基于频率感知混合Transformer的加注动作识别方法研究OA

In recent years,Transformer models have demonstrated significant advantages in modeling long-term dependencies within skeleton sequences,especially in the domain of skeleton action recogni-tion.However,current Transformer-based methods typically rely on simplistic attention mechanisms to capture spatial-temporal features,which limits their ability to learn discriminative representations for ac-tions with similar motion patterns.To address these challenges,this study introduces a frequency-aware mixed Transformer model for skeleton action recognition.This model is specifically designed to recognize similar skeleton actions with subtle discriminative movements by combining frequency-domain modeling with spatial-temporal attention mechanisms.By integrating high-frequency and low-frequency branch modules,the model enables multi-scale feature interaction,thereby enhancing its ability to distinguish similar actions.Comparative experiments between the proposed model and five mainstream models were conducted on a self-built skeleton action dataset containing seven types of standard operations.The results show that the frequency-aware mixed Transformer model achieves a recognition accuracy of 97.62%,out-performing other skeleton action recognition models.