Last modified: 2026-05-31
Abstract
Pedestrian crossing behavior outside crosswalks is an important issue in urban traffic safety, especially for elderly pedestrians and other vulnerable road users. Previous studies have investigated differences between elderly and non-elderly pedestrians in crossing behavior outside crosswalks, including crossing time, vehicle gap judgment, and risk awareness. However, hesitation behavior before crossing has often been evaluated using observational or subjective criteria, and has not been sufficiently quantified using detailed skeletal posture features such as head orientation, trunk direction, foot orientation, and body center of mass movement.
This study proposes a video-based framework for analyzing pedestrian hesitation behavior before crossing outside crosswalks. The TITAN dataset released by Honda Research Institute USA is employed because it contains urban traffic scenes in Tokyo with pedestrian age-group labels and pedestrian action attributes. Age-group labels and action attributes provided in the dataset are used for initial analysis. Crossing phases and hesitation behavior types, such as stopping, looking left and right, tentative stepping, and restarting movement, are additionally annotated according to the purpose of this research.
MediaPipe Pose Landmarker is used to detect human body landmarks from images or videos and to obtain pose information for posture-based behavior analysis. The extracted features include walking speed, stopping duration, number of left-right checks, trunk direction, foot orientation, and body center-of-mass trajectory. These features are compared across age groups to identify age-related characteristics of hesitation behavior before crossing. The proposed framework is expected to provide objective and reproducible indicators of pedestrian hesitation behavior based on skeletal posture features. A composite hesitation index will be explored using these features.