Research On Methods For Indoor Human Activity Recognition Based On Wireless Sensing

Wireless signals produce regular signal change patterns when they encounter obstacles through reflection,diffraction,and refraction phenomenon in an indoor environment.These patterns can be used to sense the indoor environment,locate target objects,and recognize human activity via signal change patterns.Wi-Fi based human activity recognition is an important research topic in wireless sensing domain and has received extensive attention from academia and industry over the recent years.This thesis focus on using two Wi-Fi signal measurements:Received Signal Strength Indicator(RSSI)and Channel State Information(CSI)to analyze and evaluate the relationship between human activity and signal change patterns in terms of indoor environment,individuals,and activity.We design and propose three methods for human activity recognition by combining signal processing techniques with machine learning algorithms to reduce the influence of indoor environment and layout,individuals,activity categories,and interference factors of Wi-Fi signals on the accuracy of human activity recognition.The main research contents and contributions of this thesis include:(1)Construct a public dataset,WIAR,for Wi-Fi-based Activity Recognition.Wi-Fi human activity data have time-varying,depending on the surrounding environment,limited equipment platforms and high-quality data with time-consuming and laborious costs problems.The above problems cannot make private activity data scientifically and rigorously reproduce the previous research and hinder activity recognition techniques exchanges and development.Based on this,the thesis constructs a high-quality Wi-Fi activity dataset named WIAR.The WIAR dataset currently includes sixteen types of daily activity data provided by ten volunteers in three indoor environments and environmental data in different indoor and outdoor environment.Based on the need for scale expansion and the limitation of individual collection costs,the scheme integrating data analysis,data evaluation,data storage and data management is designed to ensure that the WIAR dataset can gradually expand the data scale,increase the types of human activity,enrich the diversity of indoor environment and improve the quality of activity data.(2)Design a method of human activity recognition based on cross-fusion of CSI and skeleton joint data.Wi-Fi signals are highly sensitive to indoor environments and cannot obtain stable recognition accuracy.Kinect's skeleton joints can help Wi-Fi data learning movement trajectories.Skeleton joint data cannot recognize human activity due to part of skeleton joints' overlapping lead to data loss of important skeleton joint.The wall-through perception capability of Wi-Fi data can make up for this problem and ensure human activity to be recognized.In response to the above problems and analysis,we design a Human Activity recognition method named HuAc which exploits the cross fusion of CSI and skeleton joint data to recognize activity.The HuAc includes a Wi-Fi module and a Kinect module.The Wi-Fi module uses K-means to select a set of subcarriers and extracts efficient data features.The Kinect module determines missing skeleton joints,selects important skeleton joints,and extracts feature data.The fusion of Wi-Fi feature data and skeleton joint data is used as the input of SVM.Compared with the skeleton joint and CSI-based methods,the experimental results show that human activity recognition accuracy of the HuAc has increased by 3%and 1.7%respectively,and the average recognition accuracy on human activity can reach 93%.(3)Design a method of fuzzy ZOne localization and Human Activity Recognition using Wi-Fi signals named ZOHAR.Aiming at the problems of human activity recognition in a fixed position that does not conform to real-life habits,independent activity and continuous activity are hard to discern in Wi-Fi signal change patterns,the ZOHAR designs a fuzzy zone localization module and human activity segmentation module.The fuzzy zone localization module first uses Maximum Mean Discrepancy(MMD)and Indicator Of Multipath(IOM)to analyze an indoor environment and then combines with the indoor layout and the relative location between one receiver and one transmitter to determine the segmentation area of the indoor environment and builds an zone-based adaptive Wi-Fi dataset using one hierarchical clustering method.The human activity segmentation module has designed a federated approach of Discrete Wavelet Transform(DWT)and moving variance to determine the activity boundary.Using CSI amplitude distribution,CSI change rate and frequency distribution distinguish independent activity and continuous activity.There are three activity attributes which are periodicity,direction and time interval of human activity can segment continuous activity.Zone information and activity data as inputs of the Sparse Representation Classification(SRC)algorithm realize fuzzy zone localization and human activity recognition.The experimental results show that the ZOHAR scheme obtains an average accuracy of 85.3%in the location of the fuzzy zone,and the accuracy of human activity recognition reaches 93.35%.(4)Design a LSTM-CNN Encoder-Decoder model named LCED for human activity recognition toward diverse individuals.Aiming at the poor migration problem of activity recognition models on individuals due to individual differences in habits and body types in a single indoor environment,the proposed LCED reduces the impact of differences between individuals on the accuracy of human activity recognition.The LCED model uses the Long-Short Term Memory(LSTM)as an encoder to obtain important information from the temporal CSI data corresponding to an activity to form a feature space.The feature space is transformed into a feature image representation,and the Convolutional Neural Network(CNN)is used as a decoder to learn effective feature data from the feature image representation.Compared with the classic classification algorithms,the overall recognition accuracy of the LCED model is improved by about 10%.The recognition accuracy of 16 types of human activity reaches 95%,and the accuracy difference is reduced by 3%on average.