PICT: Precision-Enhanced Road Intersection Recognition Using Cycling Trajectories

Abstract

To recognize road intersections using cycling trajectories accurately is vital to the quality of the digital map that cycling navigation apps use. However, the existing approaches mainly identify road intersections based on motor vehicles’ trajectories, and they fail to tackle unique challenges posed by cycling trajectories: (i) Cycling trajectories of minor intersections and their adjacent road segments are quite sparse. (ii) Turning behaviors occur at different areas in intersections of various sizes. To address the above challenges, in this paper, we propose a precision-enhanced road intersection recognition method using cycling trajectories, called PICT . Initially, to enhance the representations of minor intersections, a grid topology representation module is designed to extract intersection topology. Then an intersection inference module based on multi-scale feature learning is put forward to identify the intersections of different scales correctly. Finally, extensive comparative experiments on two real-world datasets demonstrate that PICT significantly outperforms the state-of-the-art methods by 52.13% in the F1-score of intersection recognition.