Continuous Space Estimation: Increasing WiFi-Based Indoor Localization Resolution without Increasing the Site-Survey Effort

Noelia Hernández, Manuel Ocaña, José Maria Alonso, Euntai KimSensors 2017 · DOI

TL;DR

This paper introduces a novel approach to improve the resolution of WiFi-based indoor localization systems without requiring more effort in surveying the environment. By using Support Vector Regression (SVR) to estimate Wi-Fi signal strengths across continuous spaces, the system can pinpoint locations more accurately, even in areas not explicitly mapped during the initial setup. This is crucial for developing more practical and user-friendly indoor location-based services.

What Problem Does This Paper Address?

Indoor localization, the process of determining a device's position within a building, is essential for a wide range of applications, from navigation in public spaces to enhanced healthcare and security. While GPS works well outdoors, it's unreliable indoors due to signal obstruction. WiFi-based localization is a popular alternative because WiFi is ubiquitous and most devices have built-in WiFi capabilities.

A common and effective WiFi localization technique is "fingerprinting," which involves creating a map of WiFi signal strengths (Received Signal Strength - RSS) at various known locations. During localization, the device's current RSS is compared to this map to estimate its position. However, the accuracy of fingerprinting is directly tied to the density of the "site survey" – the process of collecting RSS data. A higher resolution (more precise location) requires a much more extensive and time-consuming site survey, which is a significant barrier to practical deployment. This paper aims to bridge this gap by increasing localization resolution without increasing the site survey effort.

Methods

The core of this paper's contribution is the Continuous Space Estimator (CSE). The CSE operates in two main stages:

  • Training Stage:
    • Discrete Information Collection: Similar to traditional fingerprinting, RSS is measured at a set of discrete, known "topological positions" within the environment.
  • Reference Surface Creation: For each WiFi Access Point (AP), the system uses Support Vector Regression (SVR) to create a continuous "reference surface." This surface estimates the expected RSS from that AP at any coordinate within the environment, not just the surveyed points. This is achieved by interpolating the discrete measurements. The paper specifically uses an epsilon-insensitive SVR (e-SVR) to handle the inherent noise and variations in WiFi signals.
  • Environment Masking: An optional mask can be applied to exclude areas that are physically inaccessible or irrelevant for localization.
  • Localization Stage:
    • RSS Measurement: The device measures the RSS from all detected APs.
  • Surface Search: The measured RSS values are then searched within the corresponding continuous reference surfaces created during the training stage. A "noise margin" is applied to account for signal fluctuations.
  • Subsurface Generation: For each AP, a "localization subsurface" is generated, indicating the probability of the device being at each coordinate based on how well its measured RSS matches the estimated RSS on the reference surface.
  • Summation and Estimation: These subsurfaces are summed up to create a final probability map. The coordinate with the highest score in this map is estimated as the device's location.

The novelty lies in using SVR to generate continuous RSS surfaces, effectively "filling in the gaps" between surveyed points, and then using these surfaces for localization.

Datasets

The experiments were conducted in a real-world environment: the third floor of the west wing of the Polytechnic School at the University of Alcalá (UAH).

  • Training Data: Measurements were collected at 30 significant topological positions. At each position, 30 fingerprints were collected and averaged to create one training fingerprint per position.
  • Test Data: Two types of test data were used:
  • Static Positions: Data collected at the same 30 topological positions used for training, but on a different day.
  • Trajectories: Data collected while a person walked along predefined paths (two different trajectories) in the environment. This data was collected on a separate day and time from the training data. The ground truth for these trajectories was generated by manually tagging known positions and interpolating between them, assuming a constant walking speed.

The paper also compares its method against existing algorithms using their respective datasets or standard implementations.

Key Insights and Ideas

  • Bridging the Resolution-Effort Gap: The most significant insight is that it's possible to achieve higher localization resolution (more precise location estimates) without a proportional increase in the effort required for site surveying. This is a critical step towards making WiFi indoor localization more practical and scalable.
  • SVR for Continuous Estimation: Using SVR to create continuous RSS "reference surfaces" is a powerful way to interpolate signal strength data. This allows the system to estimate locations in areas where no direct measurements were taken.
  • Robustness to Noise: The inclusion of a "noise margin" in the subsurface generation step is crucial for handling the inherent variability and noise in WiFi signals, leading to more stable localization.
  • AP Location Independence: A key advantage highlighted is that the system does not require prior knowledge of the APs' physical locations, simplifying deployment in new or unknown environments.
  • Trained vs. Untrained Areas: The CSE performs well in both surveyed (trained) and unsurveyed (untrained) areas, demonstrating its versatility. While accuracy is slightly better in trained areas, the method significantly improves accuracy in untrained areas compared to traditional methods that would simply fail or provide a poor estimate there.

Key Results

  • Trajectory Localization: For devices in motion (trajectories), the CSE method achieved a mean error reduction of around 62% and 67% compared to the next best algorithms (RF and SVM, respectively).
  • Topological Position Localization: When locating devices at known topological positions, the CSE's mean error was comparable to SVM and better than RADAR, and only slightly worse than RF. Crucially, when the CSE's output was "snapped" to the nearest topological position, its performance became 36% better than SVM and 18% better than RF.
  • Reduced Training Data: The study showed that the number of training positions could be reduced by almost half (from 30 to 18) with only a minor increase in mean error, provided the positions were uniformly distributed. Even with as few as 8 training positions, the CSE still outperformed other algorithms trained with more data.
  • Noise Margin Impact: The analysis indicated that a noise margin of 10 dB or above was sufficient to provide optimal results, balancing signal variation with localization accuracy.
  • Surface Resolution: The mean error decreased as the resolution of the reference surfaces increased, up to a coordinate size of 15 cm, which is close to the wavelength of 2.4 GHz WiFi signals.

Limitations and Future Work

  • Site Survey Not Eliminated: While the CSE significantly reduces the effort of site surveying, it does not eliminate it entirely. The paper suggests that designing an automated fingerprint database generation process and a maintenance system would be a valuable future research direction.
  • Hierarchical Localization: The authors plan to improve the CSE by incorporating a hierarchical localization approach.
  • Sensor Fusion: Future work includes integrating data from other sensors (e.g., Li-Fi, accelerometers, compass) to further enhance localization accuracy, especially for moving devices.
  • AP Selection: Investigating methods for selecting a subset of APs to reduce the computational load for reference surface estimation and RSS searching is also planned.
  • Movement Detection: Incorporating movement detection (e.g., distinguishing between moving and stationary states) could help filter trajectory estimations and improve accuracy.

Summary updated

BibTeX
@article{hernandez2017continuous,
  author = {Noelia Hernández and Manuel Ocaña and José Maria Alonso and Euntai Kim},
  title = {{Continuous Space Estimation: Increasing WiFi-Based Indoor Localization Resolution without Increasing the Site-Survey Effort}},
  journal = {Sensors},
  year = {2017},
  doi = {10.3390/s17010147},
}
Figure