GPS Trajectory-Based Traffic Congestion Detection and Classification Using Delaunay Triangulation and Graph Neural Networks

Traffic congestion degrades urban mobility and economic productivity, yet existing sensor-based monitoring systems detect only where congestion occurs without distinguishing its underlying infrastructure causes. This paper presents a fully automated, detector-free framework that classifies congestion events by their operational mechanisms using raw GPS trajectory data. To overcome the limitations of density-based clustering and manual labeling, we introduce a velocity-adaptive spatial transformation method based on Delaunay Triangulation that converts unstructured trajectory points into semantic graphs while preserving topological relationships essential for pattern recognition. A Graph Neural Network (GNN) classification framework then categorizes these graphs into three operationally distinct congestion types: highway bottlenecks, signalized intersections, and stop-controlled intersections. Experimental validation on 34,459 vehicle trajectories comprising over one million GPS observations from the Detroit metropolitan area demonstrates the effectiveness of this geometric deep learning approach. Among six evaluated architectures, GraphSAGE-LSTM achieved the highest performance with 92.7% classification accuracy, 0.93 precision, and a 0.93 weighted F1-score. These results establish that integrating Delaunay-based geometric clustering with inductive graph learning provides a scalable, infrastructure-independent solution for semantic traffic state classification at the metropolitan scale.