Lightweight OpenPose with MobileNet v1 backbone and ResNet-50 for badminton technical action recognition in teaching

Abstract

The rapid development of computer technology and increasing focus on physical fitness have spurred the application of deep learning in sports training. To address the low accuracy and high computational cost of badminton technical action recognition methods, this study proposes a lightweight teaching movement recognition system combining MobileNet v1 optimized OpenPose and ResNet-50. The system employs a self-collected dataset of 4216 samples across four action categories (preparation, trigger, hitting, and swing), with a 70/20/10 training-validation-test split and cross-validation for robust evaluation. Compared with baseline systems including improved SVM, VGG-19-based OpenPose, and improved SlowFast, the system achieves an identification speed of 8.82 fps and a recognition accuracy of 97.9%, and average F1 score of 98.2%, while maintaining the lowest CPU occupancy at 24.8%. The system precisely identifies and labels badminton technical actions, providing a reference for teaching and training.