NR-UIO: NLOS-Robust UWB-Inertial Odometry Based on Interacting Multiple Model and NLOS Factor Estimation

Jieum Hyun, Hyun Myung2021 · DOI

TL;DR

This paper introduces NR-UIO, a novel UWB-IMU sensor fusion algorithm that robustly estimates robot pose in challenging indoor environments. It tackles the issue of non-line-of-sight (NLOS) UWB signal interference by introducing an NLOS factor estimation within an Interacting Multiple Model (IMM) framework, significantly improving localization accuracy.

What Problem Does This Paper Address?

Accurate indoor localization is crucial for robots operating in environments where GPS is unavailable, such as disaster sites. Traditional methods like vision or LiDAR struggle in smoky or feature-poor conditions. Ultra-wideband (UWB) sensors offer a promising alternative due to their ability to penetrate some obstacles and provide precise ranging. However, UWB signals are susceptible to significant errors in non-line-of-sight (NLOS) conditions caused by walls and other obstructions, degrading localization performance. Existing solutions often rely on pre-installed maps or known anchor locations, or they are limited to simulations. This paper aims to develop a robust UWB-IMU fusion system that can accurately estimate robot pose, including orientation, even in the presence of NLOS UWB signals, without requiring prior knowledge of anchor positions.

Methods

The core of the proposed NR-UIO (NLOS-Robust UWB-Inertial Odometry) system is a tightly coupled UWB-IMU sensor fusion approach. It leverages an Interacting Multiple Model (IMM) filter to handle the dynamic nature of robot motion, employing both a constant velocity (CV) and a coordinated turn (CT) model.

Key novelties include:

  • Initial UWB Anchor Position Estimation: The system can estimate the initial positions of UWB anchors without prior knowledge. This is achieved by using the relative positions of multiple UWB tags on the robot and solving a least-squares problem using the Levenberg-Marquardt algorithm.
  • NLOS Factor Estimation: A crucial contribution is the estimation of an NLOS factor for UWB range measurements. This is done in a tightly coupled manner using IMU data and a Unscented Kalman Filter (UKF). The estimated NLOS factor is then used to either compensate for or reject erroneous UWB range data, making the system robust to NLOS conditions.
  • Tightly Coupled UWB-IMU Fusion: Unlike loosely coupled methods, this approach integrates IMU measurements (gyroscope and accelerometer) and UWB range measurements directly into a single state vector. This state vector includes not only the robot's position and orientation but also IMU biases (gyroscope scale factor and bias, accelerometer bias) and the estimated UWB anchor positions.
  • IMU Pre-Integration: To handle the different data rates of UWB and IMU sensors, IMU data is pre-integrated between UWB measurements, providing a more accurate propagation of the robot's state.

Datasets

The experiments were conducted in a real indoor environment at the Safety Robot Experiment Complex in Pohang, South Korea. The paper does not mention the use of pre-existing datasets for training or evaluation. Instead, it describes a custom experimental setup involving a UAV platform equipped with UWB tags, an IMU, and a LiDAR for ground truth measurement.

Key Insights and Ideas

  • The IMM filter is essential for handling complex robot motion: The paper highlights that a single kinematic model is insufficient for accurate state estimation when robots exhibit dynamic movements. The IMM framework, by combining multiple models, significantly improves robustness.
  • Tightly coupled fusion is superior for IMU bias estimation and NLOS handling: Integrating UWB and IMU data directly allows for the simultaneous estimation of IMU biases and the UWB NLOS factor, which is difficult to achieve with loosely coupled methods.
  • NLOS factor estimation is critical for UWB robustness: The ability to identify and compensate for NLOS UWB signals is the key to achieving reliable localization in cluttered indoor environments. The proposed method demonstrates that this can be done in real-time.
  • Initial anchor position estimation removes a significant prior requirement: By enabling the system to estimate anchor positions, the practical deployment of UWB localization is simplified, as it doesn't require precise pre-surveying of anchor locations.

Key Results

  • The proposed NR-UIO algorithm demonstrated superior localization performance compared to a traditional tightly coupled UWB-IMU EKF method. The EKF-based method failed to follow the ground truth trajectory when the UAV entered a corner.
  • Quantitative results show significant improvements in Root Mean Square Errors (RMSE) compared to ground truth. For example, the NR-UIO achieved an RMSE of 0.2133 m in the y-axis, compared to 0.4748 m for the UIO (without NLOS estimation) and 1.3779 m for the EKF.
  • When compared to a state-of-the-art (SOTA) offset compensation method (OC-UIO), NR-UIO also showed better performance, achieving an RMSE of 0.2133 m compared to 0.3086 m for OC-UIO.
  • The paper visually demonstrates the effectiveness of the NLOS factor estimation, showing how the compensated UWB range data closely follows the ground truth, unlike the raw UWB data which exhibits significant deviations in NLOS conditions.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the SOTA offset compensation method has limitations, such as requiring the UWB offset to remain constant for several steps and assuming UWB measurements are not significantly affected by noise.
  • The current implementation of the NLOS factor estimation relies on the assumption that the IMU measurements are relatively accurate between UWB updates.
  • Future Work:
    • The authors suggest exploring more advanced outlier detection methods, such as using Student-t distributions for UWB measurement distributions.
  • They also plan to investigate alternative, more cost-effective solutions for altitude estimation, as the current setup uses a 1D LiDAR costing around $130.
  • Future research will also focus on utilizing other methods, such as machine learning algorithms, for estimating the NLOS bias.

Summary updated

BibTeX
@misc{hyun2021nr,
  author = {Jieum Hyun and Hyun Myung},
  title = {{NR-UIO: NLOS-Robust UWB-Inertial Odometry Based on Interacting Multiple Model and NLOS Factor Estimation}},
  year = {2021},
  doi = {10.3390/s21237886},
}
Figure