Since 2009, pedestrian fatalities caused by traffic accidents in the United States have increased by 78%, exceeding 7,000 deaths annually. To reduce injuries, a deeper understanding of injury mechanisms based on real-world accident data is required. Recently, accident reconstruction techniques utilizing recorded video from vehicle-mounted cameras have gained increasing attention in accident analysis. However, injuries resulting from contact with the road surface are often occluded due to camera blind spots, hindering accurate capture. Additionally, the low frame rate of many vehicle-mounted cameras limits the ability to accurately analyze the detailed kinematics of pedestrian impacts.
To address these limitations, this study proposes a reconstruction method that estimates pedestrian motion in occluded regions by matching observed behavior from vehicle-mounted camera recorded video with simulation results.
First, to reconstruct three-dimensional pedestrian motion from vehicle-mounted camera video, two-dimensional skeletal keypoints were estimated using MMPose applied to video. To integrate depth information, DepthAnythingV2, a monocular depth estimation model, was applied to estimate 3D keypoints from the same video. Due to initially low depth estimation accuracy, DepthAnythingV2 was pre-trained using stereo videos simulating pedestrian motion. This pre-training improved model performance and enabled more reliable 3D pose reconstruction.
Next, to evaluate the matching accuracy between camera-derived and simulation-based pedestrian motion, a novel metric was introduced. This metric incorporates the displacement of the center of mass in the chest region and the directional similarity of chest-facing vectors in the forward and downward directions, quantified using cosine similarity. These features are normalized and weighted to compute a composite score for motion matching.
By integrating image analysis and simulation technologies, the proposed method enables more accurate reconstruction of pedestrian-vehicle collisions. It also contributes to the advancement of post-crash rescue systems, such as Advanced Automatic Collision Notification (AACN), ultimately supporting efforts to reduce pedestrian fatality rates.