Research And Implementation Of Fall Detection Based On Ubiquitous Wireless Signals

With the trend of global aging,the safety problem of elderly people living alone has become an important research content,especially the fall detection has vital research significance.In this thesis,WiFi and sound wave,two most common ubiquitous signals in daily life,are selected to realize real-time fall detection.Due to the difficulty of collection of fall action samples,the number of fall samples in the study is often less than that of non-fall samples,which makes the detection model limited.In addition,there are many types of falls in real life,and there are many actions similar to falls,which increase the difficulty of fall detection.Ubiquitous signals are easily affected by the environment.When the environment changes,the performance of the fall detection model based on ubiquitous signals decreases.Most of the existing methods adopt the training method of domain adaptation to adapt the model to the new environment,but this method requires the data samples in the new environment to participate in the training,and the cost of collection and training is high.At present,most of the research on real-time fall detection is still in the theoretical stage,which does not take into account equipment stability,computing resource consumption and the validity of real-time activity data fragments in the actual scene.In order to solve the above problems,this thesis designs fall detection methods for WiFi and acoustic wave respectively.The main work is as follows:(1)A virtual fall sample generation method is proposed to make up for the shortage of fall samples and increase the diversity of fall samples.For WiFi,the amplitude data collected by each antenna pair were reconstructed and trained to retain the characteristics of the data of each antenna pair.For acoustic waves,2D convolutional and deconvolution autoencoders are used for reconstruction training.(2)A method for separating action and environment features is proposed.The source domain features were extracted by deep convolution model,and the feature vector was divided into two parts.The upper part was brought into the fall and non-fall binary network training,and the lower part was brought into the domain classification network training,so as to separate action information from environment information.This framework is applicable to both WiFi and acoustic wave.(3)A real-time fall detection algorithm is proposed.For WiFi,multiple activity segments are captured in real time by using the current environment amplitude variance as the benchmark of activity occurrence,and the final detection result is obtained by voting of multiple activity segment detection results.For acoustic wave,the acoustic data is collected in real time by interval sampling,and the detection results are obtained by cross detection.Experiments are carried out in various scenarios.The average accuracy of crossenvironment real-time fall detection based on WiFi and acoustic waves is 84.5%and92.5%respectively,which is 14.1%and 7.5% higher than that of directly using the source domain model detection method.At the same time,this thesis designs and implements a WiFi and acoustic fall detection system for actual application scenarios,and completes the corresponding function test and performance test of the system.