TL;DR
This paper introduces a novel fusion approach for indoor robot localization, combining Visual-Inertial Odometry (VIO) with 5G Channel State Information (CSI) and geomagnetic data. By leveraging a deep learning model for 5G/geomagnetic fusion and an Error State Extended Kalman Filter (ES-EKF) for sensor fusion, the method significantly improves localization accuracy and robustness in challenging indoor environments.
What Problem Does This Paper Address?
Accurate and reliable indoor localization for mobile robots is a critical challenge. While Visual-Inertial Odometry (VIO) offers high short-term accuracy, it suffers from cumulative errors, leading to drift over time. Conversely, global positioning techniques like Wi-Fi, 5G, and geomagnetic localization provide drift-free positioning but generally lack the precision required for demanding applications. Existing methods often struggle to compensate for the weaknesses of individual sensors, especially in dynamic or low-light indoor environments. This paper aims to bridge this gap by creating a robust fused localization system that combines the strengths of VIO with the global referencing capabilities of 5G CSI and geomagnetic data.
Methods
The proposed method employs a multi-stage fusion strategy:
- 5G CSI and Geomagnetic Fusion for Global Localization:
- Signal Preprocessing: 5G CSI data (amplitude and phase) is processed, and subcarrier aggregation is performed to reduce data dimensionality while retaining key signal characteristics. Geomagnetic data is rotated to the navigation coordinate system to mitigate pose-induced interference.
- Deep Learning Fusion: An Error Back Propagation Neural Network (BPNN) is trained to fuse the preprocessed 5G CSI and geomagnetic data. This BPNN learns the complex, non-linear mapping between these sensor inputs and global coordinates, providing a more accurate and reliable global position estimate.
- VIO for Local Localization and Coordinate Transformation:
- VIO Estimation: A standard VIO algorithm (VINS-mono is used in experiments) provides high-frequency local pose and position estimates.
- Coordinate Transformation: A least-squares approach is used to determine the transformation parameters (rotation and translation) to align the VIO's local coordinate system with the global coordinate system established by the 5G/geomagnetic fusion.
- ES-EKF for Fused Localization:
- Sensor Fusion: An Error State Extended Kalman Filter (ES-EKF) is employed to fuse the globally referenced VIO position estimates with the global position estimates derived from the 5G/geomagnetic fusion. The ES-EKF effectively integrates these two sources of information, leveraging the VIO's dynamic accuracy and the global sensors' drift-free nature to produce a stable and precise final trajectory.
Datasets
The paper does not explicitly mention pre-existing datasets for training or evaluation. Instead, the authors collected their own data using a custom experimental setup. This involved driving an unmanned vehicle along predefined trajectories in two distinct indoor scenarios: a conference room and a covered corridor. The data collected included images and IMU data from a camera, 3D vector information from a magnetometer, and CSI data from a 5G receiver.
Key Insights and Ideas
- Synergistic Fusion: The core insight is that combining VIO with 5G CSI and geomagnetic data offers a significant advantage over using any single sensor or even pairwise fusions. VIO provides the necessary dynamic accuracy, while 5G CSI and geomagnetism offer crucial global referencing to combat VIO's drift.
- Deep Learning for Robust Global Referencing: Using a BPNN for 5G CSI and geomagnetic fusion is a key innovation. This allows for learning complex, non-linear relationships that are difficult to model analytically, leading to more accurate global position estimates, especially in environments with signal variations.
- Error State Kalman Filtering for Stability: The ES-EKF is crucial for effectively integrating the local VIO estimates with the global 5G/geomagnetic estimates. Its error-state formulation is well-suited for handling the uncertainties and dynamics of sensor fusion, resulting in smoother and more reliable trajectories.
- Impact of Lighting on VIO: The experiments highlight how VIO performance is significantly degraded in low-light conditions, underscoring the need for robust global referencing mechanisms like the one proposed.
Key Results
The proposed fused localization method demonstrated significant improvements in indoor positioning accuracy and robustness:
- 5G/Geomagnetic Fusion: The combined 5G and geomagnetic localization achieved an average error of 0.49 m and a maximum error of 1.47 m, outperforming individual 5G (0.59 m average, 2.76 m max) and geomagnetic (0.67 m average, 3.01 m max) localization.
- Overall Fusion Performance:
- In a conference room scenario, the fused method achieved an average error of 0.61 m, a 69.0% improvement over the VINS-mono algorithm (1.97 m average).
- In a covered corridor scenario, the average error was 0.72 m, a 67.2% improvement over VINS-mono (2.21 m average).
- Trajectory Improvement: The fused trajectories were observed to be smoother and significantly closer to the ground truth compared to VIO alone, especially in challenging lighting conditions or over longer paths where VIO drift becomes pronounced.
Limitations and Future Work
- Limitations:
- The paper acknowledges that while the ES-EKF-based method provides real-time localization, its accuracy is still lower than some optimization-based methods.
- The current work primarily focuses on 2D planar localization.
- The performance of VIO is still affected by lighting conditions, and while the fusion mitigates this, extreme conditions can still impact the overall accuracy.
- Future Work:
- The authors plan to investigate methods that combine the real-time capabilities of filter-based approaches with the accuracy of optimization-based methods.
- They aim to extend the localization to 3D space, leveraging advancements in space technology.
- Future research will also consider using UAVs as platforms to achieve three-dimensional localization of robots.
Summary updated