Extracting 3D Maps from Crowdsourced GNSS Skyview Data.

3D maps of urban environments are useful in various fields ranging from cellular network planning to urban planning and climatology. These models are typically constructed using expensive techniques such as manual annotation with 3D modeling tools, extrapolated from satellite or aerial photography, or using specialized hardware with depth sensing devices. In this work, we show that 3D urban maps can be extracted from standard GNSS data, by analyzing the received satellite signals that are attenuated by obstacles, such as buildings. Furthermore, we show that these models can be extracted from low-accuracy GNSS data, crowdsourced opportunistically from standard smartphones during their user's uncontrolled daily commute trips, unleashing the potential of applying the principle to wide areas. Our proposal incorporates position inaccuracies in the calculations, and accommodates different sources of variability of the satellite signals' SNR. The diversity of collection conditions of crowdsourced GNSS positions is used to mitigate bias and noise from the data. A binary classification model is trained and evaluated on multiple urban scenarios using data crowdsourced from over 900 users. Our results show that the generalization accuracy for a Random Forest classifier in typical urban environments lies between 79% and 91% on 4 m wide voxels, demonstrating the potential of the proposed method for building 3D maps for wide urban areas.