Multi-scale Graph Convolutional Network for Intersection Detection from GPS Trajectories.

To facilitate the reconstruction of high-quality road networks, intersections as the key locations provide valuable information about the network topology. However, only a few efforts have been made on the data-driven automatic detection of intersections from, e.g., large-scale GPS trajectories. To bridge the gap, we propose a machine learning based intersection detection approach based on large-scale real-world GPS trajectories of drivers from the Grab ride-hailing service. Instead of representing locations with vector descriptors, we innovatively propose a graph representation that models a location together with its local surroundings to improve the descriptiveness of the location descriptors. Moreover, we present a multi-scale graph convolutional network (GCN) to generate robust graph-level descriptors, followed by logistic regression to discriminate intersections from non-intersections. The experimental results show that our proposed multi-scale graph model outperforms the conventional multi-scale vector representation by 8.5%. Appealingly, the proposed graph representation can be considered as a general location descriptor, which can be used in a variety of geo-based applications other than intersection detection for location modeling.