Author(s): HONG, E., Institution: WASHINGTON UNIVERSITY IN ST. LOUIS, Country: UNITED STATES, Abstract-ID: 2079

INTRODUCTION:
Artificial intelligence has found widespread application across various domains, including sports education where machine learning (ML) techniques are applied to analyze players movements thereby offering valuable insights for enhanced performance.
We introduce a golf swing analysis system that harnesses the power of machine learning while addressing critical shortcomings commonly encountered in existing ML-based AI tools. In our work, a ML-model is deployed on a comprehensive dataset comprising diverse golf swings and discerns correlations between the trajectory of the golfers shoulders and the resulting ball flight. The acquired data serves as the foundation for predictive analytics to offer actionable insights into the quality of the swing. Notably, the rotational angles derived from the players motion serve as invaluable indicators of the golf swings dynamics.

METHODS:
Most state-of the art AI tools rely on ML-based pose detection to extract coordinates of crucial body points, but these tools often yield inaccurate results, particularly when analyzing movements of high-level players. This can be due to the fact that these tools are typically trained on datasets comprising predominantly of data from average individuals. For instance, professional players fully turn their shoulders up to 180 degrees during their swings, a rarity among non-professionals.
We can either re-train the tool with datasets containing instances of abnormal poses or selectively rectify inaccurate results based on body dynamics to address the problem. Our approach adopts the latter strategy and uses a formula to determine the horizontal rotation angle by comparing shoulder widths at setup and follow-through positions as follows.

shoulder_turn (°) = arccos (shoulder width at follow-through/shoulder width at setup position) x 180/π

In case that the tool fails to accurately detect shoulder positions, we identify the last correct shoulder coordinates during the swing and project the correct shoulder positions for the remainder of the swing based on body dynamics analysis.

RESULTS:
In our experimentation, player’s swings are analyzed by MediaPipe, a ML-based pose detection tool, and the subsequent ball trajectories are measured using TopTracer, a commercial ball tracing system. Their comparisons reveal a compelling correlation between the shoulder rotation angle and the flight path of the ball.
We also confirmed that MediaPipe generated occasional errors in pose analysis. Our post-processing tool automatically detected them and corrected such errors. The amount of error recovery ranges from 36° to 42° in shoulder rotational angle. This ensures that the data obtained is reliable and reflective of the true dynamics of the golfers swing.

CONCLUSION:
While expanding the dataset holds promise, its effectiveness in rectifying errors requires further investigation and validation. Also, exploring the impact of augmenting the dataset on error reduction can be a promising direction for future work.