomous driving scenarios, including periods of GNSS interference and partial outages. Results demonstrate improved positioning accuracy, faster filter convergence, and enhanced robustness compared with conventional EKF- and fixed-parameter UKF-based fusion methods. The findings highlight the proposed adaptive UKF approach as a practical and scalable solution for resilient PNT in autonomous vehicle applications operating under degraded and contested sensing environments.

Reliable and accurate navigation in urban environments presents a significant challenge for autonomous systems, primarily due to multipath propagation and signal obstruction that degrade Global Navigation Satellite System (GNSS) performance. This paper introduces and evaluates a comprehensive sensor fusion framework designed to mitigate these effects and enhance positioning robustness in dynamic urban scenarios.
The proposed method utilizes an Unscented Kalman Filter (UKF) to integrate three distinct data sources: GNSS pseudorange measurements, acceleration and orientation data from an Inertial Measurement Unit (IMU), and distance data from wheel-odometry sensors. Unlike traditional Extended Kalman Filter implementations, the UKF is employed to better handle the nonlinear system dynamics inherent in ground vehicle navigation.
A central component of this approach is a dynamic Fault Detection and Exclusion (FDE) mechanism. This algorithm continuously monitors the residuals between the predicted state from the UKF and the observed GNSS measurements (so called innovations) within a sliding window. By analyzing the variance of these innovations against a predefined threshold, the system identifies and excludes satellite measurements affected by multipath before they corrupt the position solution.
To validate this methodology, we conducted a measurement campaign using a mobile rover equipped with a multi-frequency GNSS receiver, an IMU, and custom wheel-odometry sensors. The rover traverses a predefined path through a deep urban canyon, which is a challenging environment where satellite visibility is restricted by buildings on both sides. The performance of the proposed algorithm is assessed by comparing the calculated position solutions against a post-processed differential carrier phase reference trajectory.
The purpose of this paper is to evaluate the effectiveness of the Sliding-Window UKF-FDE in real-world dynamic conditions. We aim to analyze whether the integration of odometry and IMU data via a UKF, combined with innovation-based fault detection, provides a stable and accurate navigation solution when GNSS signals are compromised. The study specifically investigates the algorithm's ability to maintain low horizontal and Euclidean error rates and its potential to generalize across different satellite geometries compared to standard GNSS positioning methods.

A navigation system comprising an IMU (Inertial Measurement Unit) driving a strapdown INS (Inertial Navigation System), a GNSS (Global Navigation Satellite System) receiver, and a wheel tick sensor is a rather simple yet powerful setup, found at the heart of many positioning solutions for automotive applications.

On the algorithmic side of fusing information from all these sources, the chosen state space and the filter (a.k.a. observer or estimator) play crucial roles. Vector spaces may provide the easiest and possibly the most intuitive choice for the state space. However, due to the nature of orientation, this will not be the most rigorous and performant choice. EKF (Extended Kalman Filter) may also be the most common choice as the filter, expanding the application of standard Kalman filters as the optimal estimators for linear systems to nonlinear systems in a simple manner. This simplicity can come at a price, though, when facing high degrees of nonlinearity. The combination of vector spaces and EKF, although very common and adequate for many use cases, leaves room for improvement in achieving higher accuracy and consistency given the hardware.

UKF (Unscented Kalman Filter) and DDF (Divided Difference Filter) are two other types of filters, which handle nonlinearities in a fundamentally different manner from EKF. They are often considered to perform better than EKF for highly nonlinear cases, albeit without proof in general. Lots of modifications have been done to all these filters to move from vector spaces to manifolds for state spaces, which improves filter performance in terms of accuracy, consistency, and even convergence.

Invariant filtering takes one step further, and targets a consistent modelling and filtering scheme, where not only the manifold structure of the state space is respected, but also the entirety of the state space and its structure as a group is designed in a way that the filter at the end possesses desired convergence and consistency properties, often out of reach of previously mentioned nonlinear filters.

