A State Optimization Model Based on Kalman Filtering and Robust Estimation Theory for Fusion of Multi-Source Information in Highly Non-linear Systems

Muhammad Adeel Akram, Peilin Liu, Muhammad Owais Tahir, Waqas Ali, Yuze WangSensors 2019 · OpenReview

TL;DR

This paper introduces a novel Robust Reweighted Iterative Kalman Filter (RRIEKF) designed to improve state estimation accuracy in highly non-linear systems, especially when dealing with noisy and outlier-prone multi-source data. By combining robust estimation techniques with an iterative Kalman filtering approach, the RRIEKF demonstrates superior performance in tracking dynamic states and integrating diverse sensor inputs compared to existing methods.

What Problem Does This Paper Address?

Accurate state estimation is crucial for many real-world applications, from target tracking to autonomous navigation. Traditional Kalman filters and their variants (like EKF and IEKF) excel in linear systems with Gaussian noise but struggle significantly when faced with non-linear dynamics or unreliable sensor data. Measurement errors, system model deviations, and especially outliers can lead to filter divergence and inaccurate state estimates. While existing robust methods exist, they often have limitations in handling complex non-linearities, multirate sensor integration, or require prior knowledge about noise characteristics. This paper aims to bridge this gap by proposing a robust, iterative filtering approach that can effectively handle both system non-linearities and measurement outliers without requiring extensive prior information.

Methods

The core of this work is the Robust Reweighted Iterative Kalman Filter (RRIEKF). It builds upon the Iterative Extended Kalman Filter (IEKF) by incorporating robust estimation principles. The RRIEKF introduces several key components:

  • Robust State Prediction Model: This model uses a batch-mode approach to maintain data redundancy and suppress outliers during the prediction phase. It also incorporates an adaptive online scheme for handling process error covariance.
  • Robust Covariance Estimation: This recursive process refines the error covariance estimation by considering both current and prior observations, making it more resilient to noise and model errors.
  • Subsampling Optimization: For systems with abundant data, a subsampling process is employed to optimize candidate solutions iteratively, increasing the redundancy of error-free estimates and selecting the best estimate based on predefined conditions.
  • Weighting Scheme: A crucial element is the iterative reweighting of observations. Outlying or noisy measurements are down-weighted, reducing their influence on the state estimate. This is achieved using robust statistics, such as Tukey's biweight function, to adaptively assign weights.

The RRIEKF iteratively refines the state estimate by applying these robust techniques within a Kalman filtering framework, aiming to achieve a more accurate and stable estimation, especially in challenging environments.

Datasets

The paper evaluates the proposed RRIEKF using:

  • Numerical Simulations: Two scenarios are simulated:
  • A simple accelerative and turning maneuver with additive Gaussian white noise (1 dBW and 5 dBW).
  • A non-linear multivariate system with added Gaussian distributed noise at various power levels (0, 5, 10, and 20 dBW).
  • Complex scenarios with introduced additional noise and outliers are also used.
  • Real Sensor Data: Data from Inertial Measurement Units (IMU) and Ultra-Wideband (UWB) sensors are integrated. These datasets represent real-world scenarios with noisy, multirate sensor measurements, including cases with calibrated sensors, sensors with added noise, and sensors with outliers.

Key Insights and Ideas

  • Robustness is Key for Non-linear Systems: The paper strongly emphasizes that for highly non-linear systems, robustness against outliers and model uncertainties is as critical as accuracy. Standard filters like EKF and IEKF, while good, are shown to be brittle in the face of significant noise or outliers.
  • Iterative Reweighting is Powerful: The core idea of adaptively down-weighting unreliable observations in an iterative manner is a highly effective strategy for improving estimation performance in noisy environments. This allows the filter to "learn" which measurements to trust more.
  • Batch Processing for Prediction: Using batch-mode processing for state prediction enhances data redundancy, which is crucial for suppressing outliers and improving the reliability of the initial state guess.
  • Multirate Sensor Fusion is Challenging: Integrating sensors with different sampling rates (like IMU at 100 Hz and UWB at 25 Hz) presents unique challenges. The proposed RRIEKF is shown to be effective in handling such multirate fusion, particularly when dealing with noisy sensor inputs.

Key Results

  • Improved Accuracy: Numerical simulations and real-world data analysis consistently show that the RRIEKF achieves higher accuracy and lower Root Mean Square Errors (RMSE) compared to EKF, UKF, and IEKF, especially under conditions of high noise and outliers. For instance, in one simulation with 20 dBW noise, the RRIEKF achieved an RMSE of 3.3963, significantly lower than EKF (30.0964) and IEKF (13.7304).
  • Enhanced Robustness: The RRIEKF demonstrates superior robustness, maintaining stable and accurate estimates even when faced with significant noise (up to 20 dBW) and observation outliers, where other filters degrade considerably.
  • Effective Multirate Fusion: The algorithm successfully integrates IMU and UWB sensor data, providing accurate position and speed estimations even when the individual sensor data is noisy.
  • Visual Evidence: Figures 2, 4, 5, 6, 7, and 8 visually confirm the RRIEKF's superior performance, showing smoother trajectories, smaller estimation errors, and better adherence to real paths compared to other filters.

Limitations and Future Work

  • Computational Cost: While the RRIEKF offers improved performance, iterative processes and robust estimation techniques can increase computational complexity. The paper mentions subsampling as a way to manage this for systems with many observations.
  • Future Work: The authors plan to extend the proposed algorithm for localization by integrating Global Navigation Satellite System (GNSS), IMU, and visual sensor measurements. They also aim to conduct a comparative analysis with other state-of-the-art estimation techniques in these complex integrated systems.

Summary updated

BibTeX
@inproceedings{akram2019state,
  author = {Muhammad Adeel Akram and Peilin Liu and Muhammad Owais Tahir and Waqas Ali and Yuze Wang},
  title = {{A State Optimization Model Based on Kalman Filtering and Robust Estimation Theory for Fusion of Multi-Source Information in Highly Non-linear Systems}},
  booktitle = {Sensors},
  year = {2019},
  url = {https://openreview.net/forum?id=cFxL0_5mnff},
}
Figure