Research On Wireless Sensing Technology Based On Environment Adaptive

In recent years,the development in artificial intelligence,sensor and wireless communication technologies is getting accelerated.Such fast-paced improvements have led to the increasing attention in technologies such as context awareness,behavior recognition and new human-computer interaction.The identity recognition,as an important part of the wireless recognition,primary focuses on the human identity.In the recognition process,the machine is able to recognize the subject's identity in a specific monitoring environment.Compared with other identification technologies,Wi-Fi based identification technology enables users to interact with the intelligent environment in a more convenient way without device dependence due to its advantages of low cost and easy deployment.However,as the target population to be identified in the actual scenario is usually dominated by a large base group,the accuracy of the model will decrease when the environment changes.In order to better adapt to people's psychological feelings,it is particularly important to seek a more comfortable and zero-burden identification method.We make full use of the periodic disturbance of human walking gait behavior to Wi-Fi channel state to obtain all gait feature information received by different antennas.According to the reconstructed CSI data format--frequency energy graph,a compression mechanism is designed to lightweight neural network.Moreover,combined with transfer learning,we build a light weight transfer model which can automatically extract human gait feature and quickly adapt to a variety of different hardware and acquisition environment.Our main research contains are described as follows:(1)For the existing Wi-Fi based gait identity recognition model,traditional manual feature extraction face challenges such as insufficient extraction of gait features and low accuracy of multi-person gait recognition.To solve the problems,we propose a new data analysis model--frequency energy graph.Then we introduce frequency energy graph to neural network model to realize multi-person identification with high accuracy in Wi-Fi environment.Different from traditional gait feature extraction methods which only extract temporal feature,our method not only extracts temporal feature of the gait,but also extracts spatial feature among subcarriers,greatly improving identification ability.The experimental results show that in an open environment,recognition accuracy of our model reaches 98.7% with 40 people,which shows good robustness.(2)Existing gait identification model has problems such as high storage cost,large parameter scale,difficult to adjust,and difficult to deploy on limited hardware platforms.In order to work out the above problems,we learn from the principle of feature extraction by convolution network and design a Balloon mechanism.The Balloon mechanism reduce the network's dependence on the number of neurons in each layer by transferring original data features to different network layers,so as to realize neurons clipping.Based on Balloon mechanism,we propose a light weight gait recognition algorithm named LWID.By channel information integration and convolution operation,the algorithm successfully compresses model scale.Experimental results show that LWID has higher accuracy and better robustness with less parameters and shorter training time compared with the existing gait identity recognition models.(3)The variety of multipath effects of Wi-Fi signal in different environment is complicated and indistinguishable.As a result,models trained in one environment face the risk of failure in the new environment,and a model needs to be retrained when the task changes.Based on transfer learning,we design a model calibration method that does not require a large amount of labeled data.By this method,a model being completely trained in an environment adapts quickly to the new environment by fine tuning a small amount of new data.Moreover,time that is required to retrain the network due to changing tasks is greatly reduced.So the proposed method realizes rapid calibration of the model across time,space and mobile devices.