Research on User Action Recognition Method Based on parallel CNN-BiLSTM neural network

Abstract

With the maturity of artificial intelligence algorithms, the recognition of human actions is gradually used to detect human body conditions and other directions. Human activities and the movement of objects change the multipath characteristics of the wireless channel, leading to changes in the channel state information (CSI). Aiming at the problems of poor flexibility and low accuracy of traditional user action recognition methods, a user action recognition method based on parallel convolutional neural network-bidirectional long short-term memory network (CNN-BiLSTM) is proposed. This method uses discrete wavelet transform to filter out the noise in the CSI action data, then extracts the amplitude characteristics of the CSI data, and finally uses a deep learning network to recognize the action. CNN can learn the spatial feature information of actions, while the BiLSTM network can learn information with complex temporal features. In the experiment, 7 actions of 5 people were collected and evaluated, including falling, picking up, walking, sitting, etc. By extracting the features of these actions and using the Softmax function to classify, the effect of user action recognition is achieved, with an average accuracy rate of 98.7%. The results are compared with a single CNN and a singal BiLSTM network respectively, indicating the parallel CNN-BiLSTM model has high classification accuracy, strong learning ability, and robustness under different users and multiple action categories.