Integrated WiFi/PDR/Smartphone Using an Adaptive System Noise Extended Kalman Filter Algorithm for Indoor Localization

Xin Li, Jian Wang, Chunyan Liu, Liwen Zhang, Zheng-Kui LinISPRS Int. J. Geo Inf 2016 · DOI

TL;DR

This paper tackles the challenge of accurate and stable indoor localization by fusing Wi-Fi fingerprinting with Pedestrian Dead Reckoning (PDR). They introduce an adaptive Extended Kalman Filter (EKF) that dynamically adjusts its noise parameters based on pedestrian movement, significantly improving accuracy and reducing common issues like signal instability and PDR error accumulation.

What Problem Does This Paper Address?

Indoor navigation is crucial for many applications, but unlike outdoor GPS, it faces significant challenges. Wi-Fi fingerprinting, a common indoor localization technique, suffers from signal interference, instability, and "jumping" in position estimates. Pedestrian Dead Reckoning (PDR), which uses inertial sensors, offers high short-term accuracy but accumulates errors over time. Existing hybrid approaches often struggle with computational burden or fail to adapt to dynamic environmental conditions. This paper aims to create a robust and accurate indoor localization system by effectively fusing Wi-Fi and PDR, overcoming their individual limitations.

Methods

The proposed solution involves several key components:

  • Wi-Fi Fingerprint Clustering with Affinity Propagation (APC): To reduce the computational load of Wi-Fi fingerprint matching, the authors employ APC. This algorithm clusters reference points (RPs) in the fingerprint database, considering both signal strength and spatial proximity. APC is chosen over K-means for its better adherence to spatial structures and more uniform class distribution.
  • Feature Fusion: A novel fusion feature, MixDis, is introduced to combine signal distance and geographic distance between RPs. This helps to better represent the spatial relationship and mitigate noise from signal instability.
  • Adaptive Noise Extended Kalman Filter (EKF): The core of the fusion is an EKF that adaptively adjusts its system noise parameters. This adaptation is based on the pedestrian's movement status (straight walking vs. turning). By dynamically changing the noise covariance matrix (Qk), the filter can better handle the varying dynamics of pedestrian motion, especially at corners, thus improving stability and accuracy.
  • Wi-Fi and PDR Integration: The system uses Wi-Fi for absolute positioning updates and PDR for continuous, high-rate position tracking. The EKF fuses these two sources, using Wi-Fi updates to correct PDR drift and PDR to bridge gaps between Wi-Fi updates.

Datasets

The experiments were conducted in a real-world indoor environment: a fourth-floor corridor at the School of Environment and Spatial Informatics, China Mining University of Technology.

  • Wi-Fi Data: D-LINK wireless routers (DIR-600NB) were installed, providing 64 Access Points (APs). Wi-Fi signal strength data was collected over five days to analyze time-variation effects. Fingerprint data was collected over 10 consecutive days, with 10 samples per RP in four directions each day, resulting in 40 groups of signal strength per RP.
  • PDR Data: An inertial sensor (likely part of the Samsung Galaxy Note 3 smartphone) sampled at 50 Hz, providing data for gait detection, step-size calculation, course estimation, and ultimately, PDR.
  • Benchmark Data: The paper compares its proposed method (AWEPDR) against several baseline methods: Wi-Fi only, PDR only, and a conventional EKF fusion (WEPDR).

Key Insights and Ideas

  • Dynamic Noise Adaptation is Key: The most significant insight is that simply fusing Wi-Fi and PDR isn't enough. The system's performance, especially at corners, is greatly improved by dynamically adjusting the filter's noise parameters based on how the pedestrian is moving. This allows the filter to be more responsive to turns without overreacting to minor movements.
  • APC for Efficient Fingerprinting: Using Affinity Propagation Clustering for fingerprint data significantly reduces the search space for matching, leading to substantial improvements in time efficiency without sacrificing accuracy.
  • Fusion of Spatial and Signal Features: Combining both geographic distance and signal strength differences in the clustering feature (MixDis) provides a more robust representation of RP relationships, leading to better clustering results.
  • Wi-Fi Instability is a Major Hurdle: The experimental data clearly shows the inherent instability of Wi-Fi signals, with significant fluctuations and "rebound" phenomena in positioning results. This underscores the need for robust fusion techniques.

Key Results

  • Accuracy Improvement: The proposed adaptive noise EKF fusion (AWEPDR) achieved an average positioning error of 2.32 m, a significant improvement over Wi-Fi alone (4.09 m).
  • Time Efficiency: APC clustering reduced the number of points to be searched by 65%-80%, greatly improving time efficiency. The average positioning time for APC was 70 ms, comparable to K-means (76 ms) but more stable.
  • Reduced Error Fluctuation: The AWEPDR method showed a smaller error fluctuation compared to other methods.
  • Corner Performance: The adaptive noise adjustment effectively weakened divergence at corners, a common problem in other fusion methods.
  • Quantitative Error Reduction: The average positioning error of AWEPDR was 56.2%, 37.5%, and 85.9% smaller than Wi-Fi, PDR, and WEPDR respectively.

Limitations and Future Work

  • Environmental Dependence: The performance of Wi-Fi fingerprinting is highly dependent on the specific environment and the density of APs. The proposed method relies on a well-established fingerprint database.
  • PDR Error Accumulation: While the fusion mitigates PDR error, it's not entirely eliminated, especially over very long distances without Wi-Fi updates.
  • Computational Cost: While APC reduces the matching burden, the overall fusion process still requires significant computational resources, especially for real-time processing on mobile devices.
  • Future Work: The paper doesn't explicitly detail future work, but potential directions could include exploring more advanced clustering techniques, investigating adaptive filtering for different sensor types, and optimizing the system for lower-power consumption on mobile platforms. Further validation in more complex indoor environments with varying layouts and higher AP densities would also be beneficial.

Summary updated

BibTeX
@article{li2016integrated,
  author = {Xin Li and Jian Wang and Chunyan Liu and Liwen Zhang and Zheng-Kui Lin},
  title = {{Integrated WiFi/PDR/Smartphone Using an Adaptive System Noise Extended Kalman Filter Algorithm for Indoor Localization}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2016},
  doi = {10.3390/ijgi5020008},
}
Figure