TL;DR
This paper introduces a novel semi-supervised learning (SSL) framework for Wi-Fi RSSI fingerprinting-based indoor localization. By leveraging the Mean Teacher model with techniques like AP selection and batch-level noise injection, it significantly improves localization accuracy and adaptability, especially in dynamic environments, while reducing the reliance on costly labeled data.
What Problem Does This Paper Address?
Indoor localization using Wi-Fi RSSI fingerprinting is a promising alternative to GPS, but it faces several challenges. Firstly, collecting large amounts of accurately labeled data (RSSI measurements paired with precise locations) is time-consuming and expensive. Secondly, supervised learning (SL) models trained on this data struggle to generalize well to new environments or adapt to dynamic changes in the wireless signal landscape, leading to performance degradation over time. Existing semi-supervised learning (SSL) methods have shown promise but haven't been fully explored for this specific application, particularly concerning scalability and continuous adaptation in real-world deployments.
Methods
The core of the proposed solution is a Mean Teacher-based Semi-Supervised Learning (SSL) framework. This framework integrates several key components:
- Mean Teacher Model: This is the foundation, using a teacher-student model architecture where the teacher's weights are an exponential moving average (EMA) of the student's weights. This approach generates more stable target labels for unlabeled data compared to methods that average predictions.
- AP Selection: A feature selection technique that identifies the most informative Wi-Fi access points (APs) based on the number of unique RSSI values. This reduces dimensionality and enhances generalization.
- Model Pre-training/Cloning: The student and teacher models are initialized by pre-training a base model on labeled data and then cloning its weights. This helps alleviate the "cold start" problem and speeds up training.
- Batch-level Noise Injection: To generate synthetic unlabeled data from labeled data, a truncated additive white Gaussian noise is injected into the RSSI features. This technique, while common in other domains, is applied here to improve robustness and generalization in Wi-Fi fingerprinting.
The framework supports two training scenarios: static training (using a hybrid labeled/unlabeled dataset during the offline phase) and dynamic training (continuously retraining the model with unlabeled data from deployed users during the online phase).
Datasets
- UJIIndoorLoc Database: A widely-used benchmark dataset for multi-building and multi-floor indoor localization. It contains RSSI measurements from 520 APs across three buildings. The authors used its validation set for both validation and testing.
- XJTLU Dynamic Database: A subset of this database, collected over 44 days on a single floor, was used to better evaluate dynamic training scenarios. This dataset is noted as being under construction and not yet publicly released.
Key Insights and Ideas
- Leveraging Unlabeled Data is Crucial: The paper strongly emphasizes that unlabeled data, whether from crowdsourcing or active users, is key to overcoming the limitations of purely supervised approaches in indoor localization.
- Mean Teacher for Stability: The Mean Teacher model's EMA-based weight averaging provides more stable and reliable pseudo-labels for unlabeled data, leading to better generalization.
- Dynamic Training for Real-World Adaptability: The ability to continuously retrain the model with new, unlabeled data from users during the online phase is vital for maintaining accuracy in dynamic wireless environments.
- Synergy of Components: The combination of AP selection, pre-training, and noise injection works synergistically with the Mean Teacher to improve robustness and reduce the need for extensive labeled data.
Key Results
- Improved Accuracy: The proposed SSL framework consistently outperformed conventional SL methods.
- On the UJIIndoorLoc database, mean 3D errors were reduced by 7.403% for CNNLoc and 7.748% for SIMO-DNN in static training.
- In dynamic training with the XJTLU database, maximum 2D error reductions reached up to 49.227% (Period 0) and 39.755% (Period 1).
- Reduced Data Requirements: Experiments showed that using only 50% of labeled data with the SSL framework could achieve better performance than using 100% labeled data with SL, significantly reducing data collection costs.
- Robustness: The framework demonstrated improved robustness, evidenced by reduced interquartile ranges (IQRs) in error distributions, indicating more consistent performance.
- Scalability: The Mean Teacher approach is noted for its better scalability to large datasets and online learning compared to other SSL methods like Temporal Ensembling.
Limitations and Future Work
- AP Selection Limitations: The current AP selection algorithm focuses on feature consistency for SSL and doesn't consider factors like AP load or signal-to-noise ratio (SNR). Future work could involve a more adaptive, multi-dimensional AP selection strategy.
- UJIIndoorLoc Database Limitations: This database lacks real temporal dynamics and has significant spatial coverage differences between its subsets, limiting the evaluation of dynamic training. The XJTLU dynamic database, while more suitable, is still under development.
- Multi-Device Heterogeneity: The framework currently does not account for RSSI deviations caused by different devices.
- Future Directions: The authors suggest designing more advanced AP selection algorithms, generating high-fidelity synthetic unlabeled data (e.g., using multi-output Gaussian processes), and extending the framework to large-scale scenarios with multi-device heterogeneity.
Summary updated