Page 72 - 2024S

UEC Int’l Mini-Conference No.52 65

Skeleton-based Action Classification in Baseball

2
2
Sergio HUESCA ∗1, 3 , Hiroki TAKAHASHI , Gibran BENITEZ , and Mariko NAKANO 3
1 UEC Exchange Study Program (JUSST Program)
2 Deparment of Informatics
The University of Electro-Communications, Tokyo, Japan
3
Instituto Polit´ecnico Nacional, Mexico City, M´exico

Keywords: Skeleton-based Action Classification, Baseball Analytics, Deep Learning, Spatial-Temporal
Graph Convolution Networks (ST-GCN), Human Pose Estimation.

Abstract

This research presents the implementation of a skeleton-based deep learning model to classify actions
within baseball games from various perspectives. Leveraging the MLB-Youtube dataset, which includes
video clips annotated with play outcomes and pitch types, the study employs Human Pose Estimation
keypoints for the pitcher, catcher, batter, and umpire using OpenPose. Implementing a deep learning
model based on Spatial-Temporal Graph Convolution Networks (ST-GCN), we accurately model the
spatial and temporal relationships within the game, enabling precise classification of play outcomes.
This approach facilitates the automatic generation of actionable statistics, significantly contributing
to the improvement of player performance and umpire decision-making. Additionally, the research
proposes methodologies for detecting pitch types, estimating the strike-zone, and classifying “ball” or
“strike” calls, thereby enhancing baseball analytics. The findings are anticipated to provide valuable
insights and tools to improve performance and decision-making in baseball games.

∗
The author is supported by JASSO Scholarship.