GNSS NLOS Identification using Sky-view Images and Machine Learning to Enhance Localization Performance in GNSS-challenged Environments

In complex urban environments, Global Navigation Satellite System (GNSS) positioning accuracy suffers from Non-Line-of-Sight (NLOS) signals and multipath effects. In this study, we propose a method that combines visual information and machine learning to identify and remove NLOS signals to enhance localization performance in GNSS-challenged environments. We generate LOS/NLOS labels using data from an inertial navigation system (INS) and images from a fisheye camera, and then extract three key features, Signal-to-Noise Ratio (SNR), elevation and azimuth to train machine learning models for signal classification. Classification performance of seven different machine learning models was evaluated for comparison. Experimental results show that deep and ensemble learning models, especially Multilayer Perceptron (MLP) algorithm, achieve higher classification accuracy but lower computational efficiency. Also feature importance analysis highlights elevation as the most influential factor. The GNSS Single Point Positioning (SPP) accuracy is improved significantly with the removal of satellites classified as NLOS, and the positioning accuracy improvement with the MLP algorithm is 22.16% and 29.55% in the horizontal and vertical direction, respectively.