Related Experiment Video
Updated: Jul 11, 2026
05:41
A Step-by-Step Implementation of DeepBehavior, Deep Learning Toolbox for Automated Behavior Analysis
Published on: February 6, 2020
Deep learning approaches for head pose estimation in sports impacts
Thomas Aston1, Georgios Machtsiras2, Filipe Teixeira-Dias1
1Institute for Infrastructure and Environment (IIE), School of Engineering, The University of Edinburgh, Edinburgh, UK.
Sports Engineering
|May 25, 2026
Summary
Benchmarking deep learning models for head pose estimation in football headers, this study found full-body mesh recovery models superior for quantifying head acceleration events in sports. These models show promise for semi-automated videogrammetric analysis.
Area of Science:
- Biomechanics
- Computer Vision
- Sports Science
Background:
- Videogrammetry can quantify head acceleration in sports, but its accuracy with deep learning pose estimators is unclear due to limitations in standard datasets.
- Sports collisions involve large rotations, rapid motion, and frequent occlusion, challenging current pose estimation techniques.
Purpose of the Study:
- To benchmark three deep learning models for monocular head pose estimation during controlled football headers.
- To evaluate model accuracy under conditions mimicking sports collisions, including occlusion and dynamic movements.
Main Methods:
- Ten participants performed linear and rotational football headers.
- Synchronized 1000 Hz infrared motion capture provided ground-truth head orientations.
- Dual 50 Hz video cameras (frontal and side views) captured data for three models: direct head pose regressor, face reconstruction, and full-body mesh recovery.
Main Results:
- All models achieved low mean geodesic errors (4°-8°) and incremental geodesic errors (<4°).
- The SAM 3D full-body mesh recovery model demonstrated the lowest mean errors (4.59° and 1.99°) and better performance during occlusion and impact phases.
- Full-body approaches outperformed head-only estimators, especially under occlusion and dynamic conditions.
Conclusions:
- Modern full-body human mesh recovery models are effective for head pose estimation in sports, outperforming dedicated head pose estimators.
- Deep learning, particularly full-body mesh recovery, shows significant potential for semi-automated videogrammetric reconstruction of head acceleration events in sports.
- Model accuracy is influenced by factors like viewing angle, head rotation, and facial visibility.
