TL;DR
This paper introduces a novel "Hybrid Wireless Fingerprint" (HW-fingerprint) that combines traditional RSSI data with a new "Ratio fingerprint" to better capture dynamic indoor environments. By feeding this enhanced fingerprint into a Convolutional Neural Network (CNN), the authors achieve improved indoor localization accuracy compared to existing methods.
What Problem Does This Paper Address?
Accurate indoor localization remains a significant challenge. While GPS works well outdoors, its signals are too weak to penetrate buildings effectively. Existing indoor localization methods, particularly those relying on Wi-Fi Received Signal Strength Indicator (RSSI) fingerprints, suffer from accuracy degradation due to environmental changes. Factors like moving objects, doors opening/closing, and even subtle shifts in humidity can alter RSSI values, making historical fingerprint data less reliable over time. This paper aims to address this by creating a more robust fingerprint that is less susceptible to these environmental fluctuations.
Methods
The core innovation is the Hybrid Wireless Fingerprint (HW-fingerprint). This is constructed in two main parts:
- RSSI Fingerprint: This is the standard approach, using the measured RSSI values from various Wi-Fi Access Points (APs).
- Ratio Fingerprint: This is a novel addition. It's created by calculating the ratio of RSSI values between different important APs. The rationale is that while absolute RSSI values might change due to environmental factors (like humidity affecting signal attenuation), the relative strength between APs might remain more stable, providing a more consistent characteristic of the environment.
To build the Ratio fingerprint, APs are first selected based on a "contribution weight" (derived from how frequently they are detected). Then, the ratios of RSSI values between these selected APs are computed.
Finally, the HW-fingerprint combines the standardized RSSI fingerprint and the Ratio fingerprint. This combined feature vector is then fed into a Convolutional Neural Network (CNN). The CNN is chosen for its proven ability to extract complex spatial features from data, treating the fingerprint as an image-like input. The CNN learns to map these complex HW-fingerprints to specific locations within the indoor environment.
Datasets
The experiments were conducted in a real indoor laboratory environment at Guangdong University of Technology. Data was collected over 15 days. The first day's data was used for training the CNN model, while data from the subsequent days was used for testing. The paper does not mention the use of any pre-existing public indoor localization datasets.
Key Insights and Ideas
- Ratio fingerprint as a stabilizer: The key insight is that the ratio of RSSI values between APs can act as a more stable feature than absolute RSSI values, helping to mitigate the impact of environmental changes like humidity.
- CNN for complex feature extraction: Treating the HW-fingerprint as an image allows CNNs to leverage their powerful feature extraction capabilities to learn subtle patterns that might be missed by simpler machine learning algorithms.
- HW-fingerprint improves existing algorithms: The proposed HW-fingerprint isn't just for CNNs; it also demonstrably improves the performance of traditional algorithms like KNN and SVM, highlighting its general utility.
- Thresholding is important: The selection of APs for the Ratio fingerprint is crucial, and the paper shows that a higher contribution threshold (0.9) leads to better accuracy, suggesting that focusing on more significant APs is beneficial.
Key Results
- Improved Accuracy: The HW-fingerprint significantly improved the average daily location accuracy for all tested algorithms:
- KNN: improved by 3.39%
- SVM: improved by 8.03%
- CNN: improved by 9.03%
- CNN Superiority: The CNN, when using the HW-fingerprint, outperformed both SVM and KNN:
- CNN was 4.19% higher than SVM.
- CNN was 16.37% higher than KNN.
- Distance Error Reduction: The HW-fingerprint also reduced the average distance error for all algorithms, with CNN achieving the lowest error (3.9118 m) when using the HW-fingerprint.
- Negligible Runtime Impact: The increase in testing time when using the HW-fingerprint was negligible, indicating no significant performance overhead for real-time applications.
Limitations and Future Work
- Limited Environment Size: The authors acknowledge that their experiment environment was relatively small. They suggest that in smaller areas, prediction errors can arise due to the close proximity of Reference Points (RPs).
- AP Selection Method: The method for selecting important APs for the Ratio fingerprint was based on statistical analysis of training data. This can lead to issues if these "important" APs become unavailable or their signals are lost during subsequent testing periods, introducing noise into the online HW-fingerprint.
- Future Work: The authors propose considering larger environments and focusing on more robust methods for stable AP selection and improving overall positioning performance. They also mention that in future work, they will focus on the work of stable AP selection and methods to increase the positioning performance of the system.
Summary updated