Research And Application Of Wireless Localization Technology Based On Deep Learning

In recent years,with the development of wireless technology and the popularization of mobile Internet use,WiFi access points(APs)have become ubiquitous infrastructure in smart environments.Recent studies have found that WiFi signals are used in a variety of device-free human perception tasks in the form of channel state information,including indoor localization tasks.WiFi indoor localization technology has become one of the mainstream technologies in the research of indoor localization because it can make full use of the existing widely covered WLAN infrastructure and coordinate with the use of portable terminals such as smart phones for positioning,and it is low in cost,convenient and easy to operate,has the potential for large-scale commercial promotion.In this paper,we use the Nexus 5 mobile phone,which supported by the Nexmon project,to obtain CSI data.Compared with the widely used Intel 5300 network interface card,the CSI data that can be obtained has more subcarriers,which means that the resolution of the CSI data is higher.After obtaining the CSI data,this paper mainly processes the CSI data through the related technology of deep learning,so as to complete the corresponding indoor localization tasks.The main achievements achieved in this paper are as follows:1.By analyzing the physical properties of CSI feature map,an embedding method that conforms to the physical properties of CSI feature map is proposed,replacing the original Patch Embedding module in the Swin network,and this embedding method can improve the network’s recognition ability for CSI feature map.2.The MFA-ResNetViT network for indoor localization and activity recognition joint classification is designed and implemented.Using the construction idea of MobileViT network for reference,a fusion network ResNetViT is constructed by integrating the multi-head self attention mechanism into the convolutional neural network ResNet.A multi-scale feature aggregation module is used to aggregate the network specified layer outputs and serve as the input part of the activity recognition classifier.Experimental results show that compared to MobileViT network and ResNet network,MFA-ResNetViT network can improve the recognition accuracy of joint classification task;3.Design and implement a joint task of indoor reference points classification and track detection based on the CNN-Conformer network.The experimental results show that the CNN-Conformer network can process this joint task of track detection and reference points classification based on different reference points.