A deep embedded clustering method for location-specific driving safety profiling using trajectory data

Understanding driver behavior is critical for enhancing road safety and enabling proactive interventions. While prior studies have largely focused on driver-specific profiles (e.g., aggressive, cautious, safe), such approaches overlook the fact that crashes are often concentrated at specific crash hostspot and corridors. Hence, this study adopts a location-based perspective, analyzing how drivers behave at particular roadway segments to identify their driving patterns. The objectives are: (i) to develop a robust machine learning framework that classifies location-specific driving behavior according to its risk level, and (ii) to create a unified model capable of analyzing multiple vehicle classes together, eliminating the need for separate models for each type. Trajectory data from different vehicle classes, collected in Chennai, India, form the basis of this analysis. Initially, each vehicle type was examined independently to capture mode-specific behaviors at various locations. The datasets were then combined to investigate cross-modal behavior patterns. Principal Component Analysis (PCA) was applied to reduce dimensionality, and four clustering techniques: K-Means, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), Mean Shift, and Deep Embedded Clustering (DEC) were employed to classify location-based behaviors into 'Aggressive,' 'Cautious,' and 'Safe' categories. The clustering outcomes were systematically evaluated using Silhouette Score, Davies-Bouldin Index (DBI), and Calinski-Harabasz Index (CHI). The findings show that DEC performs best, while DBSCAN yields the weakest clustering results. The unified location-based model demonstrates strong potential for large-scale deployment in real-world scenarios, offering valuable insights into risky driving hotspots and enabling targeted interventions to improve driver awareness and roadway safety.