Hopular-Based GNSS Signal Reception Classification Method for LOS/NLOS Detection in Urban Environments

Global Navigation Satellite System (GNSS) positioning performance in challenging environments such as urban canyon or indoor environments suffer significant degradations, where frequent none-line-of-sight (NLOS) signals and multipath significantly lower the GNSS positioning accuracy. Consequently, the mitigation of NLOS and multipath effects is important to achieve accurate positioning results. To mitigate the effects from NLOS and multipath signals, the accurate classification of GNSS signal types is required. Recently, the GNSS signal reception classifiers based on deep learning models are drawing more attention due to higher accuracy, better efficiency, and greater convenience. In this paper, a Hopular-based deep learning model is proposed for post-processing GNSS signal classification applications using four GNSS features derived from the raw GNSS measurements: Carrier-to-noise ratio (C/N0), Time-differenced Code-Minus-Carrier (time-differenced CMC), Loss of Lock Indicator (LLI) and Satellites-To-Receiver elevation. The raw GNSS measurements are collected at the two separate locations (Location A & B) under the urban canyon environment in Calgary downtown, using a u-blox ZED F9P receiver. Each measurement is accurately labeled as either line-of-sight (LOS) or NLOS measurement, using a precisely calibrated omnidirectional fish-eye camera with a 360-degree field-of-view lens. Both multi-features and single-feature tests are conducted to evaluate the performance of the Hopular-based model; and their results are compared to another two state-of-the-art machine learning models: Support Vector Machine (SVM) and Gradient Boost Machine (GBM). The trained Hopular-based deep learning model provides a 89.80% and 96.75% classification accuracy of LOS/NLOS signals using all four GNSS features, for dataset A and dataset B respectively. Where the classification accuracy of SVM and GBM models are only 82.66% and 83.71% for dataset A; 80.19% and 82.10% for dataset B. For the dataset A and B, the Hopular-based model has improved 6.09% and 14.65% classification accuracy compared to using GBMs; and 7.14% and 16.56% compared to using SVMs.