Multipath Parameters Estimation in Physically Based Synthetic Environment Using Robust Deep Neural Regression

Global Navigation Satellite System (GNSS) signals reception may suffer from multipath propagation. Specifically, in deep urban environment the disturbance can be strong. It may lead to a degraded position solution at the expense of the final user. Despite the development of dedicated techniques like Multipath Estimating Delay Lock Loop (MEDLL) in Townsend et al. (1995), multipath error mitigation is difficult to achieve and still need to be improved. In this work, we propose a Deep Learning (DL) method to estimate the multipath parameters generated by a GNSS signal propagation simulator. As in Munin et al. (2020) we focus on multiple correlator outputs to construct the datasets. The correlator outputs form images containing both signal and multipath information. The number of correlator outputs is flexible in order to generate images with different size, so with various resolutions. We assume that a multipath can be entirely characterized by four parameters namely delay, Doppler frequency shift, magnitude and phase. The disturbed signals are retrieved from a synthetic environment, and more specifically from urban canyons where the multipath phenomenon is worse. A Convolutional Neural Network (CNN) model is trained on physically based synthetic data with various images sizes to assess the proposed CNN algorithm performances. Moreover, a soft-labelling method is applied to enhance the CNN regression task. The data labels take the form of distributions and a specific histogram loss function based on Kullback-Leibler (KL) divergence will be applied during the CNN training.