Intersection Identification Using Large-Scale Vehicle Trajectory Data

Abstract

AbstractIntersections are critical facilities in traffic control and management with vital impact on traffic mobility and safety. Identifying the intersections in a network is the first step. As a major source of traffic data, trajectory data based on Global Positioning System (GPS) locations from probe vehicles are widely used. Thus, how to automatically extract the information and identify the intersections out of massive and noisy GPS data is important. This study uses machine learning method associated with empirical rules to find the pattern in vehicle trajectories and detect intersections at a city scale. First, the turning points are extracted from massive GPS points by examining the bearing changes. Then, a density-based clustering method is applied to cluster all the turning points as potential intersection area. To remove the false detections, two types of indices are developed: (a) direction discrepancy index (DDI) and (b) turning discrepancy index (TDI). By examining the two indices in each potential detection, clusters with high discrepancy in travel direction and turning direction distribution are filtered out. The methods are applied on a collection of millions GPS records. The raw trajectory data are noisy with inconsistent record intervals and imprecise locations. The results are compared with Open Street Map (OSM) database. Although the results generated from GPS data do not cover some points due to data availability, they can detect the points that are missed by OSM database. This method is easy to scale with few interventions; thus, it could be a good supplement to the open source geographical database.KeywordsVehicle GPS trajectoryIntersection identificationLarge-scale network