Research On Wireless-based Human Behavior Sensing In Smart Spaces

With the rapid development of mobile internet and Internet of Things,pervasive computing is moving from concept to reality.Smart spaces,as the most concrete and typical manifestation of pervasive computing,exploit the sensing devices embedded in them to perceive human,object and environment deeply and roundly,thus providing various intelligent services.Among these,user behavior perception is the crucial link to realize these services.Wireless sensing technology leverages the human-induced signal fluctuations to recognize human behaviors.The compelling advantages of wireless sensing,e.g.non-line-of-sight,passive sensing,low cost,easy to deploy,and powerful expansibility,enable it a promising alternative for human behavior sensing in smart spaces.In recent years,due to the wide popularity of Wi Fi networks and RFID technology,wireless-based human behavior perception has attracted much attention.While many efforts have been made,wireless sensing still faces many problems when it comes to practical applications.Firstly,most of existing works require the external factors(mainly including the human subjects and the environments)to keep consistent in both training and testing processes,which is difficult to be held in real settings.Secondly,existing works on action recognition only focus on isolated single action,rather than continuous action sequence.However,human actions in many settings are fluent and continuous.It is hard to spot and segment each isolate action from the continuous signal.In addition,labeling the continuous action sequence in the training process is also a very intractable problem.Thirdly,the high sensitivity of wireless signal to user and environment as well as the low resolution of commercial RFID and Wi Fi signal in time and space makes wireless sensing fail in multi-user settings,e.g.retail stores and trauma room.Therefore,exploring more general wireless sensing methods in real settings will do great significance for the deployment and popularity of wireless sensing technology in smart spaces.This thesis takes the two important issues in behavior perception,i.e.identity perception and action perception,as the researching clue,and focuses on the wireless-based human identification and continuous action sequence recognition,as well as multisensory behavior recognition.In the meantime,this thesis makes deep analysis on the sensing mechanism of wireless signals,and optimizes the sensing method,device deployment and signal preprocessing accordingly,so that the system can work in real settings more effectively.The main contribution of this thesis is given as follows:(1)This thesis makes a deep exploration on the mechanism of wireless based human behavior sensing.We construct the wireless sensing model and analyze the influence of human behaviors on signal amplitude,signal phase and channel state information.Besides,we comprehensively research and summarize the wireless sensing ability under different device deployments,users and environments,as well as behaviors.(2)Human identification is the premise of behavior perception.Compared to other biometric identification,gait-based identification can work in a non-contactable and non-invasive way,and is thus more geared to wireless sensing.To relax the requirement of existing works on the consistence of walking cofactors(e.g.carrying a backpack or appearance changes)in training and testing processes,this thesis proposes a RFID-based human identification method,RFree-ID,which can discern people irrespective of walking cofactors.The key insight is that the RFID reader and tags can serve as a radio gate,and the walking-induced RF signal fluctuations from tags are capable of perceiving different walking patterns when people cross this gate.More importantly,spatially separated tags can provide abundant temporal and spatial information for amplifying discrepancies among people and minifying the influence of walking cofactors.We build the signal reflection and diffraction model when people walk through the radio gate,and elaborate an adaptive multi-tag space fusion algorithm to realize human identification.Extensive experiments under various settings validate the reliability and robustness of RFree-ID.(3)Action recognition is the core of behavior perception and many human activities in real life are composed of continuous action sequences.This thesis takes sign language recognition for example to explore action sequence recognition and proposes a wireless based technical framework RFsign.RFsign is implemented on RFID and Wi Fi,respectively.First,RFsign leverages multi-tag of RFID and multi-antenna and multi-channel of Wi Fi to profile the time-space change of sign gestures.Then,RFsign builds the space-time model of action sequences based on deep learning network,which can handle environment and user diversity very well.On the top of the network,RFsign employs Connectionist Temporal Classification(CTC)to get around temporal segments and achieve end-to-end continuous sign language recognition.(4)For the behavior perception of multi-user in smart spaces,this thesis makes an exploration on leveraging multisensory,including wireless sensing,to recognize the identity and action of each user.This thesis takes the customer behavior perception in retail stores for example,and proposes a hybrid RFID and smartwatch method,IDeye,for behavior perception.The key insight of IDeye is that RFID tag and smartwatch can identify the item and the user,respectively,and both of them can profile the user-item interaction.Therefore,the fusion of RFID and smartwatch can enable the recognition of user,item and user-item interaction.To deal with the data heterogeneity of multisensory,IDeye constructs two action sequences via the two modalities,respectively,and enables the user-item pairing based on the similarity of these two sequences,which can also decrease the error rate from each modality and enhance the robustness of the system.