Deep learning approaches for head pose estimation in sports impacts

Abstract Videogrammetry can quantify head acceleration events in sport, but because standard datasets lack the large rotations, rapid motion, and frequent occlusion characteristic of sports collisions, the accuracy of modern deep learning pose estimators in this context remains unclear. This study addresses this gap by benchmarking three models for monocular head pose estimation during controlled football headers: a direct head pose regressor, an end-to-end face reconstruction model, and a full-body human mesh recovery model. Ten participants performed linear and rotational headers. Synchronised 1000 Hz infrared motion capture provided ground-truth orientations, while dual 50 Hz video cameras supplied frontal and side views. Model outputs from standardised detections were temporally smoothed and evaluated using geodesic and incremental geodesic error metrics. All models achieved single-digit mean geodesic (4°–8°) and incremental geodesic (<4°) errors. SAM 3D yielded the lowest mean errors (4.59° and 1.99°, respectively) and showed lower sensitivity to occlusion and temporal impact phase. By uniquely comparing head-only and full-body approaches, results demonstrate that modern full-body human mesh recovery models outperform dedicated head pose estimators under the heavy occlusion and dynamic conditions typical of sports collisions. Errors increased for side-view footage, rotational trials, and low facial visibility. These findings support using deep learning, particularly full-body mesh recovery, for semi-automated videogrammetric reconstruction of head acceleration events.