Design And Implementation Of Action Recognition System Based On COTS Millimeter Wave Device

As an important technology in the field of human-computer interaction,human action recognition has bright application prospects in security monitoring,medical care,somatosensory games.Currently,human action recognition methods are mainly based on wearable devices or videos.However,the method based on wearable device requires users to carry specific devices with them,which will bring inconvenience to the user's daily life;the video-based method brings privacy issues,these shortcomings limit the usages of these methods.With the popularization of wireless devices and Wi-Fi,Wi-Fi-based action recognition systems have gradually attracted the attention of researchers.The Wi-Fi-based action recognition system overcomes the shortcomings of the above-mentioned methods and can make use of currently widely deployed Wi-Fi equipments,and has the advantages of low cost and ease of use.This thesis designs and implements an indoor action recognition system based on CSI information of millimeter wave devices,which recognizes human actions by analyzing the changes in CSI signals caused by human activities.Compared with the currently commonly used Wi-Fi signals that work at 2.4GHz and 5GHz,millimeter-wave Wi-Fi that work at 60 GHz is more sensitive to changes in the environment,can improve recognition accuracy.Channel State Information(CSI),as a physical layer information with higher granularity than RSSI and better performance of environmental information,has high perception accuracy.The main work of this paper is as follows:(1)In order to extract CSI information from COTS millimeter wave devices,the underlying firmware of the device was cracked to realize online adjustment of antenna parameters in the beamforming process and provide the underlying foundation for calculating CSI through the SNR.(2)Propose an algorithm for gaining action intervals.The preprocessed data contains non-action parts.In order to gain the action interval,the system proposes a cutting algorithm using average absolute error(MAD)and sliding window.This algorithm first calculates the MAD value of the sliding window,and compares it with the threshold to determine whether it is the starting point(or ending point)of the action waveform,and further divides the sliding window into two smaller windows to compare with the threshold again,which can more accurately locate the CSI signal interval corresponding to human activity.(3)This thesis use algorithms based on support vector machines(SVM)and convolutional neural networks(CNN)to realize the function of action recognition.SVM-based algorithms use histogram of oriented gradient and gray-level co-occurrence matrix to extract image features,and use one-versus-one method to achieve multi-classification;the CNN-based algorithm uses the residual network ResNet-18 for action recognition.Experiments show that the method based on SVM achieves an accuracy of 94.2%,and the method based on CNN achieves an accuracy of 95.4%.The recognition results of different testers and different action positions are also discussed.