Personalizing AI-Based Safety Perception Assessment for Urban Streets Using Personality Traits

Abstract

We present an interpretable human-centered AI framework that incorporates Big Five personality traits to model perceived street safety at both scene and individual levels, addressing the limitations of conventional "one-size-fits-all" safety scoring. Our dataset includes 20 daytime and nighttime street images from Nagoya’s Shinsakae district. A pairwise comparison experiment with 69 participants yielded 13,110 individual safety judgments, and personality traits were measured using the BFI-10 scale. We used a Mask2Former model pre-trained on ADE20K to extract six CPTED-relevant visual features. We first constructed a baseline linear model using averaged human responses, then extended it by adding personality traits as moderators that adjust visual cue weights. Results show greenery and pedestrians enhance perceived safety, while graffiti, litter, and vehicle-dominated streets reduce it. Notably, personality systematically modulates sensitivity to these cues, and our personalized model consistently outperforms the baseline. This framework has implications for explainable urban AI, personalized navigation, and inclusive city design aligned with SDG 11.