Estimating road segments using natural point correspondences of GPS trajectories

This work proposes a fast and straightforward method, called natural point correspondences (NaPoCo), for the extraction of road segment shapes from trajectories of vehicles. The algorithm can be expressed with 20 lines of code in Python and can be used as a baseline for further extensions or as a heuristic initialization for more complex algorithms. In this paper, we evaluate the performance of the proposed method. We show that (1) the order of the points in a trajectory can be used to cluster points among the trajectories for road segment shape extraction and (2) that preprocessing using polygonal approximation improves the results of the approach. Furthermore, we show based on "averaging GPS segments" competition results, that the algorithm despite its simplicity and low computational complexity achieves state-of-the-art performance on the challenge dataset, which is composed of data from several cities and countries.