Research On Technology Of Human Activity Recognition Based On Lightweight

With the continuous development and improvement of mobile communication and embedded technology,the integration of communication,computer,multimedia and other applications has been promoted,making people’s lives more smart.There are a large number of Internet of Things devices appearing in people’s lives.They provide various kinds of information and services for people’s life,study and work through a variety of human-computer interaction ways,so as to simplify people’s life style and save people’s time and energy.As an important way of human-computer interaction,human activity recognition technology based on Wi-Fi has gained wide attention in the field of the Internet of Things.However,most of the existing methods need to use large-scale datasets and deeper networks,ignoring the limitations of CPU computing power and memory resources of Internet of Things devices,making the current methods not applicable to Internet of Things devices.Therefore,for the practical application of human activity recognition technology based on Wi-Fi,it is great significance to study a lightweight,low-cost and applicable human activity recognition method for the ubiquitous Internet of Things devices.Based on the above issues,this thesis studies the following research:1.In this thesis,the human activity recognition technology based on SVM is introduced in detail,including data collection,preprocessing,feature extraction and classifier classification.On this basis,an aerial handwriting recognition system is implemented.2.To solve the problem of too many parameters in the network model,a lightweight temporal convolution network is proposed.A Wi-Fi based lightweight human activity recognition model is implemented based on TCN and residual structure in MobileNetV2.The experimental results show that the average accuracy of the model for five actions is 95.2%at 24 locations,but the computational cost is only 6%of the baseline model TCN.3.To solve the problem that the network model depends on largescale datasets and reduce the computing cost of Internet of Things devices in the data collection and processing phase,a lightweight complex-temporal convolution network is proposed.The network takes the original CSI data as input directly,and extracts the amplitude and phase information from the CSI at the same time,so the network can reduce the input data while obtaining the same amount of information.The experimental results show that the average accuracy of the model is 96.6%for five actions at 24 locations,and the performance is still outstanding at a small number of training samples,a small number of subcarriers and a low sampling rate.