融合注意力机制的多视图卷积手势识别研究OA北大核心CSTPCD

Gesture recognition based on surface Electromyography(sEMG)plays an important role in human-computer interactions.However,improving the accuracy of gesture recognition is a challenging task because of the nonlinearity and randomness of sEMG.To this end,this paper proposes a multiview convolutional gesture recognition model that incorporates an attention mechanism.First,a multiview input is constructed by extracting the classical feature set of the sEMG signal using a 200 ms sliding window.Second,Efficient Channel Attention(ECA)is used to weight the multiview features in the channel dimension,to strengthen effective features and weaken ineffective ones.Finally,multiview convolution is used to extract the high-dimensional myoelectric features with attention weights,thereby fusing them using the high-level feature fusion module to reduce data dimensionality and improve model robustness.The models were trained and evaluated on three public EMG datasets,namely NinaPro DB1,NinaPro DB5 and NinaPro DB7,obtaining an average recognition accuracy of 87.98%,94.97%,89.67%,respectively over a 200 ms sliding sampling window;the average voting accuracy for the entire gesture movement was 97.38%,98.41%,97.09%,respectively,and the average information transfer rate was 1308.71 bit/min.Compared with traditional machine learning methods and state-of-the-art deep gesture recognition methods that have been developed in recent years,the present model has higher recognition accuracy for both unimodal myoelectric and multi-modal gesture recognition,proving its effectiveness and generality.