Research On Activity Recognition And Position Estimation Using Amplitude Decomposition Of CSI Measurement

With the rapid development of wireless communication,wireless sensing has received great attentions for both academic and industry.Indoor localization and activity recognition of a person are widely used in many aspects of our daily life.In order to solve the problems of sensing in complex environment,this thesis studied activity recognition and position estimation algorithm using amplitude decomposition of CSI measurement and machine learning.The main work is described as follows:(1)The basic knowledge of human activity recognition and position estimation is studied.Firstly,the common measurements for activity recognition and position estimation are introduced.Then the existing CSI based activity recognition and position estimation methods by machine learning are described.Finally,an experimental platform is built to provide the foundation for the following research.(2)A CSI measurement decomposition based activity recognition and position estimation algorithm is proposed.In the off-line phase,the static information of CSI amplitude at each reference position is measured.At the same time,the CSI amplitude is obtained when target at different reference positions with different activities.Then,the CSI image is constructed with the CSI amplitude.Next,the convolutional neural network(CNN)is used for classification learning and the position classification model is obtained.For another,by subtracting the static information of corresponding reference position,the dynamic information of CSI for each activity can be obtained.Thus,the CSI amplitude dynamic information image is constructed.The CNN is used for classification learning and the activity classification model is obtained.In the online phase,the received CSI amplitude is used to construct CSI amplitude image.And then the position classification model is used for position estimation.For another,by subtracting the CSI amplitude static information of the estimated position,the CSI amplitude dynamic information image is constructed for activity recognition by the activity classification model.Through CSI amplitude decomposition,the background information of CSI can be eliminated and the performance of activity recognition can be improved.Experimental results show that the proposed algorithm can achieve better performance than the existing methods.(3)A CSI measurement decomposition and attention mechanism based activity recognition and position estimation algorithm is studied.Based on the proposed activity recognition and position estimation algorithm,three attention mechanism structures,including Non-local Neural Networks,SENet and DANet,are combined with the CNN for off-line activity classification and position classification training,respectively.In the proposed algorithm,attention mechanism can strengthen features and focus on important features which can improve the efficiency of off-learning.Experimental results show that the attention mechanism can improve the performance of activity recognition and position estimation dramatically.