Core Concepts
The authors propose an innovative solution for GNSS positioning by regulating the cost function and accurately estimating measurement errors using a Graph Neural Network.
Abstract
In urban environments, GNSS receivers face challenges due to blocked line-of-sight signals, leading to errors in distance estimation. The authors replace error estimation heuristics with a deep learning model based on Graph Neural Networks. By analyzing the cost function of multilateration, they derive an optimal method to utilize estimated errors. Empirical results show significant improvements in localization accuracy against recent baselines. The proposed solution involves regulating the cost function, estimating measurement errors using a GNN, and selecting robust measurements for improved positioning accuracy.
Stats
Our approach guarantees improvements from 40% to 80% in horizontal localization error.
Dataset contains roughly 120,000 GNSS epochs collected from diverse cities.
LSTM-based neural network shows mean absolute error reduction of 41% compared to our proposed GNN model.
Quotes
"Our approach guarantees that the multilateration converges to the receiver’s location as the error estimation accuracy increases."
"We propose using a graph neural network model to jointly process information across GNSS measurements in an epoch."