TL;DR
This paper introduces an "outlier-insensitive" Kalman filter (OIKF) that robustly handles noisy sensor data by modeling outliers as random variables with unknown variances (NUV). The OIKF effectively balances the influence of potential outliers, preserving Kalman filter optimality for clean data while significantly improving accuracy in the presence of outliers.
What Problem Does This Paper Address?
The Kalman filter (KF) is a cornerstone algorithm for tracking the state of dynamical systems using noisy observations. It's highly effective for linear Gaussian systems, minimizing mean-squared error (MSE). However, real-world sensor data often contains "outliers" – erroneous measurements that deviate significantly from the norm. These outliers can severely degrade the KF's performance, leading to inaccurate state estimates. Existing methods for handling outliers often require careful tuning, can discard valuable information, or are computationally expensive. This paper aims to develop a robust Kalman filter that can effectively handle outliers without sacrificing performance on clean data and with reduced computational complexity.
Methods
The core innovation is the Outlier-Insensitive Kalman Filter (OIKF). Instead of simply discarding outliers, the OIKF models each potential outlier as a normally distributed random variable with an unknown variance (NUV). This NUV component is added to the standard observation noise. The key is to estimate these unknown outlier variances online.
The paper proposes two methods for estimating these NUV variances:
- Expectation-Maximization (EM): This algorithm, adapted from previous work on smoothing with outliers, estimates both the first and second-order moments of the state.
- Alternating Maximization (AM): This is a novel approach for the filtering task. It's simpler than EM, relying only on first-order moments (specifically, the state estimate from the KF). This makes it less computationally demanding and, as shown empirically, more robust.
Both EM and AM aim to find a Maximum A Posteriori (MAP) estimate for the outlier variance. A crucial property of this approach is its "sparsifying" nature: when an outlier is not strongly detected, its estimated variance is effectively zero, causing the OIKF to revert to the standard KF. When an outlier is detected, its estimated variance is incorporated into the observation noise covariance, balancing its contribution to the state update.
Datasets
- Synthetic Dataset: A dynamic WNA (White Noise Acceleration) model was used to generate synthetic data. This allowed for controlled experiments with varying levels of outlier intensity and noise.
- Real-World Dataset: The Michigan NCLT (North Campus Long-Term) dataset was used. This dataset contains noisy GPS readings from a Segway robot, providing a realistic scenario for evaluating the filter's performance in a practical application.
Key Insights and Ideas
- Modeling Outliers as NUV: The idea of treating outliers not as errors to be discarded but as a probabilistic component with unknown variance is a powerful way to integrate outlier information into the estimation process.
- Sparsity for Robustness: The MAP estimation of outlier variances naturally leads to a sparse solution, meaning most of the time, outliers are effectively ignored, and the filter behaves like a standard KF. This preserves optimality for clean data.
- AM's Simplicity and Robustness: The Alternating Maximization (AM) approach is a significant contribution. By avoiding the computation of second-order moments (state covariance), it achieves comparable or better performance than EM with reduced complexity and improved robustness.
- Reverting to Classic KF: A key desirable property is that when data is clean, the OIKF seamlessly reverts to the optimal Kalman filter, ensuring no performance degradation.
Key Results
- Improved Accuracy: Both OIKF-EM and OIKF-AM significantly outperform existing robust Kalman filter variants (like WRKF and ORKF) and the standard KF in terms of Mean Squared Error (MSE) and Root Mean Squared Error (RMSE), especially in the presence of outliers.
- OIKF-AM Efficiency: OIKF-AM demonstrated comparable performance to OIKF-EM but with a substantial reduction in runtime (around 40% less).
- Optimal Performance on Clean Data: For data free of outliers, the OIKF-AM achieves optimal MSE, matching the performance of the standard KF.
- Robustness to Noise Levels: The OIKF-AM showed superior performance across various levels of observation noise variance.
Limitations and Future Work
- Computational Complexity: While OIKF-AM is more efficient than OIKF-EM, it still involves iterative estimation within the KF update step. For extremely high-dimensional or real-time critical applications, further optimization might be needed.
- Prior Selection: The paper uses a uniform prior for the NUV variance, which is parameter-free. However, the authors suggest that investigating alternative priors could lead to different effective loss functions (e.g., Huber cost) and potentially further enhance performance.
- Data-Driven Fusion: The authors express interest in exploring the fusion of OIKF with trainable, data-driven variants of the Kalman filter (e.g., neural network-aided filters) to handle partially known state-space models more effectively.
Summary updated