Pose-Based Action Recognition in Tennis using MediaPipe and LSTM

Pose recognition in tennis is an essential aspect for analyzing playing techniques and evaluating athlete performance. This study develops a tennis pose recognition system that integrates MediaPipe for pose feature extraction with Long Short-Term Memory (LSTM) networks for movement classification. The research dataset consists of 2,010 images of tennis movements across four categories: backhand, forehand, ready position, and serve, annotated in COCO format. MediaPipe successfully extracted pose landmarks from 1,782 images (88.7%), generating 33 pose landmarks flattened into a 99-dimensional feature vector. The LSTM model is designed with a 3-layer LSTM architecture and 2 dense layers, trained using a stratified train-test split with an 80:20 ratio. Model evaluation uses various metrics including accuracy, precision, recall, and F1-score. The results show that the system achieves 90.20% accuracy, with the best performance in the ready position category (F1-score: 91.28%) and the lowest in the forehand category (F1-score: 88.89%). The model demonstrates good computational efficiency with a memory footprint of 714.39 KB, enabling deployment on mobile devices. This study contributes to the development of automated sports analysis systems and demonstrates the feasibility of integrating MediaPipe-LSTM for real-time tennis pose recognition applications.