A Study Of Human Walking Sensing Technique Based On Channel State Information

With the innovation and development of embedded technology,wireless commu-nication technology,and Internet technology,the integration of applications of com-puter,Internet,mobile communication and digital multimedia has been promoted.More and more intelligent devices have gradually been appearing in people's daily lives,and constantly provide various customized information and services for people.However,in the traditional desktop computing or mobile computing modes,users must interact with devices through keyboards,mouses,touch screens and so on.To simplify the interactive procedure,the concept of ubiquitous computing has been pro-posed,which aims to integrate computing into the physical space and makes hardware devices disappear from people's sight so that people can pay more attention to the tasks they work on.From the perspective of human-computer interaction,this means that human-computer interfaces need to be extended from traditional input devices to the whole 3D physical space,namely,converting from "explicit interaction" of the traditional desktop computing or mobile computing modes to "implicit interaction".Context-awareness is an essential part of implicit interaction,and the ways of perception generally involve contact sensing and contactless sensing.Contact sens-ing basically relies on embedded sensors or wearable devices to sense the situational information of users,such as locations,behaviors and physiological indicators.How-ever,since contact sensing requires users to wear or contact with specific devices,it still can't achieve real implicit interaction to a certain degree.While contactless sensing doesn't require direct contact between users and devices.At present,the ma-jor contactless sensing methods are based on computer vision and speech.In recent years,Wi-Fi signals have been used for contactless context-awareness.Affected by different obstacles,Wi-Fi signals may travel through the line-of-sight(LOS)path or paths of reflection,diffraction and scattering to reach the receiver.Human activities can always change the propagation paths of Wi-Fi signals,and as a result,the received signal strength(RSS)or channel state information(CSI)of superimposed signals at the receiver side will change accordingly.By analyzing the change patterns of internal parameters of received signals,the Wi-Fi-based context-awareness technique is en-abled to realize human motion sensing.Compared with the vision-based or speech-based methods,the Wi-Fi-based sensing method isn't constrained by light conditions,and has lower power consumption and wider perception range than the speech-based systems.More importantly,it won't lead to the leakage of privacy data like video or audio.Besides,the wide deployment of Wi-Fi devices in the surroundings makes them an ideal basis for constructing ubiquitous context-awareness applications.Focusing on the subject of human walking motion sensing,this dissertation sys-tematically studies the human walking sensing technique based on the channel state information of Wi-Fi signals.The main work and contributions are as follows:1.Propose a walking detection approach based on spectrum energy.Through theoretical analysis and experimental validation,we find that,in indoor human daily activities,walking has a relatively large range of motion(compared with the mo-tions like waving,sleeping)and long duration(compared with the motions like sitting down,falling).The frequency of CSI amplitude fluctuation induced by the movement of human legs or feet during walking is within the 30?60 Hz band,of which the weighted spectrum energy is derived as the walking detection indicator.In terms of the energy as well as motion duration,the proposed approach can recognize walking motions with the average true positive rate(TPR)of 96.41%and false positive rate(FPR)of 1.38%.Since there is no need to extract complex motion features and train motion classifiers as in the existing motion recognition approaches,the proposed ap-proach has more concise processing flow and lower computational complexity,which can realize efficient and fast walking detection,and has high practicability in practice.2.Propose a step counting approach based on discrete wavelet transform and short-time energy.Given the identified CSI data fragments of walking,the subcar-riers with large variances of CSI amplitudes are dynamically selected from each CSI stream.The data of the selected subcarriers is subsequently decomposed into sev-eral wavelet coefficients with different frequency and time scales by discrete wavelet transform(DWT),and the short-term energy of the detail coefficients corresponding to the speeds of legs or feet is calculated.In order to eliminate the energy imbalance caused by the change of signal propagation distance,the moving standardization tech-nique is employed to rebalance the short-term energy and obtain more stable statistics of the effective peaks(the number of peaks is roughly counted as step count).The de-rived step counts of multiple subcarriers from multiple CSI streams are combined to output the final step counting result.The performance of the proposed approach is verified by building experimental platforms in two different indoor scenarios.The ex-perimental results show that the proposed approach separately achieves the average step counting accuracies of 90.2%and 87.59%for 15 subjects in the two scenarios,and it's robust to the environmental changes.Compared with the existing step counting methods,which directly utilize the low-frequency components of CSI amplitudes or are based on the torso speed,the proposed approach can achieve accurate and reli-able step counting for walking motions including in-place walking,and has a larger sensing range.3.Propose an attention-based gait recognition and walking direction estimation approach.Since different individuals usually have different body shapes and unique walking gaits,the unique movement pattern of each body part of an individual during walking can induce special change pattern of the CSI amplitudes of Wi-Fi signals.In addition,changes in the distance and direction of human walking relative to the transceiver can lead to different change trends of CSI amplitudes over time.By mining the inherent change patterns and trends of CSI amplitudes related to walking gait and direction with the attention-based RNN Encoder-Decoder,the proposed approach can jointly realize recognition and estimation of human gait and walking direction.The validity of the approach is evaluated by conducting walking experiments in three different indoor scenarios.The results demonstrate that the proposed approach can achieve the average F1 scores of 97.32%to 89.77%for gait recognition from a group of 4 to 10 subjects and 97.41%for direction recognition from 8 walking directions,and the average accuracies of these two tasks both reach 98%.Compared with the existing approaches,the proposed approach can adaptively concentrate on meaningful slices of CSI data sequence with the help of attention mechanism,thus the processes of gait cycle detection and segmentation can be removed in gait recognition.Meanwhile,given different tasks,the proposed approach can dynamically select specific data slices and automatically generate deep and high-quality gait features or direction features,which means there is no need for the extraction of artificial features.