Integrated WiFi/PDR/Smartphone Using an Unscented Kalman Filter Algorithm for 3D Indoor Localization

Guoliang Chen, Xiaolin Meng, Yunjia Wang, Yanzhe Zhang, Peng Tian, Huachao YangSensors 2015 · DOI

TL;DR

This paper presents a novel 3D indoor localization system that fuses WiFi fingerprinting and Pedestrian Dead Reckoning (PDR) using an Unscented Kalman Filter (UKF). By leveraging smartphone sensors and an improved K-means clustering for WiFi, the system achieves significantly higher accuracy and real-time performance compared to individual methods, enabling a more intuitive 3D visualization of indoor environments.

What Problem Does This Paper Address?

Indoor localization remains a significant challenge, especially for complex environments where traditional GPS is unavailable. Existing WiFi fingerprinting methods often struggle with high computational costs and poor real-time performance on resource-constrained smartphones. Pedestrian Dead Reckoning (PDR), while providing high update rates, suffers from cumulative errors over time. This paper addresses the need for a robust, accurate, and real-time indoor localization system that overcomes the limitations of these individual technologies by effectively fusing them. Furthermore, it tackles the challenge of visualizing this localization data in a 3D space, which is computationally demanding for mobile devices.

Methods

The proposed system integrates two primary localization techniques:

  • WiFi Fingerprinting:
    • Improved K-means Clustering: To reduce the search space and improve retrieval time, an enhanced K-means clustering algorithm is used to partition the WiFi fingerprint database. This significantly speeds up the matching process without sacrificing accuracy.
  • K-Nearest Neighbor (K-NNSS) with Improved Similarity: The traditional K-NNSS algorithm is employed for positioning. A novel similarity calculation is introduced that considers not only the Received Signal Strength (RSS) but also the number of common Access Points (APs) detected, aiming to better reflect spatial relationships.
  • Pedestrian Dead Reckoning (PDR):
    • Auto-Correlation Analysis for Step Counting: A new step counting method based on auto-correlation analysis of accelerometer data is proposed. This method is more robust to variations in phone orientation and pedestrian motion compared to traditional peak detection algorithms.
  • Step Length and Heading Estimation: Standard methods for estimating step length (frequency model) and heading direction (combination of magnetic compass and gyroscope) are used, with map information employed to constrain heading direction.
  • Fusion with Unscented Kalman Filter (UKF):
    • The core of the integration lies in the Unscented Kalman Filter (UKF). The UKF is chosen over the Extended Kalman Filter (EKF) to avoid linearization errors and the need for Jacobian matrices, making it more suitable for the nonlinear system dynamics of indoor localization. The fusion is performed directly in the step domain, simplifying the process and reducing complexity.
  • 3D Visualization:
    • The system utilizes the Unity 3D platform for real-time 3D visualization of the indoor environment and the user's position. This allows for an intuitive and realistic representation of the location data.

Datasets

  • Experimental Data: The experiments were conducted in a specific corridor on the fourth floor of the School of Environment Science and Spatial Informatics at China University of Mining and Technology.
  • WiFi Fingerprint Database: 390 sampling points were collected, with an average sample interval of 3.5 meters. Each point contained RSS values from 216 available APs.
  • Test Points: 15 randomly selected test points were used for evaluation, with each point being located 10 times to assess accuracy.

Key Insights and Ideas

  • Complementarity of WiFi and PDR: The paper highlights the inherent strengths and weaknesses of WiFi fingerprinting (no cumulative error but potentially slow) and PDR (high update rate but cumulative error), demonstrating their strong complementarity for robust indoor localization.
  • Clustering for Efficiency: The application of K-means clustering to the WiFi fingerprint database is a crucial step for improving real-time performance on mobile devices by drastically reducing the search space during the matching phase.
  • Robust Step Counting: The auto-correlation method for step counting offers a significant improvement in reliability by being less susceptible to variations in how a user holds or moves their phone.
  • UKF's Suitability for Nonlinear Systems: The choice of UKF over EKF is justified by its ability to handle nonlinearities inherent in sensor fusion and motion modeling without the computational burden of Jacobian calculations.
  • 3D Visualization as a User Experience Enhancer: The integration of Unity 3D for 3D visualization transforms raw localization data into an easily understandable and engaging user experience, which is often a bottleneck for mobile applications.

Key Results

  • WiFi Fingerprint Clustering:
    • Positioning time was reduced by an average of 51%, with a maximum reduction of 64%.
  • Average positioning error decreased from 1.92 m to 1.12 m.
  • Maximum positioning error was reduced from 4.09 m to 0.59 m.
  • PDR Auto-Correlation:
    • The auto-correlation algorithm showed better performance in counting steps across various mobile phone positions and pedestrian motions compared to peak detection. For instance, in "In Hand, Changing frequency of walking," auto-correlation yielded 196 steps versus 224 for peak detection, with the true steps being 200.
  • UKF Fusion:
    • The fused system significantly reduced cumulative errors. For example, at corner C, PDR-only error reached 6.8 m, while the fused system error was under 1.7 m.
  • At the end of a 133.5 m path, PDR positioning error was 6.0 m, while the fused system error was only 1.5 m.
  • Figure 14 visually demonstrates that the UKF fusion method achieved more accurate location results compared to PDR or WiFi positioning alone.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the WiFi fingerprinting approach can still have large errors in some cases, especially when the initial filter value is based on a large WiFi positioning error.
  • The step length estimation is noted as being difficult to set precisely due to individual differences in height and weight, relying on coefficients from a previous study.
  • The accuracy of heading direction estimation can be affected by sensor errors and non-aligned phone orientation.
  • Future Work:
    • The authors suggest that the UKF filter algorithm effectively fuses WiFi fingerprint positioning and PDR positioning, and that correcting WiFi fingerprinting can fix the cumulative error of PDR. This implies further refinement of the fusion process and error correction mechanisms.
  • The development of a 3D indoor positioning system based on Unity 3D is presented as a way to express complicated indoor space geographic information and enrich user experience, suggesting potential for further enhancements in visualization and user interaction.

Summary updated

BibTeX
@article{chen2015integrated,
  author = {Guoliang Chen and Xiaolin Meng and Yunjia Wang and Yanzhe Zhang and Peng Tian and Huachao Yang},
  title = {{Integrated WiFi/PDR/Smartphone Using an Unscented Kalman Filter Algorithm for 3D Indoor Localization}},
  journal = {Sensors},
  year = {2015},
  doi = {10.3390/s150924595},
}
Figure