TL;DR
This paper introduces a novel Kalman filter framework that unifies handling of non-linear systems, outliers, and non-uniform sampling, while also enabling fixed-lag smoothing and automated parameter estimation. This integrated approach allows for more robust and accurate state estimation in challenging real-world scenarios where traditional Kalman filters struggle.
What Problem Does This Paper Address?
The Kalman filter (KF) is a powerful tool for signal processing, but its effectiveness is often limited by strict assumptions: linearity of the system and Gaussian noise. When these assumptions are violated, performance degrades. While extensions exist for non-linear systems (e.g., EKF, UKF), outlier handling, and non-uniform sampling, these are typically addressed independently. This separation means that the full benefits of these extensions cannot be exploited simultaneously. For instance, dealing with non-uniform sampling often forces a trade-off between smoothing accuracy and real-time processing. Furthermore, accurately estimating the KF's crucial parameters (like noise covariance matrices) is difficult, especially in the presence of outliers and without reference measurements. This paper aims to bridge this gap by presenting a unified framework that integrates these features, enabling more robust and versatile state estimation.
Methods
The core of the paper is a novel Kalman filter framework built upon the Extended Kalman Filter (EKF) for handling non-linearities. The key innovations and integrations are:
- Outlier Handling: A threshold-based outlier detection mechanism is introduced, evaluating the Mahalanobis distance of the innovation. Measurements exceeding a predefined threshold are marked as invalid and excluded from the filtering process.
- Non-Uniform Sampling and Resampling: The EKF's prediction capabilities are leveraged to generate state estimates at desired output times, even when observations are irregularly spaced. This allows for a constant output rate, mitigating issues with non-uniform input. Intermediate prediction steps are used to minimize linearization errors.
- Fixed-Lag Smoothing: The framework incorporates fixed-lag smoothing to improve accuracy by utilizing past information, while maintaining a constant delay. This is achieved by modifying the standard Rauch-Tung-Striebel smoother to handle variable sampling rates and potential missing (outlier-rejected) measurements.
- Automated Parameter Estimation: A crucial contribution is the automated estimation of process noise covariance (Q), measurement noise covariance (R), and the outlier detection threshold (α). This is formulated as an optimization problem, minimizing a modified log-likelihood function. A genetic algorithm (GA) is employed for global optimization, avoiding local minima often encountered with gradient-based methods. A penalty term is introduced into the objective function to penalize the rejection of valid measurements while encouraging the rejection of outliers.
Datasets
The paper utilizes two types of datasets for evaluation:
- Simulated Data: Two datasets (D1 and D2) were generated to mimic object movement in a 2D plane with constant turning rate and acceleration. These datasets were designed with different process noise parameters and included artificially introduced outliers (5% of data) and non-uniform sampling.
- Real Data: 2D position measurements from a local positioning system (LPS) observing a quad bike on a race track were used. This data was non-uniformly sampled and corrupted by outliers. Ground truth was provided by a Real-Time Kinematic Global Positioning System (RTK-GPS).
Key Insights and Ideas
- Unification is Key: The most significant insight is that integrating outlier handling, non-uniform sampling, and smoothing into a single Kalman framework unlocks synergistic benefits that are lost when these are treated independently.
- Outlier Penalty for Robust Estimation: The introduction of a penalty term in the parameter estimation objective function is a clever way to simultaneously estimate noise parameters and the outlier threshold. This penalty discourages discarding valid data while still allowing for outlier rejection.
- GA for Robust Parameter Estimation: Using a genetic algorithm for parameter estimation is crucial for finding a global optimum, especially when dealing with complex objective functions influenced by outliers. This is a significant improvement over methods that might converge to local minima.
- Smoothing Benefits Even with Outliers: The results demonstrate that fixed-lag smoothing can significantly improve accuracy, even when dealing with data that contains outliers, provided the framework correctly identifies and handles them.
- Outlier Rejection as Implicit Noise Cutting: In some cases, the rejection of outliers and the filtering of noisy data can lead to results that are as good as, or even better than, filtering data that was initially clean but still subject to Gaussian noise. This suggests that robust filtering can effectively "cut off the tails" of the underlying noise distribution.
Key Results
- Parameter Estimation Accuracy: Method A (with outlier consideration in parameter estimation) achieved significantly more accurate estimates of noise parameters compared to Method B (without outlier consideration), which tended to overestimate parameters by up to two magnitudes.
- Improved Filtering Accuracy: Filtering with the proposed method (A) resulted in substantially lower Root Mean Square Errors (RMSE) compared to Method B, especially for real data. For real data, Method A achieved an RMSE of 0.13 m (real-time) and 0.07 m (delayed filtering), while Method B achieved 0.40 m (real-time) and 0.24 m (delayed filtering).
- Robust Outlier Detection: The outlier detection method demonstrated high sensitivity (over 94%) and specificity (over 96%) on both simulated and real datasets, indicating its robustness.
- Fixed-Lag Smoothing Benefits: Fixed-lag smoothing further reduced RMSE for both methods, particularly for Method A, highlighting its effectiveness in improving accuracy with a constant delay.
- Real-World Applicability: The framework showed robust operation on real-world data, with the filtered trajectory closely matching the reference data when outliers were handled, and deviating significantly when they were not.
Limitations and Future Work
- Computational Cost of GA: The paper acknowledges that the global optimization approach using a genetic algorithm can be computationally expensive, especially for high-dimensional systems.
- EM Algorithm Robustification: Future work could explore robustifying the Expectation-Maximization (EM) algorithm for parameter estimation to potentially reduce computational costs.
- Transfer to UKF/SRUKF: The authors suggest that transferring the presented combination of modifications to the Unscented Kalman Filter (UKF) and Square-Root Unscented Kalman Filter (SRUKF) would be a valuable area for future investigation.
- Further Robustification Methods: The paper notes that the proposed filtering procedure can be combined with other robustification methods, suggesting potential for further improvements.
- Parameter Estimation Speed-up: Investigating ways to speed up parameter estimation, particularly for high-dimensional systems, is identified as a key area for future research to enhance the method's viability.
Summary updated