Real-Time Badminton Action Recognition based on Media Pipe and Motion Bidirectional Encoder Representation Transformer

In recent years, accurate recognition of athletic actions in sports is crucial for performance analysis, training feedback, and intelligent coaching systems. This research proposes a lightweight and effective methodology for recognizing badminton actions using pose-based video analysis, by incorporating Motion Bidirectional Encoder Representation Transformer (MotionBERT). Initially, the dataset is collected from real-world badminton videos which includes realistic and complex motion patterns. Further, the collected data is preprocessed by employing uniform frame sampling and temporal noise filtering to preserve motion continuity and reduce estimation. Furthermore, human pose detection and estimation are performed using Media Pipe that assist to provide accurate 3Dimensional joint coordinates robustly from video frames. Thus, these pose sequences are then classified using MotionBERT transformer-based model that comprehends human motion through attention mechanisms. Finally, the proposed MotionBERT demonstrates the significance of integrating real-time pose estimation with transformer-based recognition to enable accurate and scalable action recognition in sports environments.