Passenger terminals are evolving into complex nodes where land and sea transport intersect. However, the safety and convenience of pedestrian environments in access roads are often compromised owing to the development of surrounding commercial areas and inadequate safety facilities. Existing safety assessments for pedestrian environments have primarily relied on qualitative or post-incident analyses centered on static structures, possessing fundamental limitations in quantifying real-time risks posed by dynamic obstruction elements. To address these issues, in this study, we propose an AI-based pedestrian obstruction analysis framework. To ensure high-fidelity data acquisition for safety assessment, we utilized optical sensors (action cameras) as wearable sensing units to capture real-time pedestrian dynamics in complex terminal environments. We constructed a dataset of pedestrian obstruction objects based on first-person walking videos recorded with an action camera worn by a pedestrian along the access roads of the Mokpo Port Passenger Terminal. In particular, we adopted the Swin Transformer architecture as the backbone for the Mask R-CNN instance segmentation model in order to leverage its previously reported strengths in multiscale object recognition and generalization in complex scenes. Furthermore, we developed the obstruction rate (OBR) measurement algorithm, which utilizes pixel-level mask information of identified objects to calculate their occupancy within designated walking areas. The OBR algorithm was applied to two distinct zones near the terminal, capturing structural differences between sidewalks and mixed pedestrian–vehicle areas. The resulting zone-wise OBR distributions provide a quantitative basis for comparing pedestrian safety conditions and identifying high-risk segments along terminal access routes. In this study, we demonstrate the feasibility of AI-based pedestrian obstruction analysis as a quantitative safety assessment tool and suggest future directions for its integration into real-time monitoring systems and policy decision-making.