TL;DR
This paper tackles the challenge of accurately quantifying the uncertainty in wireless fingerprinting for indoor localization. By using machine learning to predict this uncertainty, the authors demonstrate a significant improvement in the accuracy and reliability of integrated indoor navigation systems.
What Problem Does This Paper Address?
Indoor localization is crucial for location-based services, but Global Navigation Satellite Systems (GNSS) like GPS are unreliable indoors. Wireless fingerprinting, which uses received signal strength (RSS) from access points (APs) to determine location, is a popular alternative. However, the performance of wireless fingerprinting can be inconsistent, making it difficult to assign an accurate "weight" to its location updates when fused with other sensors like inertial measurement units (IMUs) in systems like Extended Kalman Filters (EKFs). This inconsistency arises from various factors like signal fluctuations and environmental complexity. The paper aims to address this by developing a method to predict the uncertainty of wireless fingerprinting location estimates, allowing for more intelligent sensor fusion.
Methods
The core of this work lies in predicting the uncertainty of wireless fingerprinting location estimates using machine learning. Two primary ML approaches are employed:
- Gaussian Distribution (GD)-based Method: This method models the likelihood of a received signal strength (RSS) at a reference point (RP) using a Gaussian distribution. The uncertainty is then derived from the likelihoods of the most probable RPs.
- Artificial Neural Network (ANN)-based Method: Specifically, a Multi-Layer Perceptron (MLP) with two hidden layers and 30 neurons per layer was chosen after evaluating different architectures for a balance between accuracy and computational cost. The ANN takes RSS measurements as input and directly outputs the predicted location uncertainty.
These predicted uncertainties are then used to adaptively set the measurement noise parameters in an Extended Kalman Filter (EKF) that fuses data from an Inertial Navigation System (INS) (dead-reckoning) and wireless fingerprinting. This adaptive setting is a key innovation, moving beyond fixed noise values.
Datasets
The study utilized data collected from indoor walking tests conducted at the MacEwan Student Hall at the University of Calgary.
- Training Data: Collected using five Android smartphones (Samsung Galaxy S4, Galaxy S7, Huawei P10, Lenovo Phab 2 Pro, and Nexus 9 tablet) equipped with inertial sensors and WiFi receivers. Two testers walked five training trajectories, resulting in 12,867 WiFi fingerprints. Each fingerprint included MAC addresses and RSS values from available APs.
- Testing Data: Collected along two pre-designed testing trajectories, lasting approximately 30 minutes each. This resulted in 7,692 WiFi fingerprints.
- Reference Trajectories: Generated using a Lenovo Phab 2 Pro smartphone that also captured RGB-D images. Simultaneous Localization and Mapping (SLAM) was used to compute these reference trajectories, providing ground truth locations.
Key Insights and Ideas
- Uncertainty is Key: The paper strongly emphasizes that simply providing a location estimate from wireless fingerprinting is insufficient. Understanding and quantifying the uncertainty associated with that estimate is critical for robust sensor fusion.
- ML for Uncertainty Prediction: Machine learning, particularly ANNs, can effectively learn the complex relationship between raw RSS measurements and the resulting location uncertainty, outperforming simpler statistical methods like the GD-based approach in some aspects.
- Adaptive Noise for EKF: Dynamically adjusting the measurement noise in an EKF based on predicted wireless fingerprinting uncertainty significantly improves the overall localization accuracy, especially in areas where fingerprinting performance is poor.
- ANN vs. GD: While both ML methods showed promise, the ANN approach offered a more stable prediction of uncertainty compared to the GD method, which was more sensitive to individual reference points.
Key Results
- The proposed ML-based adaptive noise setting for the EKF significantly reduced indoor localization errors.
- Compared to a constant noise setting (DR/WiFi-CN), the DR/WiFi-GD strategy reduced RMS errors by 27.9% and 95% errors by 32.3%.
- The DR/WiFi-ANN strategy reduced RMS errors by 23.3% and 95% errors by 28.7%.
- Both ML-based strategies (GD and ANN) effectively mitigated large position errors that occurred with the constant noise approach.
- The DR/WiFi-ANN strategy achieved the smallest maximum error among the integrated methods.
- The study demonstrated that the majority of location errors could be kept within 4 meters using the ML-predicted uncertainty.
Limitations and Future Work
- Computational Cost: While the ANN architecture was optimized for balance, training and inference still incur computational costs, which might be a consideration for resource-constrained devices.
- Generalizability: The study was conducted in a specific indoor environment. Further research could explore the generalizability of the proposed methods to different indoor settings and wireless infrastructures.
- Data Requirements: The ML models require sufficient training data to learn the uncertainty patterns effectively. The quality and coverage of the training data are crucial.
The authors suggest that investigating location uncertainty alongside the location solution is essential for achieving more reliable indoor localization. Future work could focus on exploring more advanced ML architectures, different types of wireless signals, and diverse indoor environments to further enhance the robustness and accuracy of such systems.
Summary updated