TL;DR
This paper introduces a novel adaptive Kalman filter for INS/GNSS navigation systems that tackles performance degradation caused by measurement anomalies and non-Gaussian noise. By combining an adaptive unscented Kalman filter (AUKF) with a maximum versoria criterion (MVC) cost function, the proposed method significantly improves navigation accuracy and robustness in challenging environments.
What Problem Does This Paper Address?
Accurate and reliable navigation is crucial for both military and civilian applications, from precision-guided weapons to autonomous vehicles. Inertial Navigation Systems (INS) offer autonomy but suffer from accumulating errors. Global Navigation Satellite Systems (GNSS) provide high accuracy but are susceptible to signal degradation and multipath effects, especially in urban canyons or tunnels. Integrating INS and GNSS (INS/GNSS) is a common solution, but this fusion is challenged by non-Gaussian noise and occasional measurement anomalies that can arise from sensor imperfections and environmental interference. Traditional Kalman filtering methods, which often assume Gaussian noise, struggle to handle these issues, leading to degraded navigation performance and potential filter divergence.
Methods
The core of this work is the development of an Adaptive Unscented Kalman Filter (AUKF) based on the Maximum Versoria Criterion (MVC), termed MVCAUKF.
- Maximum Versoria Criterion (MVC): This criterion is introduced as a novel cost function for optimization. Unlike traditional methods that primarily use second-order moments (like the minimum mean-square error criterion), the MVC considers higher-order moments of estimation errors. This makes it more suitable for handling non-Gaussian measurement noise and outliers. The versoria function, used in the MVC, has a thick-tailed characteristic and avoids computationally expensive exponential operations, offering robustness and efficiency.
- Adaptive Unscented Kalman Filter (AUKF): Building upon the Unscented Kalman Filter (UKF), which is adept at handling nonlinear systems, the MVCAUKF incorporates an adaptive mechanism. This allows the filter to dynamically adjust its parameters in real-time to cope with system process uncertainty and measurement anomalies, thereby enhancing robustness.
The MVCAUKF algorithm consists of two main parts: state prediction and measurement update, following the standard Kalman filter structure but incorporating the MVC cost function and adaptive adjustments.
Datasets
The paper evaluates the proposed MVCAUKF method using data collected from a vehicular navigation experimental platform. This platform included:
- A MEMS-INS/GPS-integrated navigation system.
- A NovAtel SPAN-LCI high-precision integrated navigation system as a reference.
The experiments were conducted on a 700m trajectory over 120 seconds, involving various maneuvers (sharp turns, lane changes, acceleration/deceleration) and challenging road conditions (flat, bumpy, uphill, downhill) to simulate complex navigation environments with potential GNSS signal degradation and non-Gaussian noise. The authors performed 12 repeated experiments to mitigate single-run uncertainties.
Key Insights and Ideas
- Beyond Gaussian Assumptions: The paper highlights that real-world navigation systems, especially in complex environments, frequently encounter non-Gaussian noise and measurement anomalies. This underscores the limitations of traditional filters that rely on Gaussian assumptions.
- MVC's Robustness: The MVC, by considering higher-order error moments and utilizing the versoria function, offers a more robust approach to filtering compared to standard quadratic criteria. Its "thick-tailed" characteristic is particularly effective against outliers.
- Adaptability is Key: The integration of an adaptive mechanism within the AUKF allows the filter to dynamically adjust to changing noise characteristics and system uncertainties, a crucial feature for maintaining performance in dynamic and unpredictable environments.
- Efficiency of Versoria: The absence of exponential operations in the versoria function contributes to a lower computational load, making the MVC approach more practical for real-time applications, especially on resource-constrained platforms.
Key Results
The MVCAUKF demonstrated superior performance compared to traditional methods (EKF, UKF) and a contemporary robust filter (MCCUKF) in complex navigation scenarios.
- Accuracy Improvement:
- Reduced RMSE of east velocity by up to 70.96% (vs. EKF), 57.14% (vs. UKF), and 43.75% (vs. MCCUKF).
- Reduced RMSE of north velocity by up to 65.52% (vs. EKF), 54.54% (vs. UKF), and 33.33% (vs. MCCUKF).
- Reduced east position error by up to 71.82% (vs. EKF), 60.94% (vs. UKF), and 48.88% (vs. MCCUKF).
- Reduced north position error by up to 65.87% (vs. EKF), 52.36% (vs. UKF), and 38.04% (vs. MCCUKF).
- Robustness: The MVCAUKF maintained good estimation performance and remained robust even with uncertain process noise and in the presence of measurement anomalies.
- Computational Cost: While exhibiting improved accuracy and robustness, the MVCAUKF achieved a shorter computation time compared to the MCCUKF, attributed to the conciseness of the versoria function. The total computation time for the MVCAUKF was 23.27 seconds, compared to 28.23 seconds for MCCUKF over the 120-second experiment.
Limitations and Future Work
- Limitations: The paper acknowledges that while the MVCAUKF shows improved accuracy and robustness, it comes at the cost of increased computation compared to simpler filters like EKF and UKF. The performance of the MVC-based filter is also sensitive to parameter tuning (e.g., the versoria radius 'r').
- Future Work: The authors suggest future research should focus on more advanced noise characterization methods to explicitly model the non-Gaussian and time-varying properties of GNSS measurement noise. This would further refine the adaptive mechanisms of the proposed algorithm and potentially lead to even greater improvements in complex environments.
Summary updated