Font Size: a A A

Research On Activity Recognition Technology Based On CSI And Deep Learning

Posted on:2018-12-07Degree:MasterType:Thesis
Country:ChinaCandidate:F S ChenFull Text:PDF
GTID:2348330521451513Subject:Engineering
Abstract/Summary:PDF Full Text Request
Based on the Wi-Fi widely separated in the world,Wi-Fi-based wireless activity recognition has attracted more and more research efforts,and device-based activity awareness is no being used as the most important solution of the commercial at now.Variety of acceleration sensors had been used on many of devices.Direction sensors and other sensors is now very mature.However,with the understanding of the wireless signal more and more profound,through the device(Intel 5300NIC)to obtain the physical layer of signal information: channel state information(CSI)as a more granular than the RSSI signal information for Wireless signal sensing provides a theoretical basis.Which is concerned by many researchers,the main research content of the paper is based on CSI wireless activity recognition perception research,the main contributes are as follows:First of all,as a CSI-based wireless activity recognition system,a system for learning the classification of activities by means of depth learning is proposed.The Wi-Fi Activity Sensor program includes the data preprocessing phase,activity detection phase,learning stage and classification state.In the activity detection model stage,a correlation-based model is used to detect the time points of activity and the time interval of activity.Cleverly solved the waveform in the stable time due to changes in the environment and changing problem.In the activity identification state,the innovation through the depth learning of the network,the activity learning and training,through learning activities of the CSI signal information to replace the current mainstream of the fingerprint,which through th train network after the classification of activities.Base on the Pearson correlation coefficient method,an activity detection model base on Pearson correlation is proposed.By analyzing the waveform of the wireless signal,the correlation between the before and after signals is used to judge the occurrence time of the activity and the length of time than takes place.The activity detection model can adapt to the situation that the target is active when the waveform is inconsistent due to the changed of environment.To obtain the CSI information in the time window of the target activity.In order to be able to classify accurately,this paper uses the Alex Net deep learning network to carry out the target activity learning and use the VGG-Net group as the contrast network.After lots of data learning,training and experimentation,it is found that for the current experimental setup,The number of data can now only receive 30 sub-carrier,a single data sample size of the data some restrictions,so the network can not be too deep,nor too shallow,if too deep due to small samples lead to over-fitting,and the network is too shallow will lead to Training level is not enough,can not achieve a good training to learn and classification,it was found that Alex Net's final classification effect is more excellent.The chapters after this experiment.The system has been tested in all directions,including the difference in the transmission data rate,and the different power of different networks,the accuracy of the system for the stability of the evaluation.For the current system to achieve and average of 96% of the average recognition rate and average of 3% false positives.The above scenes are in a single wireless AP under the scene,especially our system can be high frequency,such as 802.11 ac or 24 GHz frequency,to achieve higher accuracy.
Keywords/Search Tags:Channel state information(CSI), Pearson product-moment correlation coefficient, Convolutional Neural Network, Alex Net
PDF Full Text Request
Related items