Research And Implementation Of Indoor Human Perception Based On WiFi Channel State Information

In the era of intelligent information,indoor human perception technology is developing rapidly.While the needs of a large number of indoor human perception applications spring up,they also constantly raise higher requirements.As the sensor devices cannot meet reliability,safety,practicability and universality at the same time,the research gradually turns to indoor sensing based on WiFi signals.WiFi devices are low-cost,safe and reliable,and have already been deployed in large quantities indoors.As common network transmission equipments,WiFi devices can also be used as indoor sensing signal sources.WiFi received signal strength indication(RSSI),as a routine sensing data source in the network link layer,has been gradually eliminated due to its vulnerability to multipath effects and data instability.Channel state information(CSI)is a more sophisticated sensing data source in network physical layer,which can resist certain multipath interference and provide richer indoor environment changing characteristics.As CSI data can be accessed from commercial WiFi devices,the focus of WiFi-based perception research has shifted to CSI.In this thesis,CSI data are used to perceive indoor humans,focusing on four main applications:people counting,identity recognition,location tracking and activity recognition.Perception content can satisfy most indoor applications,answering four questions:how many people,who are they,where are they and what are they doing.To count the people,the Doppler frequency shift feature is proposed to identify the number of people passing through the entrance,and the rotation segmentation algorithm is proposed to solve the problem of continuous passing.In terms of identity recognition,the statistical features of the CSI data are excavated,and the Isolation Forest is used for the authentication problem,which is to identify unknown and known identities,and to classify the specific known identities by using support vector machine(SVM).In location tracking,a deep neural network(DNN)regression is exploited to estimate the target location,while particle filtering and map matching are combined to optimize the location results.In activity recognition,the recurrent neural network(RNN)is applied to recognize human activities,and the recognition effects of three different scenarios are compared:line of sight,non-line of sight and through-the-wall.The experimental results show that the accuracy is 95.6%for the proposed people counting algorithm in the case of discontinuous traffic.If the pedestrians separated by more than 8 meters,the continuous traff-ic accuracy is 90%.In the experiment of identity recognition,the recognition accuracy of identity authentication is more than 93.3%for only one legitimate user and 88.9%for two legitimate users.The recognition accuracy of classification is up to 96%for 6 different persons.In the location tracking problem,experiments are carried out in our laboratory and a hall respectively.The positioning accuracy of both scenes is 0.54 meters.For activity recognition,6 common activities are identified.The recognition accuracy is 98.3%for line of sight,96.6%for non-line of sight and 95%for through-the-wall scenes.Experiments show that the proposed method can effectively solve the indoor human perception problem,and the evaluation results outperform the state of the art.