In this work, we apply the invariant filtering approach to the problem at hand and provide an extensive real-world evaluation, as well as comparison to UKF and DDF on manifolds. The evaluation will be based on a large number of test drives in open sky and urban environments, with centimeter-level independent reference data available. Aspects such as convergence time, tolerance to initialization error, accuracy, and consistency will be examined. To the best of our knowledge, this would be the first effort of such kind, helping to understand the actual benefits of invariant filtering for a common application in real-world scenarios.

In this paper, we evaluate the performance of a tightly coupled GNSS/INS/vision system designed for a small Unmanned Ground Vehicle (UGV) in different urban environments. The UGV is intended to operate on the sidewalks of a major Danish city and automatically detect garbage, potholes and illegally parked bicycles using a 3D lidar. The UGV is travelling at a maximum speed of 6 km/h. The reliability of the positioning solution is obviously critical for such an application, as the UGV must remain on the sidewalk at all times for legal and safety reasons. The designed navigation system is tested in four different areas, where the UGV was manually driven approximately 100 meters back and forth in each area. The four areas represent open-sky, rural, suburban and urban. The sensors on the UGV are a GNSS receiver, a MEMS-based IMU, an upward facing fisheye camera and a 3D lidar. The Tightly Coupled GNSS/INS is utilizing relative carrier-phase measurements to achieve accuracy to the cm level. The urban and deep urban areas pose a significant challenge, as these environments are prone to multipath and Non-line-of-sight (NLOS) reception, which severely degrades the GNSS measurements. The paper will analyze the challenges associated with the four different test areas and evaluate the best parameters for the fusion of GNSS, INS and vision data. There will be a particular emphasis on outlier rejection methods, such as innovation filtering, NLOS detection and measurement weighting using contextual information from the upward camera. We will further also investigate different combinations of GNSS signals and make a user equivalent range error analysis. The 3D lidar on the UGV will be used as an odometry reference to the GNSS/INS/camera solution, as this sensor typically works well in urban environments.

GNSS-based positioning for land applications remains challenging in urban and rail environments due to signal obstruction and multipath effects. Conventional single point positioning (SPP) typically relies on elevation- or carrier-to-noise density ratio (C/N₀)-based weighting schemes, which inadequately model local signal conditions. Recent work has shown that upward-facing fisheye cameras can be used to classify GNSS signals as line-of-sight (LOS), non-line-of-sight (NLOS), or vegetation-obstructed (VEGE), which can be used to develop enhanced weighting schemes.

Building on the CAPLOC2025-3S weighting scheme introduced by Marais et al. (2025) [1], this paper proposes a vision-Inertial Measurement Unit (IMU)- aided sensor fusion framework for improved land-based GNSS positioning. Our approach is based on a tightly coupled Extended Kalman Filter (EKF). The filter uses GNSS pseudoranges and fisheye-camera-derived weights in the measurement update, and the angular rates and accelerations measured by the IMU in the time update or prediction step.

The objective of this paper is to show the benefits of integrating IMU data into the fisheye-camera-based weighted GNSS SPP solution. To do so, the EKF-based sensor fusion framework is evaluated using an urban dataset. A Septentrio Mosaic Receiver with a single patch antenna, an XSens IMU, and a fisheye camera were mounted on top of a vehicle to collect the data. Moreover, a Septentrio AsteRx SBi3 Pro+ dual-antenna receiver with IMU and Real-Time Kinematics (RTK) access was mounted on the vehicle to generate a ground truth.

The paper analyzes the achievable positioning accuracy of the tightly-coupled vision-IMU-aided EKF sensor fusion framework by comparing the solution to the ground truth. The solution is benchmarked against the CAPLOC2025-3S weighted least squares algorithm. Results include a thorough analysis of error time series (horizontal positioning error, cumulative error distribution) for both the EKF-based and the CAPLOC2025-3S WLS solution. The results demonstrate the potential of vision-IMU-aided GNSS sensor fusion to enhance land-based positioning performance in challenging urban and rail scenarios.


[1] J. Marais, T. Guillemaille, C. Meurie and C. Menier, "Fisheye camera-based local effects mitigation for better accuracy in land transportation applications," 2025 IEEE/ION Position, Location and Navigation Symposium (PLANS), Salt Lake City, UT, USA, 2025, pp. 861-869, doi:10.1109/PLANS61210.2025.11028202 .