TL;DR
This paper introduces a novel Adaptive Robust Unscented Kalman Filter (AMUKF) that combines a fading factor for state estimation and the Maximum Correntropy Criterion (MCC) for measurement estimation. This approach significantly improves robustness against outliers in target tracking systems, outperforming existing methods like UKF and AHUKF in both simulated scenarios with measurement and state outliers.
What Problem Does This Paper Address?
Accurate state estimation is crucial for many real-world applications, particularly in target tracking. Traditional Kalman Filters (KF) and their nonlinear extensions like the Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF) perform well under ideal conditions. However, their performance degrades significantly when sensor data is contaminated by outliers, which are common in practical scenarios. Outliers can lead to inaccurate state estimates, potentially causing system failure. While existing robust filters exist, they often have limitations, such as not effectively handling both state and measurement outliers simultaneously or being computationally complex. This paper aims to develop a robust and efficient filtering algorithm that can effectively suppress the impact of outliers in both the state and measurement models of nonlinear systems.
Methods
The core of this paper is the development of the Adaptive Robust Unscented Kalman Filter (AMUKF). The AMUKF builds upon the strengths of the Unscented Kalman Filter (UKF), known for its good performance and computational efficiency in nonlinear systems. The novelty lies in the integration of two key components:
- Fading Factor for State Model: A fading factor, denoted as $\alpha_0$, is introduced into the state model. This factor adaptively adjusts the influence of the predicted state based on the innovation sequence. When the state model error is large, the fading factor can be greater than 1, effectively increasing the utilization of new measurements and reducing the impact of dynamics model errors. This helps in strongly tracking the state estimate.
- Maximum Correntropy Criterion (MCC) for Measurement Model: The Maximum Correntropy Criterion is employed as the cost function for the measurement model. MCC is known for its robustness to outliers because it is based on the correntropy, which is a measure of the similarity between two random variables. By maximizing correntropy, the filter becomes less sensitive to large deviations in measurements, effectively suppressing the impact of measurement outliers.
The paper derives the cost function for the AMUKF by combining these two elements and then embeds this into the UKF framework. The algorithm proceeds through standard UKF steps (Unscented Transform, time update, measurement update) but incorporates the fading factor and MCC in the covariance and gain calculations.
Datasets
The paper does not introduce new datasets. Instead, it uses two simulated scenarios for evaluation:
- Radar Tracking System: A standard two-dimensional system tracking a maneuvering vehicle using distance measurements. This scenario includes tests with measurement outliers and a combined case of state and measurement outliers.
- Mars Entry Model: A more complex model simulating a vehicle's entry into Mars' atmosphere, considering atmospheric disturbances and measurement outliers from two orbiters and surface beacons.
Key Insights and Ideas
- Synergistic Robustness: The combination of a fading factor for state estimation and MCC for measurement estimation provides a powerful synergistic effect. The fading factor helps maintain accurate state tracking even with model uncertainties, while MCC robustly handles noisy or outlier measurements.
- Adaptive Nature: The fading factor is adaptive, allowing the filter to dynamically adjust its reliance on past information versus new measurements, which is crucial for tracking maneuvering targets or systems with changing dynamics.
- Unified Outlier Suppression: Unlike some previous methods that might focus on either state or measurement outliers, AMUKF effectively addresses both simultaneously, leading to more reliable estimates in challenging environments.
- UKF as a Strong Foundation: Leveraging the UKF's inherent strengths in handling nonlinearities makes the proposed AMUKF computationally feasible and appropriate for a wide range of practical applications.
Key Results
The paper presents simulation results demonstrating the superiority of the AMUKF over the standard UKF and an Adaptive Huber Unscented Kalman Filter (AHUKF).
- Case 1 (Measurement Outliers): AMUKF showed significantly better tracking performance and smaller position and velocity errors compared to UKF and AHUKF.
- Case 2 (State and Measurement Outliers): AMUKF again outperformed both UKF and AHUKF, demonstrating its robust ability to track targets even when both the process and measurement models are corrupted by outliers.
- Root Mean Square Error (RMSE): Table 2 shows that AMUKF achieved the lowest RMSE for state estimation (x, x, y, y) compared to UKF and AHUKF in the second simulation case.
- Mars Entry Model: Simulations with the Mars entry model also confirmed AMUKF's superior performance in handling measurement outliers, with significantly smaller state errors compared to UKF, and slightly better performance than AHUKF.
Limitations and Future Work
- Single Fading Factor: The paper acknowledges that for complex multivariable systems, a single fading factor might be insufficient. The authors suggest that replacing the scalar fading factor with a matrix of fading factors could be a direction for future research, similar to work by Yang et al. [25].
- Stability Analysis: The authors state that in the future, they will study the stability of the adaptive robust unscented Kalman filter based on the maximum correntropy criterion.
- Theoretical Convergence: While simulations show convergence, a formal theoretical proof of convergence for the AMUKF under various outlier conditions might be a valuable extension.
Summary updated