A synthetic dataset with ground-truth 3D joint trajectories for validating joint-level pedestrian trajectory extraction

Résumé fourni par la source

This study presents a synthetic dataset designed to support the validation of joint-level trajectory extraction for pedestrian dynamics measurements. Existing pedestrian datasets commonly provide pedestrian-level trajectories, but rarely include frame-wise 3D body-joint coordinates that can be used to evaluate joint-level trajectory extraction and microscopic step measurements. To address this gap, this study developed a dataset containing 147 rendered video sequences, comprising 18,669 video frames in total, of a single animated pedestrian walking along a straight path at three walking speeds. The same walking motions were rendered from 49 virtual cameras placed at different distances, heights, and viewing angles, enabling investigation of camera configuration effects. For each video, the dataset provides frame-wise ground-truth 3D body-joint coordinates in both the world and camera coordinate systems, together with camera intrinsic and extrinsic parameters. By combining rendered videos, known camera parameters, and frame-wise body-joint coordinates, this dataset provides a controlled benchmark for developing and validating vision-based methods for extracting 3D joint-level pedestrian trajectories and deriving step measurements from video. The controlled variation in camera configurations can provide guidance for camera placement in real-world pedestrian experiments.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.