Machine Learning Assisted GNSS Direct Position Estimation for Urban Environments Applications

The GNSS direct position estimation (DPE) technique was proposed as a superior alternative to the conventional two-step scalar tracking loop (STL). Existing literature proves DPE's superiority with simulation data and theoretical bounds. However, in urban areas, its superiority to STL often falters as most of the satellites are error-affected from multipath (MP) and non-line-of-sight (NLOS) reception. The MP and NLOS signals mismatch the existing signal model of DPE, which only assumes line-of-sight (LOS) reception with additive white Gaussian noise. To that end, we aim to solve DPE's misspecified signal model from MP by integrating DPE with a Random Forest Machine Learning (RF ML) regression approach. The RF ML uses multi-correlator outputs from STL's tracking as inputs to estimate the code delays and amplitude of the reflected signals. The estimates are then used to correct the MP-affected autocorrelation function (ACF) to produce the LOS ACF. As a traditional DPE does not involve tracking, the RF ML will be integrated with a homegrown multi-correlator based DPE (Corr-DPE) which uses the correlator outputs and pseudorange estimates from STL. Results from real GNSS data point out that a RF ML-integrated Corr-DPE shows promise in offering more superior performance to STL in urban environments.