| The fall behavior in the family bathroom scene has the characteristics of strong closure,serious harm,and poor rescue timeliness.Therefore,it is of great significance to study how to detect it effectively.In recent years,Wi-Fi sensing technology has become an emerging direction of sensing research due to its low cost,non-contact,unaffected by light,and better privacy.How to achieve a high-accuracy,high-reliability fall detection on Wi-Fi facilities is critical.At present,the research on bathroom fall detection based on Wi-Fi sensing faces challenges such as lack of relevant datasets,limited recognition accuracy,and poor robustness.Aiming at the above problems,this thesis firstly constructs and preprocesses datasets of different bathroom scenarios.Secondly,the preprocessed data is input into our designed high-precision fall detection recognition model.Finally,an effective migration mechanism is established using matrix fusion.The main research work of this thesis is summarized as follows:(1)Two high-quality bathroom fall datasets are constructed.A high-quality dataset is a prerequisite for conducting Wi-Fi sensing research.The sensing model constructed on the basis of high-quality datasets has higher reliability and authoritative experimental results.In order to ensure the diversity of the dataset,this thesis constructs two bathroom fall datasets in the home bathroom environment and the public bathroom environment respectively.In addition,the raw data is processed into a dataset that can be directly input into the subsequent neural network.The process contents include the deployment of experimental equipment,data acquisition,data preprocessing,etc.(2)Aiming at the problems of insufficient feature extraction and limited recognition accuracy in the existing fall detection algorithms based on Wi-Fi sensing,this thesis proposes a deep learning-based non-contact bathroom fall detection model named Wi SFall.Firstly,an one-dimensional time series CSI data stream is reconstructed into a two-dimensional frequency energy map to make the acquired sensing data features more detailed and accurate.Secondly,Butterworth filtering is performed on the reconstructed sensing data to remove the white noise.Finally,the filtered sensing data is input into a constructed deep learning model.Through the effective extraction and classification of sensing features,a high-precision bathroom fall detection in the Wi-Fi environment is achieved.(3)Aiming at the problems that traditional fall detection methods based on vision or special sensors are not conducive to privacy protection and poor generality,this thesis proposes a perceptual model transfer method called Wi TFall,which adopts a method of data fusion based on the Wi SFall.Firstly,on the basis of the Wi SFall,the matrix fusion and model lightweighting methods are applied to the data preprocessing stage and the deep neural network model construction stage respectively.Secondly,the sensing model and transfer learning method are combined to preliminarily realize the cross-scenario capability of the model.Finally,a complete transfer sensing mechanism is established by effectively extracting and classifying the common features of the source domain and the target domain.(4)Aiming at the bathroom fall detection algorithm,this thesis designs and realized a fall detection system based on the Python Flask web framework.The functions of the system include uploading data,calling models,connecting to background databases,and displaying results.As an algorithm effect display system,it has a simple interface and complete functions,and its visual recognition process is convenient for users to analyze the fall detection results more intuitively.In summary,this thesis constructs two bathroom fall datasets and preprocesses them.Based on the two datasets,the bathroom fall detection models Wi SFall and Wi TFall are proposed respectively.Among them,the Wi SFall mainly focuses on data filtering and fall detection,while the Wi TFall mainly focuses on data fusion and model generalization.The experimental results show that both of the two models achieve a high accuracy of about 99% for falling behavior detection in the home bathroom scenario.Compared with other similar models,our proposed model performs better and has stronger generalization ability. |
