Nonlinear Regression-Based GNSS Multipath Modelling in Deep Urban Area

Abstract

As the necessity of location information closely related to everyday life has increased, the use of global navigation satellite systems (GNSS) has gradually increased in populated urban areas. Contrary to the high necessity and expectation of GNSS in urban areas, GNSS performance is easily degraded by multipath errors due to high-rise buildings and is very difficult to guarantee. Errors in the signals reflected by the buildings, i.e., multipath and non-line-of-sight (NLOS) errors, are the major cause of the poor accuracy in urban areas. Unlike other GNSS major error sources, the reflected signal error, which is a user-dependent error, is difficult to differentiate or model. This paper suggests training a multipath prediction model based on support vector regression to obtain a function of the elevation and azimuth angle of each satellite. To extract an unbiased multipath from the GNSS measurements, the clock error of high-elevation QZSS was estimated, and the clock offset with other constellations was also calculated. A nonlinear multipath map was generated, as a result of training with the extracted multipaths, by a Support Vector Machine, which appropriately reflected the geometry of the building near the user. The model was effective at improving the urban area positioning accuracy by 58.4% horizontally and 77.7% vertically, allowing us to achieve a 20 m accuracy level in a deep urban area, Teheran-ro, Seoul, Korea.