Self-attention networks for motion posture recognition based on data fusion
Zhihao Ji, Qiang Xie · 2022
Aiming at the problems of error and poor recognition effect in the data collected by sensors in current motion posture recognition, this paper proposes a self-attention network for motion posture recognition based on data fusion to solve these problems. This method uses the attitude angle information output by the gyroscope to correct the attitude angle obtained by the acceleration sensor using kalman filtering, which effectively improves the accuracy of the attitude angle; at the same time, the attitude angle and acceleration sensor data are used to construct an attention convolutional neural long short term memory artificial neural network (CNN-LSTM) of the attention mechanism to recognize the motion state. The experimental results show that the use of data fusion method can correspond to the accuracy of physical signs, and compared with the traditional network, the accuracy of the network frame recognition proposed in this paper is improved