Pedestrian fatalities in traffic accidents continue to rise, with severe injuries often resulting from both vehicle impact and subsequent ground contact, frequently occurring outside the field of view of vehicle-mounted cameras.
This study presents a proof-of-concept (PoC) approach for reconstructing three-dimensional pedestrian motion—including occluded regions—using dashcam video. The method integrates 2D human pose estimation (MMPose) and monocular depth estimation (Depth Anything V2),the latter was fine-tuned on a custom dataset, to generate 3D skeletal coordinates.To evaluate motion matching, the reconstructed pedestrian poses were quantitatively compared with a database of vehicle collision simulations using the THUMS human body model and skeletal data representing real-world crash scenarios generated in PC-Crash. Composite similarity indices based on thoracic center of gravity trajectory and torso orientation vectors were employed for this comparison.
Preliminary results indicate that the fine-tuned system achieves an average RMSE of approximately 0.1 m for key skeletal points, enabling accurate depth estimation for 3D pose reconstruction. Matching experiments with 11 PC-Crash cases demonstrated high similarity scores, and reconstructed sequences successfully identified critical injury events such as head-to-ground contact in occluded regions, confirming the feasibility of this approach for accident reconstruction and injury risk assessment.
However, this study remains preliminary, limited to controlled indoor experiments with a single vehicle type and few subjects. Real-world crash footage and diverse vehicle geometries were not considered, and skeletal reconstruction from actual accident videos has not yet been implemented. Future work will expand the simulation dataset, refine similarity weighting, and validate the approach using real crash video. Ultimately, this technology may support forensic analysis and emergency response, but further validation is required before real-world application.