Deep Learning For Human Activity Recognition Based On Wi-Fi Signals And Wireless Image

Human activity detection and recognition is one of the hot research directions in the field of Internet of Things,which has important theoretical and application value.Although the activity detection methods based on video and wearable sensors have achieved high accuracy,there are still strong limitations and cost problems.The method based on Wi-Fi signal breaks through the limitations of traditional methods.It can be used without light and occlusion,does not involve user privacy,and is low-cost and easy to deploy on a large scale.Therefore,it is widely used in the field of human activity recognition.Wi-Fi signal-based methods are divided into Received Signal Strength Indication(RSSI)methods and Channel State Information(CSI)methods.Compared with RSSI,CSI has more detailed information and can better describe human behavior,so that the recognition model can achieve higher accuracy.In this thesis,the research and analysis of the relevant performance and mechanism of the application of deep learning in human activity detection and recognition tasks based on CSI signals are carried out.Firstly,a human activity detection and recognition method based on the time domain attention mechanism of Long Short-Term Memory(LSTM)network is proposed.Using the powerful feature extraction ability of LSTM network,the feature extracted from the hidden layer is weighted in time domain by introducing attention mechanism,so that the corresponding feature weight increases when there is activity and decreases when there is no activity,which can better express the temporal change characteristics of CSI signal caused by human activity.This thesis verifies through experiments that the LSTM network after introducing the attention mechanism has better convergence on the WAR data set and higher accuracy,up to 97.5%,which is about 4.5% higher than the accuracy of the LSTM network alone.Secondly,an attention convolutional neural network classifier based on CSI wireless images is proposed.By converting the CSI signal into a wireless image,and using a down-sampling strategy to obtain more image samples,the generalization performance of the deep network is improved.Through experimental comparison of different convolutional neural networks,it is finally determined to use ResNet18 as the basic network model.Since the human activity information in most areas of the image is not obvious,this thesis introduces the Convolutional Block Attention Module(CBAM)in ResNet18 to better extract activity features from the two dimensions of space and channel.Experimental results show that the proposed method can achieve an accuracy of 96.9%on the WAR datasets,which ensures the accuracy and reduces the difficulty of data preprocessing and greatly reduces the model training time.Finally,in view of the problem that the currently published CSI datasets are generally small,leading to large generalization errors of deep networks,a deep migration learning method for human activity recognition based on wireless images is proposed.In this thesis,the ResNet18 network model parameters pre-trained on the ImageNet datasets with tens of millions of images are used as the initial values of the network model used in this thesis,and then the model is fine-tuned through training.Through experimental verification,after fine-tuning the network model,the accuracy of the model can be further improved,and the model converges faster.At the same time,in order to verify the generalization ability of the model,this article refers to the data collection method of the WAR data set,and conducts data collection in three different indoor environments.Through experimental verification,the model still has a high accuracy rate on the data sets collected by ourselves in different environments,all reaching about 98%,which verifies that the proposed model has good generalization ability.