Dynamic Gesture Recognition Based on LSTM-CNN

The current research on using surface electromyography (sEMG) for gesture recognition mainly focuses on designing EMG signal features, decent feature designs can significantly improve the result. Nevertheless, the process of designing and selecting features can be complicated, as well as the precision of recognition for different features will be largely different even for the same model. Therefore, in this paper, we take advantage of the complementarity of Long Short-term Memory (LSTM) and Convolution Neural Networks (CNNs) by combining them into one unified architecture, which we call LSTM-CNN (LCNN). This model can directly input preprocessed EMG signal into the network for dynamic recognition of gestures. The LSTM model is used to extract timing information in signals. The CNN model can perform a secondary feature extraction and signal classification. In the experiment stage, the average recognition accuracy of LCNN can achieve 98.14%. As the experiment showed, LCNN model is feasible on dynamic gesture recognition based on sEMG signal.