Research On Human Activity Recognition Method To Eliminate Position Dependence

With the development of pervasive computing and the popularity of Wi-Fi devices,human action recognition based on Wi-Fi has been applied in security monitoring,smart home,humancomputer interaction and other fields.Due to the advantages of low cost,easy deployment,wide coverage,the ability to penetrate walls,not affected by light and not easy to leak privacy,human action recognition schemes based on channel state information in Wi-Fi signals are favored by more and more researchers.However,it is found that there are inconsistent channel state information expressions in the same category of actions with different positions and orientations,that is,the human action recognition work has the problem of location dependence,which leads to the degradation of recognition performance.In this paper,we try to use the neural network model to extract more effective features from CSI data and reduce the generalization error of action recognition accuracy,thus eliminating the location dependence effect.The research content and contributions are as follows:(1)Research on link selection and action interval division algorithm.Aiming at the difference of sensitivity of each antenna link to different types of actions,this paper firstly selects the amplitude measurements that are easier to obtain and more representative in CSI as the main analysis object of human action recognition,and uses the combination of outlier removal and S-G filtering for data preprocessing.An antenna link algorithm based on average variance was proposed to select the antenna link that was more sensitive to the target motion.Aiming at the problem of original data redundancy,this paper proposes an action region division method to determine the initial interval of the action,obtain more concise and effective data,reduce the amount of data processing and improve the recognition accuracy.(2)Research on autoencoder network model for eliminating position dependence.Aiming at the problem of position dependence in the process of human action recognition,this paper proposes an LSTM-CNN autoencoder network model.The encoder part of the autoencoder network adopts the LSTM network structure,and uses its good ability to learn time series information to convert the original time series signal into a fixed-length feature vector,and then non-linearly transform it into a feature image to extract key information.The decoder part of the autoencoder network adopts the CNN network structure,which can effectively capture the local features and spatial distribution of the image,has better feature extraction and classification performance,improves the recognition accuracy and robustness,and effectively eliminates the influence of position dependence.(3)Human action recognition experiment and result analysis.In order to highlight the influence of position dependence in action recognition,this paper sets up action categories with different positions and orientations,builds an experimental platform and collects CSI data,and constructs an action dataset that can fully reflect the position dependence.In order to verify the effectiveness of the proposed method,the performance of the antenna link selection and action interval division algorithms are evaluated.It is determined that the proposed antenna link selection method based on average variance is effective and superior to the other two common data screening schemes.The effect of LSTM-CNN autoencoder model and other models in eliminating position dependence is compared and evaluated.Experiments show that the model proposed in this study can effectively recognize the same action in different positions and orientations,and the average recognition accuracy can reach 95%,which is higher than the average accuracy of machine learning and other deep learning models,and effectively eliminates the problem of location dependence.