Enterprise AI Analysis

Our Robust Data-Driven Methodology

Our analysis pipeline begins with collecting 321 driver swing sequences from six amateur golfers in a controlled studio environment. Synchronized monocular swing videos and TrackMan-derived ball trajectory measurements form the foundation of our dataset. We then preprocess the data to extract both human pose and golf club keypoints.

Keypoint Extraction and Graph Construction

We leverage YOLO11m for robust object detection of the club-head and club-grip, and YOLO11m-pose for human body keypoints. These 15 keypoints (e.g., Left-Shoulder, Right-Wrist, Club-Head) are then integrated into a unified spatial-temporal graph. This graph effectively models both the golfer's kinematics and their interaction with the club, which is crucial for predicting outcomes like Spin Axis, Launch Direction, and Ball Speed.

To ensure high data quality, we apply linear interpolation and Savitzky-Golay filtering to smooth keypoint coordinates, addressing issues like occlusion and motion blur inherent in video data.

Superior Performance with Graph Neural Networks

Our experiments demonstrate that graph-based models, especially STGAT (Spatial Temporal Graph Attention Network), significantly outperform traditional machine learning methods and even ST-GCN in predicting golf ball flight outcomes. This underscores the power of modeling complex spatial-temporal relationships in swing mechanics.

Impact of Club Keypoints: Our ablation study further validates the critical role of integrating club keypoints. Including club dynamics consistently improved prediction accuracy and regression performance, particularly for Ball Speed, demonstrating that a holistic representation of body and equipment interaction is essential for modeling complex sports movements.

This comprehensive modeling allows our system to capture subtle variations in club-body coordination that significantly influence impact conditions and ball flight characteristics, leading to more accurate and reliable predictions.

Actionable Insights Through Explainable AI

A core strength of our framework is its interpretability, achieved through Integrated Gradients (IGs) with phase-specific baselines. This technique quantifies the importance of both body and club keypoints across eight distinct swing phases, enabling granular, biomechanically consistent feedback.

By comparing optimal and suboptimal swings, we can pinpoint exactly which keypoints (e.g., Left-Shoulder, Club-Head) and during which phases (e.g., backswing, impact) contribute most significantly to ball flight outcomes. For instance, a "worst swing" might show disproportionate reliance on the downswing and little contribution at impact, indicating disrupted body rotation.

Personalized Feedback: Our analysis reveals that even small modifications to influential keypoints, like the club-head or upper-limb joints, can lead to measurable changes in ball speed and directional control. This enables coaches to provide highly targeted, phase-specific corrective feedback, moving beyond general advice to data-driven, individualized training strategies.