Badminton Action Classification Based on PDDRNet
Xianwei Zhou, Le Ruan, Song-Sen Yu, Jian Lai, Zheng-Feng LI, Weitao Chen · Atlantis highlights in social sciences, education and humanities/Atlantis Highlights in Social Sciences, Education and Humanities · 2023
Badminton is one of the most popular sports nowadays.To assist badminton teaching, a two-stage badminton movement classification method based on PDDRNet is proposed in this paper.In the first stage, the PDDRNet model for human pose estimation is trained using the knowledge distillation architecture of the teacher student network, the student network uses the lightweight model SECANet, while SimCC is simultaneously applied to replace the heatmap for representation.In the second stage, the estimated poses from the first stage are used for feature engineering, and XGBOOST is applied to classify the underlying badminton movements.In order to verify the performance of our proposed algorithm, we leverage the MPII datasets for human pose estimation experiments, and a proprietary badminton movement dataset for badminton movement classification.The results show that on the MPII dataset, it achieves a 3.1% improvement in PCKh when compared to lite-HRNet.In the second stage, the accuracy of the badminton movement classification algorithm using XGBOOST reaches 93.5% , which is 7.60% higher than the KNNbased badminton movement classification method.