Indoor Localization Using Data Augmentation via Selective Generative Adversarial Networks

Wafa Njima, Marwa Chafii, Arsenia Chorti, Raed M. Shubair, H. Vincent PoorIEEE Access 2021 · DOI

Abstract

Accurate indoor localization is critical for various location-based services. Deep neural network (DNN) based received signal strength indicator (RSSI) fingerprinting offers high performance but requires extensive training data, making site surveys costly. This paper introduces a data augmentation framework using Generative Adversarial Networks (GANs) to generate synthetic RSSI data from a limited set of real labeled samples. By employing semi-supervised learning for pseudo-labeling and a dual-selection criterion for spatial coverage and realism, the proposed "Selective GAN" improves localization accuracy by 21.69% on simulated data and 15.36% on experimental data.

Introduction: The Data Scarcity Bottleneck

Indoor localization serves as the "GPS for indoors," enabling applications from warehouse robotics to retail navigation. RSSI fingerprinting is a leading technique where unique signal patterns from Wi-Fi access points are mapped to specific coordinates. While DNNs excel at this mapping, their performance is heavily dependent on the volume of training data. Collecting thousands of real-world signal measurements is labor-intensive and expensive. This research addresses the challenge of training high-performing models with minimal real-world labeled data.

Methodology: The Selective GAN Framework

The proposed framework utilizes a semi-supervised approach to expand small datasets intelligently. The process is divided into offline and online phases, primarily focusing on the Central Unit (CU) tasks during the offline phase.

Overview of the proposed localization framework, illustrating the offline training pipeline and online prediction phase.

1. Data Generation and GAN Architecture

A GAN is trained on the initial small set of real RSSI data. The Generator ($G$) creates synthetic RSSI vectors from random noise, while the Discriminator ($D$) learns to distinguish these from real samples.

The GAN architecture used for RSSI data augmentation, featuring fine-tune training loops for the Generator and Discriminator.

2. Pseudo-Labeling

Because the GAN generates signal data without associated coordinates, a supervised DNN is trained on the small real dataset to assign "pseudo-labels" (estimated coordinates) to the synthetic RSSI vectors.

3. Selective Filtering

The core innovation lies in the "Selective" filter, which applies two critical criteria to the generated data:

  • Criterion 1 (Environment Coverage): The indoor area is partitioned into zones. Synthetic data is selected to ensure even distribution across the entire map, preventing spatial bias.
  • Criterion 2 (Realism): The Discriminator acts as a quality-control gatekeeper. It ranks synthetic samples based on their "credibility" (loss values), and only the most realistic samples are retained.

4. Final Model Training

A "mixed" DNN is then trained using a combination of the original real dataset and the filtered, high-quality synthetic dataset.

Experimental Evaluation

The framework was validated using two distinct datasets:

  • Simulated Environment: A 400 $m^2$ area with 10 access points, incorporating realistic pathloss and shadowing. To simulate real-world conditions, 40% of signals were intentionally omitted.
  • Experimental Data: The public UJIndoorLoc database (Building 1, Floor 2), featuring 1,395 training fingerprints across 80 positions.

Results and Discussion

The "Selective-SS-GAN" approach demonstrated significant performance gains:

  • Accuracy Improvements: Localization accuracy improved by 21.69% in simulated environments and 15.36% in real-world tests compared to models using only real data.
  • Efficiency: Results indicated that a model trained on 1,000 real samples plus 1,000 "selective fake" samples achieved performance comparable to a model trained on 2,000 real samples. This suggests a potential 50% reduction in manual site survey labor.

The study highlights that quality and spatial diversity are more important than sheer volume. Simply adding more synthetic data can introduce errors if the data is redundant or unrealistic.

Conclusion

By leveraging the Discriminator as a quality filter and enforcing spatial balancing, the Selective GAN framework effectively mitigates the data scarcity problem in indoor localization. Future work will explore more advanced GAN architectures and the impact of dynamic environments on signal patterns.

Summary updated

BibTeX
@article{njima2021indoor,
  author = {Wafa Njima and Marwa Chafii and Arsenia Chorti and Raed M. Shubair and H. Vincent Poor},
  title = {{Indoor Localization Using Data Augmentation via Selective Generative Adversarial Networks}},
  journal = {IEEE Access},
  year = {2021},
  doi = {10.1109/access.2021.3095546},
}
Figure