<p>Automated classification of baseball pitch types is essential for advanced sports analytics, player evaluation, and injury prevention. Traditional methods primarily rely on ball trajectory data obtained from radar or optical tracking systems. However, biomechanical studies indicate that the pitcher’s body movements contain rich, and distinctive information valuable for pitch recognition. In this paper, we propose a skeleton-based approach that leverages the temporal and spatial dynamics of the pitcher’s motion extracted from broadcast MLB videos. Our method employs state-of-the-art pose estimation techniques combined with a YOLO-based pitcher detector to accurately isolate pitcher poses. We introduce a compact and biomechanically informed skeleton layout, TKA-11, focusing on torso, knees, and arms joints, which enhances spatial feature discriminability. To capture spatio-temporal dependencies, we design TKA-STAGCN, a Spatial-Temporal Graph Convolutional Network enhanced with Effective Squeeze-and-Excitation modules for temporal attention. We evaluated our approach on real-world broadcast footage, achieving 63.37% accuracy for the six-class pitch classification task, and 73.32% and 85.36% accuracy for the binary classification tasks of Fastball vs. Non-Fastball and Fast vs. Slow pitches, respectively, highlighting the model’s strong discriminative capability in binary settings. These results reflect the model’s ability to leverage body motion to effectively differentiate between pitch types across both coarse and fine-grained classification scenarios. Furthermore, our TKA-11 design yields an improvement of nearly 4% in accuracy compared to using full-body keypoints, demonstrating the benefit of biomechanically-informed joint selection.</p>