Machine Learning–Driven Safety Analysis of Automated Vehicles in Simulated Mixed Traffic Environments

<div>Automated Vehicles (AV) pose new challenges in road safety, multimodal interaction, and urban planning, requiring a holistic approach that prioritizes sustainability and protects all road users. The KASSA.AST project addresses this by deploying and evaluating an automated shuttle in southern Austria on three routes. The study area is a Park & Ride zone near a train station, enabling seamless transfers and higher transit use. To assess the safety impacts of the automated shuttle, four Mobility Observation Boxes (MOBs) were deployed. These AI-based systems detect and classify road users, track their trajectories and geospatial coordinates, and identify safety-critical events via Surrogate Safety Measures (SSMs). Over 10 days, a trajectory dataset captured interactions among vehicles and the shuttle. The resulting real-world dataset is a core contribution. This dataset underpins microscopic behavior modeling. Trajectory pairs yield car-following and interaction metrics (relative distance, relative speed, acceleration) to calibrate custom models for realistic mixed traffic. Simulations generate a structured interaction database with time spans, trajectories, conflict points, and SSMs (such as Time-to Collision—<i>TTC</i>, Post-Encroachment Time—<i>PET</i>, and Deceleration-rate-to-avoid-crash—<i>DRAC</i>). These outputs support detailed analysis of shuttle interactions, including near misses. To reveal patterns, clustering identified three interpretable safety-relevant regimes: (i) a low-demand background regime (<i>n</i> = 96) with low speeds and near-zero deceleration demand, (ii) a fast-and-tight regime (<i>n</i> = 33) with reduced <i>TTC</i>, elevated critical-event speeds, and high <i>DRAC</i>/<i>Modified (M)DRAC</i> demand, and (iii) an AV-regulated regime (<i>n</i> = 10) dominated by the shuttle as adversary, showing short <i>TTC</i> but stable moderate speeds (~4 m/s) and conservative headway policies. Ensemble-tree supervised learning reproduced these regimes with high accuracy and revealed that critical-event speeds and counterpart headway are the strongest discriminators, while AV role metadata contributes marginally. This integrated approach—linking field data, behavior modeling, simulation, and machine learning—provides a robust framework for assessing AV safety in urban contexts.</div>