Research On Model-Based Contact-less Human Motion Perception Algorithm

Perception and interaction technology is an important topic in the field of ubiquitous computing.In recent years,with the rise of the Internet of Things research boom,the research of perception networks centered on traditional wireless sensor networks has rapidly increased.At the same time,with the development of smart devices,new demands have been put on human-computer interaction,and contact-less sensing technology has emerged.Wireless signals are affected by environmental changes,and these effects apply to all frequency ranges of currently deployed common wireless data networks(such as Wi Fi,Blue Tooth,Zig Bee,and RFID).By continuously recording changes in physical measurements(such as signal strength or To F(Time of Flight)),contact-less sensing systems can analyze these signals to detect changes in the environment and correlate them with entities and their locations or activities.Traditional contact-less human motion perception systems generally use simple time/frequency domain statistical features in the main feature extraction step.For example,calculating signal statistic features such as mean/variance of CSI/RSSI signals in Wi Fi and Blue Tooth transmission,calculating the frequency change of the signal in the ultrasonic signal to obtain the Doppler frequency shift.Then it is further mapped to the corresponding action classification through machine learning methods.The main drawback of this type of method is that under the multi-path effect,the wireless channel transmission is very complex,so the statistical characteristics of the time/frequency domain are very sensitive to environmental changes and different habits of action executor.In order to achieve acceptable recognition accuracy,a large amount of data collection and training in different environments are often required.To solve this problem,our idea is a model-based feature extraction scheme,focusing on extracting information about channel changes caused by human motion in the signal,and connecting features to physical quantities such as speed and distance corresponding to human motion.Compared with traditional feature extraction methods,feature extraction and expression based on physical models improves the interpretability of the system and improves robustness in complex and variable environments.We proposed three basic models suitable for contact-less human motion perception,and in this thesis,we will use these three models to introduce the feature extraction technology based on physical model and the results achieved in the contact-less human motion perception system.CSI speed model: Traditional human motion perception system considers the indoor multi-path effect as a black box that cannot be explored,and the relationship between human motion and Channel State Information(CSI)signals is not clear.In this work,based on the superposition theory of wireless communication,we proposed the CSI-Speed model,that is,the correlation between the CSI signal frequency and the human motion velocity in radical direction.By using this model,we can integrate the information of CSI signals of different sub-carriers and extract the representative features that are truly related to actions.Through experiments,the motion classification system based on CSI speed model shows better robustness to different objects and environment changes.Ultrasound distance model: With the help of the Doppler model,the traditional ultrasonic gesture recognition technology can effectively determine the motion direction of the palm relative to the microphones and speaker,but it is often unable to effectively distinguish more complex finger gestures.In this work,we estimate the distance of the surrounding reflector by modulating the signal with good autocorrelation characteristics to the ultrasonic frequency band and calculating the Channel Impulse Response(CIR).Furthermore,by expressing Time-d CIR-Channel data as a 3-dimensional tensor,a deep learning model was used to distinguish multi-finger gestures that could not be effectively distinguished based on the Doppler effect.LTE long-distance propagation signal model: By interpreting and analyzing the channel reference signal(CRS)in the down-link channel of Time Division Duplexing(TDD)base station,we explore the possibility of using LTE signal as the medium of high-precision contact-less sensing system.In this work,we propose a block-principal component analysis(block-PCA)method to integrate information from different subcarriers with large amplitude and phase gaps due to the long distance propagation of LTE signals.By analyzing the processed LTE signal,we can passively monitor the human movement at a distance of up to 40 meters and tiny keystrokes at a distance of15 meters.In conclusion,we summarize the common system architecture and technical scheme of contact-less perception system,point out the main challenges in the research of contact-less perception system,and look forward to the possible research ideas and directions in the future.