Hybrid Deep Learning Models for Tennis Action Recognition: Enhancing Professional Training Through CNN-BiLSTM Integration

Abstract

Classifying tennis movements from video data presents significant challenges, including overfitting, limited datasets, low accuracy, and difficulty in capturing dynamic, real-world conditions such as variable lighting, camera angles, and complex player movements. Existing approaches lack robustness and practicality for real-time applications, which are crucial for sports analysts and coaches. To address these challenges, this paper proposes an advanced architecture that strategically integrates the Bidirectional Long Short-Term Memory Network (BiLSTM) and transfer learning from the lightweight Convolutional Neural Network (CNN) MobileNetV2. The motivation behind this work lies in enabling coaches to objectively analyze player performance and tailor training strategies based on precise movement recognition. The model is designed to enhance video representation capture, improve action classification accuracy, and operate efficiently in real-world conditions. Validation with the THETIS dataset demonstrates state-of-the-art results, achieving 96.72% accuracy and 96.97% recall, significantly outperforming existing methods. Additionally, the integration of cloud and edge computing capabilities facilitates real-time detection of tennis actions, providing immediate, actionable insights for practitioners. A motivating case study showcases how this method can effectively identify and analyze complex movements such as smashes and slices, addressing long-standing challenges in video-based tennis training. This research offers a robust and adaptable solution for classifying tennis actions, with promising implications for trainers and sports analysts seeking efficient and scalable tools for video analysis.