LSTM-Based Badminton Stroke Classification: Comparing OpenPose Keypoint Formats under Event-and-Camera Shift
Abstract
: Badminton is one of the most popular sports globally, yet coaches and sports analysts still rely on conventional methods to evaluate players from match videos. As a first step toward a comprehensive badminton analytics platform, this paper proposes an end-to-end system to classify four badminton strokes (serve, smash, backhand, and lift) using Human Pose Estimation with OpenPose combined with a Long Short-Term Memory (LSTM) classifier. Skeleton keypoints were extracted in both COCO (18 keypoints) and BODY_25 (25 keypoints) formats from broadcast match videos, grouped into 25-frame sequences, and fed into the LSTM model. On same-event test data, both keypoint formats reached 94.73% accuracy, with macro-averaged F1 scores of 94.17% for COCO and 94.03% for BODY_25. Under external validation on footage from a different tournament, recorded with a higher and more distant camera, BODY_25 retained 91.47% accuracy and 90.80% F1, whereas COCO fell to 81.39% accuracy and 79.90% F1. The results are comparable to those reported by prior deep learning approaches, while using only broadcast video without wearable sensors or controlled laboratory conditions. BODY_25 is therefore markedly more robust to this combined event-and-camera shift, with a 3-percentage-point drop compared to over 13 points for COCO, although the study design does not allow the effect of camera placement to be separated from the other differences between the two events.