Positioning researches using deep learning in urban areas

Abstract

In urban areas, Global Navigation Satellite System (GNSS) radio signals sometimes arrive at receivers with distorted signals in a Non-Line of Sight (NLOS) signal environment. Measurements of the distorted satellite radio signals cause large positioning errors. In this paper, we propose techniques that show improved performance over the conventional positioning techniques using Deep learning technique in the urban areas where positioning performance is degraded sharply. First, we propose an algorithm for estimating pedestrian position using only LOS satellites constellation classified in RNN-based LOS classifier without GNSS positioning process in urban canyon where positioning error is expected to be worst. The proposed algorithm shows about 90% confidence ratio from the experimental results. This ratio can be used to determine whether a pedestrian on the road has crossed the road. The data from the proposed algorithm can be used as useful information to correct the conventional GNSS positioning results in urban canyons. In this study, in order to greatly improve the positioning performance of the conventional LTE fingerprinting technology, which used only the signal strength of LTE radio signals, we proposed a new LTE fingerprinting using multiple LTE measurements based on CNN, a deep learning model specialized for image processing. We built a database based on real data, collected measurements, and used the measurements for the proposed fingerprinting positioning technique. The proposed technique shows about 15m positioning error performance. we prove that our proposed technique is superior to the conventional fingerprinting technology that shows about 70m of positioning error and better positioning performance than GNSS positioning performance in urban area.