TL;DR
This paper introduces a novel framework called HTL-CD for indoor positioning that tackles the challenge of changing environments and device variations. By leveraging "co-occurrence fingerprint data" and a sophisticated transfer learning approach, it achieves more accurate and robust positioning without needing to constantly retrain models with new labeled data.
What Problem Does This Paper Address?
Indoor positioning systems (FIPS) often rely on "fingerprints" – unique signal characteristics at different locations. However, these fingerprints can change over time due to environmental shifts (e.g., furniture rearrangement, new access points) or differences in the devices used for data collection (e.g., different phone models, how they're held). This "distribution discrepancy" and "feature space heterogeneity" forces frequent model retraining, which is impractical as collecting new labeled data is expensive and time-consuming. Existing transfer learning methods often struggle with this heterogeneity or require labeled data in the target domain, which is precisely what's hard to get.
Methods
The core of the proposed solution is the Co-Occurrence Fingerprint Data-Based Heterogeneous Transfer Learning Framework (HTL-CD). Here's a breakdown of its key components:
- Feature Space Projection: Source domain data (from an older, known environment) is mapped into the target domain's feature space. This is achieved using a mapping matrix
A. - MMD Restriction: To align the distributions of the source and target data, the Maximum Mean Discrepancy (MMD) is used. This minimizes the difference between the marginal (overall) and conditional (given a label) distributions, making the mapped source data more representative of the target environment.
- Co-occurrence Data: The paper defines "wide-sense co-occurrence data" as pairs of fingerprints collected at the same physical location but from different time periods or potentially different devices. This is a more practical definition than "narrow-sense" which requires strict consistency. These pairs act as bridges between the source and target domains.
- Correlation Restriction: Using the co-occurrence data, the framework calculates correlation coefficients between samples. This is crucial for preventing "negative transfer" – where trying to transfer knowledge actually harms performance. If source and target samples are highly correlated, more knowledge is transferred; if not, less is transferred, mitigating the risk of negative impact.
- Objective Function and Optimization: These components are combined into an overall objective function that is iteratively optimized to find the best mapping matrix
Aand ultimately improve positioning accuracy.
Datasets
The experiments were conducted using the UJI-DB dataset. This dataset is notable because it contains Wi-Fi RSS fingerprints collected over 25 consecutive months from a complex indoor environment (a university library). This long-term collection period naturally captures the environmental variations and device differences that the HTL-CD framework aims to address. The authors used data from the first month as the source domain and data from subsequent months as target domains.
Key Insights and Ideas
- "Wide-Sense" Co-occurrence is Key: The practical definition of co-occurrence data, allowing for variations in collection time and device, is a significant enabler. It makes the approach feasible in real-world FIPS scenarios where strict consistency is rare.
- Correlation as a "Smart Filter": The correlation restriction is a clever way to ensure that knowledge transfer is beneficial. By quantifying the relationship between source and target samples, the framework avoids blindly applying transfer learning, which can be detrimental.
- MMD for Distribution Alignment: Using MMD to align both marginal and conditional distributions is a robust approach to handle the inherent differences between the source and target domains, even without labeled target data.
- Robustness to AP Changes: The experiments clearly show that traditional methods degrade significantly when access points (APs) change or disappear. HTL-CD's ability to maintain performance under these conditions is a major advantage.
Key Results
- Improved Accuracy: The HTL-CD framework achieved at least 17.15% smaller average localization errors (ALEs) compared to existing transfer learning-based positioning methods.
- Robustness: The framework demonstrated strong performance even when faced with significant changes in APs (emerging and missing APs) over time.
- Effectiveness without Labeled Target Data: A critical finding is that HTL-CD performs well without requiring accurately labeled samples in the target domain, which is a significant practical advantage.
- Parameter Insensitivity: Robustness analyses showed that the framework's performance is not overly sensitive to the specific values of its design parameters (α and β), indicating a stable and reliable system.
Limitations and Future Work
- Limited to Specific Environments: The authors acknowledge that while the framework is robust to temporal variations, its application in different environments is still limited. Future work will explore this.
- Source Domain Sample Requirement: The HTL-CD scheme still requires an adequate number of samples in the source domain. Future research will focus on addressing scenarios with very small source domain datasets ("small-scale transfer FIPS").
Summary updated