The Application of Action Recognition Based on MPP-YOLOv3 Algorithm in Posture Correction

Abstract

Posture recognition, as a research hotspot, has been widely applied. A recognition model based on bone key point detection is proposed for the posture correction application module. Firstly, the lightweight You Only Look Once v3 Tiny network was chosen as the infrastructure, and the OpenPose algorithm in the bottom-up strategy was chosen to implement posture recognition. To reduce the computational burden of the model, the Media Pipe Blaze Pose algorithm was introduced for improvement. At the same time, by refining more bone key points, the accuracy of the model has been improved. The experiment outcomes said the recognition accuracy of Cross View in the NTU60 RGB+D dataset and NTU120 RGB+D dataset was 94.7% and 82.7%, respectively. Compared to graph Transformer networks and semantic posture recognition models, the Cross Subject metric has improved by an average of 3.5%. Therefore, the research and design model has shown better robustness in the field of posture recognition, which can help complete pose correction more efficiently.