Automatic Calibration of Road Intersection Topology Using Trajectories (CITT Framework)

Authors: Liu Guoping, Ma Nan. Source: Didi Technology Collaboration. The paper "Automatic Calibration of Road Intersection Topology using Trajectories" was accepted at ICDE 2020.

Road intersections are critical for digital maps, and GPS trajectory data contain rich topological information. To improve map accuracy, Didi and East China Normal University propose the CITT (Three‑stage Intersection Topology Calibration) framework.

1. Trajectory Quality Enhancement – Raw GPS data are noisy due to device errors and signal issues. Trajectories are segmented based on spatial‑temporal continuity, noisy points at intersections are filtered using spatial density, and the Douglas‑Peucker algorithm compresses trajectories while preserving turning features.

2. Core‑Area Detection – A quadtree partitions space with a minimum cell size of 25 m, starting search from the 200 m level. Mean‑shift and DBSCAN (direction‑based) identify grid cells with high turning density and low speed, locating the intersection center. An annular geometric model then expands outward layer by layer; density of turning points decreases and speed increases toward the edge, allowing the method to adapt to various intersection shapes (e.g., roundabouts, overpasses).

3. Influence‑Zone Topology Calibration – Around the detected center, a radius threshold defines the influence zone. All trajectories inside are extracted, turning clusters are identified using a direction‑weighted Frechet distance combined with DBSCAN, and a force‑attraction clustering refines each cluster’s central path. The reference trajectory is chosen via a Frechet‑based sampling strategy to reduce bias. Finally, a classic Hidden Markov Model (HMM) matches the fitted turning paths to the baseline road network, accelerated by constructing convex hulls for each intersection.

Experimental Results – Compared with three industry methods (CBTP, Ahmed2012, Kharita), CITT achieves the highest precision and F‑score in intersection position detection. For turning‑path shape fitting, CITT outperforms the HyMU sweeping algorithm, showing smaller deviations from ground‑truth roads. In topology calibration, CITT attains 70 % precision, detecting missing and offset turning paths with a 1:5 ratio, as illustrated by the case studies.

Future Work – The authors plan to extend CITT to lane‑level topology calibration and to develop an automated differential fusion pipeline that integrates calibrated results with the baseline map, further enhancing navigation reliability.