Multipath signals are a common problem that degrades the accuracy of GPS, and occur mainly in urban environments where high buildings are crowded. Multipath signals are very fatal for system such as autonomous vehicles and autonomous flying drones that have to operate in the downtown with high positioning performance. Several recent studies have proposed first arrival path detection techniques using deep learning. Among them, FPDNet which is SOTA estimates the code delay of the first arrival path from the auto-correlation function (ACF) output image. However, FPDNet has the disadvantage that it generalizes signals with different property, which reduces their efficiency. Furthermore, code delay values which are regression results is not interpretable. In this paper, we propose a classifier network that separates signals based on C/N0, which represents the state of the signal according to noise, and estimates the probability of first arrival path for all samples in the ACF. By separating signals based on C/N0 and teaching different networks, individual characteristics of signals according to C/N0 can be intensively learned, and the results of first arrival path detection can be interpreted by estimating probabilities. Crucially, using soft labels to be estimated, the characteristics of not only first arrival path but also multipath can be learned, so that the characteristics of the first arrival path can be learned more quickly and accurately. We propose a novel first arrival path detection technique with the above advantages, while simultaneously comparing its performance with FPDNet to demonstrate the superiority of the proposed technique. |
