Volatility-Based Short-Term Driving Style Analysis Using an Unsupervised Approach on Large-Scale Sensing Data

Driving behavior analysis is crucial for traffic behavior modeling, risk assessment, and real-time decision-making in dynamic urban environments. However, previous studies based on simulations or surveys often suffer from limited data and subjective bias, limiting their applicability to complex traffic environments. Recent advances in sensing technology and large-scale real-world driving datasets have enabled the collection of high-resolution data. This study presents a framework for quantifying and classifying short-term driving styles using vehicle onboard sensor data. By quantifying speed and acceleration volatility and employing unsupervised learning, the framework identifies distinct behavioral patterns across diverse urban scenarios. The proposed method is validated on the Argoverse 2 dataset, covering over 2000 km of urban roads across six U.S. cities. The framework effectively classifies driving styles into calm, normal, and aggressive classes and accurately identifies risky behaviors in urban traffic scenarios. The results indicate that normal driving is the most common style, accounting for over 58% of intersection and 80% of nonintersection trajectories. Aggressive driving is more prevalent at intersections. Time window analysis reveals that a 3-s duration achieves an optimal balance, capturing 8% of aggressive cases at intersections and 5% at nonintersections. Consistent clustering performance across multiple cities in the dataset further demonstrates the framework's generalizability and robustness. These findings highlight its potential for enhancing driver assistance customization, improving urban traffic efficiency, and refining risk assessment models.