TL;DR
This paper tackles the challenge of insufficient data for accurate indoor Wi-Fi localization by proposing a novel data augmentation technique using Multi-Output Gaussian Processes (MOGP). By leveraging correlations between Wi-Fi signals from different access points, their method effectively generates synthetic data, significantly improving localization accuracy in large-scale, multi-building, and multi-floor environments.
What Problem Does This Paper Address?
Indoor localization, crucial for location-based services, often relies on Wi-Fi Received Signal Strength Indicators (RSSIs) and a technique called "fingerprinting." This method involves mapping Wi-Fi signal strengths at known locations (reference points) to create a "fingerprint database." Machine learning models, particularly Deep Neural Networks (DNNs), can then use this database to estimate a user's location. However, building comprehensive fingerprint databases for large indoor spaces is incredibly time-consuming and labor-intensive. This data scarcity can lead to poor spatial representation and hinder the accuracy of localization models, especially in complex environments with multiple buildings and floors. The COVID-19 pandemic further highlighted the challenges of collecting data in such settings.
Methods
The core innovation of this paper is the application of Multi-Output Gaussian Processes (MOGP) for multi-dimensional data augmentation. Unlike traditional Single-Output Gaussian Processes (SOGP) that model one Wi-Fi access point (AP) at a time, MOGP can simultaneously model RSSIs from multiple APs. This allows it to capture the inherent correlations between signals from different APs, even across different floors or buildings.
The authors propose three "data augmentation modes":
- By a Single Floor: MOGP is fitted using data from APs on a single floor.
- By Neighboring Floors: MOGP considers RSSIs from adjacent floors to augment data for a specific floor.
- By a Single Building: MOGP is fitted using all RSSI data within an entire building.
They also explore various MOGP kernels (RBF, Rational Quadratic, Matérn family, Ornstein-Uhlenbeck) and their hyperparameters (variance, length scale) to find the optimal configuration for data augmentation. The augmented data is then used to train a hierarchical Recurrent Neural Network (RNN) for indoor localization.
Datasets
The primary dataset used in this research is the UJIIndoorLoc database. This is a publicly available, large-scale, multi-building, and multi-floor Wi-Fi fingerprint database, widely used as a benchmark in the field. The authors split the existing training/validation data into new training and validation sets (70:30 ratio) and used the validation set as their testing dataset.
Key Insights and Ideas
- MOGP's Advantage: The ability of MOGP to model correlations across multiple Wi-Fi access points is a significant advantage over SOGP, which treats each AP independently. This multi-dimensional approach leads to more effective data augmentation.
- "By a Single Building" Mode Dominance: Augmenting data at the building level ("by a single building") yielded the best localization performance. This suggests that considering the entire building's signal landscape, including inter-floor correlations, is crucial for large-scale localization.
- Kernel Choice Matters: The Matérn5/2 kernel, with specific variance and length scale hyperparameters, proved to be the most effective for this data augmentation task, outperforming other common kernels.
- Augmentation Ratio Trade-off: While augmenting data significantly (e.g., 10x the original data) can improve coverage, there's a point where excessive augmentation might lead to ignoring original data features. A ratio of 1 (doubling the data) provided the best results in their experiments.
- Interpretability of MOGP: Compared to DNNs, MOGP offers better interpretability, allowing researchers to understand how augmentation affects the data and potentially control the process more effectively.
Key Results
The proposed MOGP-based data augmentation, when applied to the UJIIndoorLoc database and used to train a hierarchical RNN, achieved a 3D localization error of 8.42 meters. This represents an improvement over the baseline RNN model without augmentation, which had an error of 8.62 meters. The "by a single building" augmentation mode specifically achieved this best result. The study also showed that using the Matérn5/2 kernel with a variance of 1 and a length scale of 10, and an augmentation ratio of 1, provided optimal performance.
Limitations and Future Work
- Building Structure Ignorance: The authors acknowledge that the UJIIndoorLoc database lacks detailed building coordinates and internal floor structure maps. This limitation means the augmentation process, while effective, cannot fully account for building and floor structures, which could further enhance positioning accuracy.
- Kernel Hyperparameter Tuning: Deciding on the optimal kernel hyperparameters for specific datasets and applications remains an open research question.
- Time-Varying Interference: The current approach does not explicitly account for time-varying interference factors (e.g., moving crowds, temporary APs), which are common in real-world indoor environments.
- Extended MOGP Models: The study focused on linear MOGP models (ICM and LMC). Future work could explore other MOGP models and kernels better suited for indoor localization and diverse databases.
- Dynamic Environments: Incorporating time-varying interference factors into the data augmentation process is suggested as a promising avenue for future research.
Summary updated