Autonomous Driving System Behavior Profiles for Pedestrian-Considerate Behavior Selection in Mixed-Traffic Environments: Empirical Calibration and Evaluation Using Naturalistic Driving Data

As autonomous driving systems (ADSs) become widespread, mixed-traffic environments will include ADSs, human-driven vehicles, and pedestrians. ADSs must select behaviors that reflect crossing-opportunity-related pedestrian consideration, distinct from physical collision avoidance, while balancing traffic efficiency and allowing ADS-specific tendencies in candidate behavior selection. This study develops a pedestrian-considerate behavior-selection model and ADS Behavior Profile (AB Profile) using naturalistic driving data. The model selects among passing, slowing, yielding, and continuing using pedestrian-consideration, traffic-efficiency, and vehicle-deceleration costs. Interaction events from the pedestrian intention estimation (PIE) dataset are divided into calibration and evaluation datasets. The Normal AB Profile is empirically calibrated, then evaluated using data not used for calibration or parameter selection. Calibration achieves Balanced Accuracy 0.761, Macro-F1 0.675, and Accuracy 0.750; evaluation yields 0.628, 0.630, and 0.705, with 31/44 events matching the behavioral reference. Balanced Accuracy and Accuracy are 0.804 and 0.800 for Crossing (pedestrian already crossing) versus 0.511 and 0.579 for Waiting (pedestrian waiting roadside). Ablation analysis shows distinct cost-component roles. Relative to the Normal AB Profile, the Aggressive AB Profile reduces delay and deceleration, whereas the Gentle AB Profile increases both. Sensitivity analysis clarifies profile-parameter roles. AB Profiles shape ADS-specific tendencies in situation-dependent candidate behavior selection rather than prescribing fixed behaviors.