TL;DR
This paper introduces a novel adaptive filter, the Maximum Correntropy Criterion Kalman/Allan Variance-Assisted FIR (mccKF/FIR) integrated filter, designed to improve the accuracy and robustness of indoor localization systems, particularly for applications like medical procedures. By adaptively switching between a robust mccKF and an Allan Variance-assisted FIR filter based on data quality, the system significantly reduces localization errors compared to traditional Kalman filters.
What Problem Does This Paper Address?
Accurate indoor localization is a critical challenge, especially in environments where GPS is unavailable. Inertial Navigation Systems (INS) are commonly used but suffer from accumulating position errors over time. While integrating INS with other sensors like Ultra-Wideband (UWB) can improve accuracy, existing data fusion methods, including standard Kalman Filters (KF), can struggle with non-Gaussian noise and environmental disturbances. This paper specifically targets the need for highly reliable and precise localization in demanding applications, such as medical interventions, where even small errors can have significant consequences. The authors aim to develop a more robust and accurate integrated localization system that can adapt to varying data quality.
Methods
The core of this work is the proposed mccKF/FIR integrated filter. This approach combines two key components:
- Maximum Correntropy Criterion Kalman Filter (mccKF): This is an enhancement of the standard Kalman Filter that uses the Maximum Correntropy Criterion (MCC) to improve robustness against non-Gaussian noise. MCC is a measure of similarity between probability distributions, and its application here helps the filter better handle outliers or heavy-tailed noise distributions that can degrade standard KF performance.
- Allan Variance-Assisted Finite Impulse Response (FIR) Filter: This component is designed as a fallback mechanism. The FIR filter is known for its robustness and ability to handle data within a defined window. In this proposed system, it's augmented with Allan Variance for improved noise estimation.
The integration works as follows:
- The system uses UWB to correct INS position errors, with the difference between INS and UWB positions serving as the measurement.
- The mccKF acts as the primary filter, fusing data from INS and UWB.
- A Mahalanobis distance is calculated to assess the quality of the mccKF's estimation.
- If the Mahalanobis distance exceeds a predefined threshold (indicating poor data quality or filter performance), the system adaptively switches to the Allan Variance-assisted FIR filter. This adaptive switching ensures resilience in challenging environments.
The paper also details the state and measurement equations for the integrated filter, along with the mathematical formulations for the MCC and Allan Variance methods.
Datasets
The paper describes experiments conducted in a real-world indoor environment at the University of Jinan, China. The testbed involved:
- A crazyflie2.0 quadrotor as the target drone.
- Six DW-1000-based UWB reference nodes with known coordinates.
- One UWB blind node fixed on the target drone to measure UWB positions.
- One IMU fixed on the target drone to provide INS-derived positions.
Two distinct reference paths were used for testing, and the results were compared against the KF, an adaptive KF (AKF), and the mccKF. No specific publicly available datasets were mentioned; the experiments utilized a custom-built testbed.
Key Insights and Ideas
- Adaptive Filtering for Robustness: The most significant insight is the adaptive switching mechanism. By intelligently switching between the MCC-enhanced KF and the Allan Variance-assisted FIR filter, the system can maintain high accuracy and robustness even when data quality fluctuates or when non-Gaussian noise is present. This is crucial for real-world applications where ideal conditions are rare.
- MCC for Handling Non-Gaussian Noise: The paper highlights the effectiveness of the Maximum Correntropy Criterion in making the Kalman filter more resilient to outliers and heavy-tailed noise distributions, which are common in sensor data.
- Allan Variance for Noise Estimation in FIR: Integrating Allan Variance into the FIR filter improves its ability to accurately estimate noise, further enhancing its performance as a fallback or complementary filter.
- Synergy of UWB and INS: The study reinforces the value of fusing UWB and INS data for precise indoor localization, with UWB providing absolute position corrections to the INS's relative measurements.
Key Results
The proposed mccKF/FIR integrated filter demonstrated significant improvements in localization accuracy:
- Test 1: The mccKF/FIR integrated method reduced localization error by approximately 32.43% compared to the standard KF and 37.5% compared to the mccKF.
- Test 2: The proposed method achieved a reduction in localization error of approximately 30% compared to the KF, reducing it from 0.18 m to 0.13 m.
- RMSE Reduction: In Test 1, the mean RMSE was reduced from 0.23 m (KF) to 0.16 m (mccKF/FIR). In Test 2, the mean RMSE was reduced from 0.18 m (KF) to 0.13 m (mccKF/FIR).
- Qualitative Performance: Figures and tables consistently show that the proposed filter's estimated positions are closer to the reference path, exhibiting less divergence and smaller fluctuations, especially during position changes. The adaptive nature of the filter was shown to be particularly effective in minimizing localization errors.
Limitations and Future Work
- Computational Complexity: While not explicitly detailed as a limitation, adaptive filtering and MCC can introduce higher computational overhead compared to a standard KF. The paper doesn't deeply explore real-time performance constraints for highly resource-limited platforms.
- Threshold Selection: The performance of the adaptive switching relies on the selection of the Mahalanobis distance threshold. The paper mentions a threshold of 0.2, but further investigation into optimal threshold selection across various scenarios might be beneficial.
- Medical Application Specifics: While the abstract mentions medical applications, the experimental validation is based on a drone localization testbed. Future work could involve direct validation in medical environments to assess its suitability for tracking instruments or patient positions.
- Generalizability: The study uses a specific INS/UWB setup. Exploring the integration with other sensor modalities (e.g., LiDAR, cameras) or in different indoor environments could broaden its applicability.
Summary updated