Research On Human Action Recognition Algorithm Based On Channel State Information

In recent years,perception methods based on wireless signals have been gradually applied to context perception and human-computer interaction.Among them,the channel state information(CSI)in the Wi Fi signal has more fine-grained identification features.In this paper,the information recognition technology is used to study the four daily behaviors of walking,squatting,running and jumping,as well as the human behaviors of real falling and suspected falling.In order to extract high-dimensional spatial features from low-dimensional features and improve the recognition accuracy,this paper proposes a human action recognition algorithm(ATLSGCN)that combines graph convolutional neural network and bidirectional long-short-term memory network based on attention mechanism.In the offline stage,the complex matrix data of the amplitude and phase of the CSI is firstly converted into the image RGB format;secondly,it is input into the graph convolutional neural network with strong feature extraction capability for training and the feature information of human actions is extracted;finally,The extracted feature information is input into the bidirectional long-short-term memory network based on the attention mechanism for secondary training to obtain the human action recognition model AT-LSGCN.In the online stage,the CSI data converted in RGB format is used as the input of the AT-LSGCN action recognition model to classify and recognize four kinds of daily human actions.Through experimental verification,the recognition accuracy of the model in three environments of hall,conference room,and laboratory with different multipath effects is 95.97%,93.53% and 90.49%,respectively,and its performance is better than the comparison algorithm.In order to effectively discriminate between real and suspected fall actions and improve the accuracy of action recognition,this paper proposes a CSI human action recognition algorithm that integrates Hilbert-Huang transform and probabilistic neural network.In the training stage,firstly,the collected channel state information CSI is first subjected to amplitude denoising and random phase elimination,and the amplitude and phase fusion signal is used as the base signal.Secondly,the Hilbert-Huang transform is used for the base signal to extract the instantaneous amplitude and instantaneous frequency of human action information as classification features to construct a feature matrix,and finally a human action recognition model that can effectively detect real falls and suspected falls is trained in the probabilistic neural network optimized by genetic algorithm.In the testing phase,the trained recognition model is used to discriminate the fall action on the input CSI data.The experimental results show that the algorithm achieves the recognition accuracy of 97.18%,94.15% and 89.97% in the above three environments,respectively,which is higher than that of the current related algorithms.The paper has 39 figures,9 tables and 51 references.