Abstract
Injury prediction in pitching depends on precise biomechanical signals, yet gold-standard measurements come from expensive, stadium-installed multi-camera systems that are unavailable outside professional venues. We present a monocular video pipeline that recovers 18 clinically relevant biomechanics metrics from broadcast footage, positioning pose-derived kinematics as a scalable source for injury-risk modeling. Built on DreamPose3D, our approach introduces a drift-controlled global lifting module that recovers pelvis trajectory via velocity-based parameterization and sliding-window inference, lifting pelvis-rooted poses into global space. To address motion blur, compression artifacts, and extreme pitching poses, we incorporate a kinematics refinement pipeline with bone-length constraints, joint-limited inverse kinematics, smoothing, and symmetry constraints to ensure temporally stable and physically plausible kinematics. On 13 professional pitchers (156 paired pitches), 16/18 metrics achieve sub-degree agreement (MAE < 1°). Using these metrics for injury prediction, an automated screening model achieves AUC 0.811 for Tommy John surgery and 0.825 for significant arm injuries on 7,348 pitchers. The resulting pose-derived metrics support scalable injury-risk screening, establishing monocular broadcast video as a viable alternative to stadium-scale motion capture for biomechanics.
Introduction
In professional baseball, biomechanics is the primary tool for injury prevention. Metrics such as hip-shoulder separation and trunk tilt are critical predictors of arm stress. However, capturing this data traditionally requires optical tracking systems costing upwards of $500,000, creating a significant "biomechanics gap" between Major League Baseball (MLB) teams and amateur or international programs. Previous monocular video attempts struggled with the noise of broadcast footage—motion blur, compression, and self-occlusion—making raw 3D pose estimation unreliable for clinical use. This work bridges that gap by transforming standard broadcast video into stable, physically plausible kinematics.
Methodology
The proposed system utilizes a multi-stage pipeline designed to convert "jittery" video poses into professional-grade biomechanical data.
- DreamPose3D & PGLM: The process begins with 3D pose estimation, enhanced by a novel Pelvis-to-Global Lifting Module (PGLM). Unlike standard pelvis-rooted estimators, PGLM uses a transformer-based architecture to estimate the global trajectory of the pelvis, effectively positioning the pitcher in 3D space.
- Biomechanics Refinement Stack (BRS): To correct markerless tracking errors, the BRS applies bone-length enforcement, constrained Inverse Kinematics (IK) using pitcher-specific joint limits, and Savitzky-Golay filtering for smoothing. This ensures the resulting motion is anatomically possible and free of high-frequency jitter.
- Injury-Risk Modeling: The pipeline extracts 18 core metrics, expanded into 114 features including workload and medical history. An ensemble of Gradient Boosting Machines and Random Forests is then used to identify high-risk pitchers.
Experimental Validation
The system was validated using three data tiers: a validation set of 156 pitches with gold-standard tracking, a deployment set of nearly 120,000 professional sequences, and an injury screening set of 7,348 pitchers.
As shown in the validation results, 16 out of 18 biomechanical metrics achieved a Mean Absolute Error (MAE) of less than 1°. Lower-body angles and elbow flexion were particularly precise, with MAEs under 0.5°. While shoulder abduction remains a challenge due to occlusion (MAE 6.2°), the overall accuracy establishes monocular video as a viable alternative to stadium-scale systems.
Key Insights and Results
A critical finding of this research is that biomechanical variability—the pitch-to-pitch change in mechanics—is a stronger predictor of injury than average values. Features related to Range and Coefficient of Variation (CV) accounted for nearly 40% of the model's predictive importance. Additionally, "P90" statistics (90th percentile stress metrics) proved highly predictive, suggesting that injury risk is driven by extreme mechanical events during peak stress.
The final automated screening model achieved an AUC of 0.811 for predicting Tommy John surgery and 0.825 for significant arm injuries, significantly outperforming baseline models using static thresholds or simple anomaly detection.
Conclusion
This work demonstrates that physics-based post-processing and structural constraints can overcome the limitations of raw neural network outputs in high-speed sports analysis. While shoulder abduction accuracy requires further refinement through multi-view integration, the current pipeline enables large-scale, cost-effective injury screening. Future efforts will focus on pitcher-specific baselines to detect individual mechanical deviations in real-time.
Summary updated