PE3DNet: A “Pull-Up” Action Quality Assessment Network Based on Fusion of RGB Image Data and Key Point Optical Flow

“Pull-up” is a regular physical fitness test in high school and college physical education, and the traditional manual supervised assessment method has the problems of strong subjectivity, large error and low efficiency in the assessment and training of “Pull-up”. In order to solve these problems, this paper proposes an action quality assessment network PE3DNet for “Pull-up” video streams, which uses OpenPifPaf to generate key points, constructs key points optical flow data using key points, and adopts P3D residual blocks to construct a dual-stream network that fuses RGB image data and key point optical flow to achieve the action quality assessment of “Pull-up”. Through the validation experiments on the self-made dataset SKD-PULL, the results show that PE3DNet achieves 92.8% Accuracy, 92.9% Precision, 91.3% Recall, and 92.1% F1 score in the standardized action of “Pull-up”, which effectively improves the accuracy of the action quality assessment of “Pull-up” program and brings higher efficiency and fairness to the sports testing process.