TL;DR
This paper introduces PrNet, a novel neural network designed to correct inherent biases in raw GNSS pseudorange measurements from smartphones. By learning to predict and remove these biases, PrNet significantly improves the accuracy of location estimation, outperforming existing methods on a challenging Android GNSS dataset.
What Problem Does This Paper Address?
Ubiquitous smartphones equipped with GNSS receivers hold immense potential for precise localization, enabling a wide range of applications. However, the accuracy of these devices is severely hampered by noisy and biased pseudorange measurements. These biases, stemming from factors like multipath propagation, non-line-of-sight signals, and hardware imperfections, are difficult to model mathematically. While traditional filtering and smoothing techniques can reduce noise, they struggle to effectively mitigate these biases. This paper tackles the critical challenge of correcting these pseudorange biases to unlock more accurate and reliable smartphone-based GNSS localization.
Methods
The core of this work is PrNet, a satellite-wise Multilayer Perceptron (MLP). PrNet takes six key features as input for each visible satellite:
- Carrier-to-noise density ratio (C/No): A measure of signal strength.
- Elevation angle: The angle of the satellite above the horizon.
- Satellite ID (Svid): A unique identifier for each satellite.
- Position estimations: Latitude, longitude, and altitude derived from a model-based approach (like WLS).
- Unit geometry vectors: Representing the direction from the satellite to the receiver.
- Heading estimations: Indicating the smartphone's direction of movement.
PrNet learns to predict the pseudorange bias for each satellite. A crucial aspect of their method is the data labeling strategy. Instead of directly measuring bias, they derive target pseudorange bias values using ground truth location data and a Rauch-Tung-Striebel (RTS) smoother. To further enhance learning, the loss function incorporates the estimation residuals of the smartphone's clock bias. This approach allows for parallel processing of varying numbers of visible satellites by using a "visible satellite mask."
Datasets
The primary dataset used for training and evaluation is the Google Smartphone Decimeter Challenge (GSDC) dataset. This dataset contains raw GNSS measurements collected from Android smartphones in both rural and urban areas. The authors specifically mention using the GSDC 2021 dataset, noting that they excluded traces without altitude ground truth, which is essential for their bias estimation method.
Key Insights and Ideas
- Pseudorange bias is a significant impediment: The paper clearly demonstrates that raw pseudorange biases can be substantial (up to 10 meters) and directly translate into large localization errors, even for sophisticated positioning methods.
- Feature engineering is key: The selection of relevant features, particularly the novel "unit geometry vectors" and "heading estimations," is shown to be crucial for PrNet's performance. These features capture environmental and motion-related information that influences signal quality.
- Data labeling is non-trivial: Deriving accurate labels for pseudorange bias from location ground truth and smoothing techniques is a critical innovation, addressing a limitation of previous methods that struggled with noisy Android data.
- Smartphone GNSS is fundamentally different: The authors highlight that Android GNSS measurements are an order of magnitude noisier than geodetic-quality ones, necessitating specialized approaches like PrNet.
- Model-based + AI is a powerful combination: PrNet doesn't replace traditional positioning engines entirely but rather enhances them. The best results are achieved when PrNet-corrected pseudoranges are fed into an RTS smoother.
Key Results
- PrNet, when combined with the RTS smoother, significantly outperforms baseline model-based methods (WLS, MHE, EKF, RTS) and other data-driven approaches (Set Transformer, PBC-RF, FCNN-LSTM).
- In Scenario I (rural fingerprinting), PrNet+RTS smoother reduces horizontal scores by up to 74% compared to vanilla model-based approaches.
- The paper demonstrates PrNet's robustness across different smartphones in rural areas, though performance degrades in urban settings with Samsung S20 Ultra, suggesting data heterogeneity challenges.
- The computational complexity of PrNet is O(M) (where M is the number of visible satellites), making it efficient for real-time applications.
Limitations and Future Work
- Urban generalization: While PrNet performs well in rural areas, its performance degrades in urban environments when tested on different smartphone models (e.g., Samsung S20 Ultra). This suggests challenges with data heterogeneity in complex environments.
- Data scarcity: The authors note that the GSDC 2022 dataset was not used due to the lack of altitude ground truth, highlighting the need for more comprehensive datasets.
- Future directions: The authors propose several avenues for future research:
- Exploring more advanced deep learning models for better temporal correlation.
- Training the pseudorange-correction network in an end-to-end manner with location ground truth.
- Collecting more open datasets with diverse sensor combinations.
- Leveraging data augmentation to improve generalization across heterogeneous data.
Summary updated