Research On Human Behavior Recognition Based On Wi-Fi Signal In Indoor Scenes

In recent years,with the accelerating trend of aging,the home life and health monitoring of the elderly,especially the elderly living alone,have received more and more attention,and the corresponding behavior sensing technology research has also received more and more attention.Among them,wireless sensing technology based on Wi-Fi signals has become an emerging research direction for human behavior monitoring with the advantages of non-contact,low cost and high privacy protection,and has received wide attention from researchers.In the foreseeable future,this technology can be widely used in important fields such as elderly monitoring,security surveillance,and smart home.However,due to environmental interference and noise problems,the existing Wi-Fi signal-based behavior recognition models have problems such as low recognition accuracy due to single feature extraction method and insufficient feature relationship extraction,and low information utilization due to discarding phase information by using only the amplitude information in the signal,which affects the overall performance.In addition,due to the massive introduction of deep learning,the pursuit of high accuracy recognition performance inevitably causes the model structure to be bloated and the computation to be too large,which affects the fast migration under the limited resources of the scene equipment.To address the above issues,the main research of this thesis is as follows.(1)To address the problems of low model accuracy due to difficult feature extraction,low recognition accuracy,and insufficient feature extraction in the research of behavior perception based on Wi-Fi signals,this thesis proposes a hybrid neural network model AHNNet for human behavior recognition based on attention mechanism.The model extracts human behavior features from channel state information in a backward and forward temporal order through a bidirectional recurrent gating network.The model extracts human behavior features in CSI in a hierarchical manner using a temporal convolutional network.To further improve the recognition performance of the model,AHNNet uses a fused attention mechanism to enhance the main features of human behavior data to achieve the effect of improving the correct recognition rate.Through experimental comparison with other high-level deep learning models,it is demonstrated that AHNNet has good environmental adaptation ability and higher recognition accuracy.(2)To address the problems of insufficient information utilization and high model computational complexity in existing behavioral recognition methods based on deep learning and Wi-Fi signals,this thesis proposes a behavioral recognition model APFNet that fuses amplitude and phase information.the model first corrects phase error by phase error in CSI,then designs a lightweight multi-data fusion network for amplitude and phase information fusion,and finally the amplitude-phase fusion features are used for human behavior recognition.Comparative experiments demonstrate that APFNet not only adapts better to data with low sampling frequency compared with recognition schemes using only magnitude information,but also significantly reduces the computational complexity of the model while maintaining a high recognition rate,and can be deployed in resource-constrained devices in the future.