A HYBRID FRAMEWORK FOR MODELING SURROUNDING-VEHICLE BEHAVIOR FOR EVALUATING AUTONOMOUS VEHICLE PERFORMANCE IN RISK-CRITICAL DRIVING SCENARIOS

Abstract

Autonomous-vehicle (AV) performance is evaluated most meaningfully not in routine traffic, but in representative risk-critical driving scenarios in which the behavior of surrounding vehicles may have important safety consequences. Because many such scenarios are difficult to observe systematically or reproduce comprehensively through field testing alone, simulation plays a central role in AV safety assessment. In this setting, however, the usefulness of simulation depends strongly on whether surrounding-vehicle behavior is represented with sufficient realism, interpretability, and behavioral richness. Existing modeling paradigms address this need only partially: classical microscopic models provide compact structure, interpretable parameters, and simulation-friendly rollout, but may become behaviorally restrictive under degraded environmental conditions or abrupt interaction changes; purely data-driven approaches offer richer future representation, but may be less interpretable and are not naturally designed for simulation-oriented behavior generation or for incorporating traffic-theory-relevant behavioral priors. This dissertation therefore develops a hybrid framework for surrounding-vehicle behavior modeling for simulation-based AV safety assessment in representative risk-critical driving scenarios. On the traffic-theory-grounded side, the dissertation develops mechanistic behavior-modeling refinements for two important classes of representative risk-critical scenarios: environmental risk and interaction-driven risk. Building on the Intelligent Driver Model (IDM) and related formulations, it introduces weather-sensitive and lane-change-aware extensions to better capture braking, gap regulation, cut-in response, and leader-transition behavior in open-loop forward simulation. To improve long-horizon behavioral realism, forward-simulation calibration is employed instead of purely one-step-ahead fitting. Optimization-based point calibration and Bayesian calibration methods are used to estimate interpretable model parameters, quantify uncertainty, and characterize driver heterogeneity under adverse-weather and lane-changing conditions. On the learned side, the dissertation investigates traffic-theory-informed trajectory prediction as a complementary module for surrounding-vehicle future generation. Rather than treating learning as a flexible replacement for classical car-following response models, it reframes surrounding-vehicle behavior modeling as direct future trajectory prediction conditioned on richer behavioral and scene context. To support this formulation, freeway drone trajectory data are adapted into scene-based prediction scenarios, and Interaction-State Prior Augmentation (ISPA) is introduced to incorporate traffic-theory-relevant behavioral priors into the trajectory-prediction model without hard-constraining it to follow a classical car-following law. The resulting model is evaluated not only with predictive-accuracy metrics, but also with analyses aligned with its prospective use as a surrounding-vehicle behavior-generation component in simulation-oriented settings. Taken together, these contributions establish a hybrid framework that combines mechanistic interpretability, calibration transparency, and rollout-oriented realism with the richer interaction realism enabled by learned future generation. The dissertation focuses on surrounding-vehicle behavior modeling as a component of simulation-based AV safety assessment. Ego-vehicle policy design, controller optimization, closed-loop autonomous-driving evaluation, and runtime AV-in-the-loop orchestration remain beyond its present scope.