It's well known that Global Navigation Satellite System (GNSS) is inaccurate in urban areas, because of many building blocked and/or reflected satellite signals. GNSS shadow matching is a promising positioning technique to improve the GNSS positioning accuracy in urbanized cities. Shadow matching algorithm is initialized by the positioning solution of conventional weighted least square estimation. Then, a grid map is defined. Shadow matching algorithm generates a building boundaries with the help of 3D city model to predict the satellite visibility (whether the direct signal transmission is visible or blocked) of the hypothesis locations in the grid map. By comparing the predicted with observed satellite visibilities at each gird point, a score for each grid point in the search area would be computed. The final positioning solution is calculated based the weighted average of the hypothesis locations, where the weighting is set based by the match score. However, it remains great challenges to identify whether the received GNSS measurements are through line-of-sight (LOS) or none-line-of-sight (NLOS) transmission in urban canyons because the NLOS signals could be received by reflection from buildings. LOS/NLOS classifier trained by machine learning algorithms would give higher recognition rate than the classifier solely based on received signal strength. This paper demonstrates the integration of GNSS shadow matching with the proposed intelligent LOS/NLOS classifier. The LOS/NLOS classifier based on signal to noise ratio (SNR), number of received satellites (NRS), elevation angle (EA), pseudorange residual (PR), pseudorange residual percentage (PRP) and normalized pseudorange residual (NRP). Different machine learning algorithms including k-nearest neighbors (KNN), neural network (NN), support vector machine (SVM), decision tree (TREE) and simple SNR classifier (SSC) algorithms will be implemented and compared. Different scenarios are also considered and used as training data, including mild or middle and deep urban areas with various building distributions. The classification accuracy of SSC considering C/N0 larger than 35 dB-Hz as LOS and smaller than 35 dB-Hz as NLOS, is 69.50% at deep urban areas, and 86.47% at middle or mild urban areas. Using data collected by a u-blox GNSS receiver in dense building areas of Hong Kong, the integrated shadow matching position solutions with the machine learning classifier are more accurate than weighted least squares (WLS)approach. The mean position error using intelligent LOS/NLOS classifier is reduced to 60% of that obtained using WLS.
|Title of host publication||IAIN 2018, Chiba, Japan|
|Publication status||Published - 28 Dec 2018|