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

This dataset contains synthetic video sequences and corresponding ground-truth 3D body-joint trajectories for validating joint-level pedestrian trajectory extraction methods. The dataset was generated in a controlled virtual environment and includes a single animated pedestrian walking along a straight path at three walking speeds: slow, normal, and fast. Each walking sequence was rendered from multiple virtual camera viewpoints with different camera distances, heights, and viewing angles. For each video sequence, the dataset provides frame-wise 3D body-joint coordinates in both world and camera coordinate systems, together with camera intrinsic and extrinsic parameters. The dataset enables systematic validation of vision-based methods for extracting 3D joint-level trajectories in pedestrian dynamics. By providing trajectories for multiple body joints, including the head, pelvis and feet, it supports the assessment of trajectory errors, gait event detection errors, and errors in step behavior measurements, such as step length, step width and step velocity. Beyond method validation, the synthetic dataset can be used to identify the most reliable camera positions for a given extraction method, the most appropriate viewpoints for specific body joints, and the best camera configurations for specific measurements. In addition, the synthetic dataset allows investigation into whether walking speed influences the accuracy and robustness of pedestrian dynamics measurements, and whether these effects differ across body joints, camera viewpoints, and extraction methods.