Research on Table Tennis Swing Recognition Based on Lightweight OpenPose

Aiming to address the issues of irregular movements and inconsistent teacher standards in the teaching process of table tennis swing, the paper proposes a table tennis swing movement recognition system based on an improved Lightweight OpenPose algorithm. The videos or pictures taken by the practitioners are preprocessed to extract the frame images of the key movements and reconstruct the human skeleton. The Lightweight OpenPose model was used in place of the OpenPose model, and MobileNet v3-small was utilized instead of the MobileNet v1 network within the Lightweight OpenPose model. Additionally, the InceptionTime time series classification network is utilized to perform action classification and compare it with the standard action, thereby correcting the learner's posture. The improved algorithm achieves a recognition accuracy of over 92.25% for table tennis swinging behavior, surpassing the accuracy of the original model. Furthermore, following the implementation of the lightweight modifications, the model becomes smaller in size and possesses fewer parameters, leading to a significant improvement in the computational speed of the model.