Research And Implementation Of Sit-up Action Recognition System Based On Channel State Informatio

Nowadays,many studies have proved that sedentary behavior will increase the risk of many malignant diseases,but most people still keep sitting for a long time in their daily life.With our life improving,people pay more attention to their own health problems.Researchers have designed a variety of activity recognition systems for this purpose.By recognition the sit-stand activity,they can detect whether the people are sedentary or not.Existing activity recognition systems often use the methods based on sensors and video images to recognize specific activities,but most of them still face many problems,such as equipment cost,difficult deployment and privacy invasion.In recent years,many researchers use Wi Fi signal to recognize human activities because of the large-scale deployments of Wi Fi network.This thesis designs and implements a system that uses channel state information(CSI)to recognize sitstand activity.The system has the advantage of low cost,easy to deploy and privacy protection.The main work of this thesis is as follows:1)In order to accurately recognize the sit-stand activity,this thesis designs and implements a sit-stand activity recognition system.The system can be divided into three main parts,namely data preprocessing,training and recognition,error correction.After filtering the outliers and noise of the original CSI amplitude data,a two-phase segmentation method is proposed to accurately segment different activities.Next,a sitting activity filtering method has been proposed to remove these activities.Then the system uses a trained convolution neural network(CNN)model to recognize sit-stand activity.Finally,an error correction method based on activity reference group(ARG)is proposed to correct the wrong recognition result of CNN.2)To address the problem that the ARG cannot be found in the error correction part of the sit-stand recognition system due to the low accuracy of CNN recognition results,this thesis optimizes the error correction algorithm.By analyzing the cohesive relation between sit-stand activities,an error correction method of sit-stand activity recognition results based on hidden Markov model is proposed in this thesis.This method uses HMM to model the error correction process,and then proposes a Viterbi algorithm combined with waveform mirror symmetry factor(WMSF)according to the mirror symmetry relationship between sit-stand activity waveforms,so as to better correct the sit-stand activity recognition results of CNN.3)In this thesis,we use several experiments to verify the effects of two-phase segmentation method,sitting activity filtering method and error correction method,and we also evaluate the robustness of our system.The experimental results show that the sit-stand activity recognition system we proposed could increase the activity recognition accuracy by 14.7% on average compared with other existing systems,and the activity recognition accuracy in line-of-sight and non-line-of-sight scenario can reach at 98.7% and 96%.Moreover,when there is interference from others or human interference,the accuracy of activity recognition in different scenarios can still be maintained at about 90%.The entire system has a certain robustness and antiinterference ability.