Program

Abstract:

GNSS signals are easily blocked or degraded due to high-rise buildings in urban areas, and positioning errors arising from reflected signals amount to as much as hundreds of meters. Various conventional GNSS techniques have been utilized to resolve this problem, but applying them to urban environments has been difficult owing to the complexity and nonlinearity of the reflected signals, which induce multipath and NLOS receptions. In our previous study, a support vector regression (SVR)-based reflection error prediction model was proposed using the relative location information of users and satellites in deep urban areas. In general, the proposed model was applicable to all multipath/NLOS/LOS signals without requiring classification of signal types and it directly eliminated ranging errors induced by the presence of reflection signals in pseudorange measurements. We generated a multipath error map based on the azimuth and the elevation of each satellite using the reflection error prediction model. An application of the error map to static users in deep urban areas exhibited a root mean square error of less than 20 m. However, application of the proposed technique to dynamic users requires the generation of reflection error maps at multiple points. Further, user location uncertainty needs to be resolved in order to identify the optimal option amongst multiple error maps in deep urban areas with excessive outliers. In this study, we generated several reflection error maps for dynamic users in an urban environment and proposed an error map selection algorithm based on residuals in environments involving user position uncertainty. In addition, a modification technique based on data obtained from a satellite with a higher elevation angle and identical azimuth was introduced to train multiple prediction models using limited training data. Training data and test data were collected on roads in Seoul with the highest concentration of high-rise buildings to ensure the effectiveness of the model in environments with severe reflection signal reception. The results of the one-hour driving test revealed that the proposed algorithm improved position accuracy by 30 % horizontally and 74 % vertically, yielding a horizontal RMS error of 23 m over 87 % of the duration of the session. The performance is similar to that in a static test, even though each model was trained using only approximately 2.5 % of the data used in static map construction.