Machine Learning in GNSS Multipath/NLOS Mitigation: Review and Benchmark

Abstract

Global navigation satellite system (GNSS) positioning accuracy is degraded in urban canyons due to signal blockages and reflections, which is still a major challenge. Recently, using machine learning to improve the accuracy of GNSS positioning in urban areas has become a new trend. This paper summarizes the works focused on GNSS multipath/non-light-of-sight (NLOS) mitigation using machine learning. The review of the studies is categorized based on the input features, algorithms, and outputs. The categorization shows that the received signal strength, elevation angle, and receiver correlator outputs from a single channel of satellite signal are the most popular input features. For the algorithm selection, the support vector machine and fully connected neural network (FCNN) are the algorithms most widely used. In terms of the outputs, most of the works made improvements in measurement status prediction, namely, LOS, multipath, and NLOS. Besides, this paper also provides an open-source dataset with four scenarios for machine learning algorithms for the GNSS multipath/NLOS mitigation. Finally, the benchmarks are established based on the proposed dataset and the FCNN and least-squares estimation to enable performance evaluation in Kaggle.