A Study Of Human Behavior Recognition And Context Awareness Technology Based On Ubiquitous Computing

The technological progress and application popularization of the Internet and intelligent hardware facilities have promoted the integration of intelligent devices and human society.The existing Internet digital platform has been striding forward to the Internet of Everything system,and the number of smart device connections in the future mobile communication filed will exceed hundreds or thousands of times compared with the current devices(smartphones,tablets and laptops).The computing environment of smart devices and the interaction environment with humans have become more and more complex,and humans have begun to have higher requirements for various new smart devices:"humans do not need to actively adapt to new smart devices,and the development of smart devices and interaction technology needs to actively adapt to human lifestyle".Ubiquitous computing,a new computing model,has been proposed,which aims to require smart devices to complete the wish of "invisible computing".In other words,when using smart devices,humans will no longer waste their attention on operating devices and paying attention to the operation process of devices,but can devote themselves to the tasks they want to complete.An important basis for realizing the vision of "invisible computing" in smart devices is context awareness.Context awareness consists of two main parts:contextual information and sensing.In human studies,contextual information is any information that can be used to characterize the human situation.Sensing can be divided into contact sensing and contactless sensing.The difference between contact sensing and contactless sensing mainly lies in whether the user has direct contact with the system that provides services.This contact is generally completed through wearable devices.In this dissertation,both contact and contactless sensing studies are carried out.Considering the extensive use of smartphones in human society and the widespread deployment of wireless devices in various households,this dissertation makes a systematic research by using the user location information,user use device information and user activity information in contextual information.Smartphones are more ubiquitous than other contact sensing means(wearables),and sensor information is more readily available.Wireless signals are less susceptible to light and obstacles than other contactless sensing methods(visual signal and ultrasonic signals,etc.),and even wireless devices are cheaper.Based on the fact that all kinds of human activities will affect the changes of signals(sensor signals and Wi-Fi signals),this dissertation focuses on the user location information,user use device information and user activity information,and makes a systematic research based on smartphone sensor signals and Wi-Fi signals.Moreover,we have implemented three different systems:AutoProfile,MotionParser and AFall.The main research work and contributions are as follows:1.AutoProfile can automatically and dynamically change smartphones’ profiles based on social scenarios,users’ motion states and the time of day contexts.AutoProfile introduces discrete wavelet transform(DWT)and Mel-frequency cepstrum coefficient(MFCC)as the features of ambient sound,which has better accuracy on the premise of smaller features.We evaluate AutoProfile’s performance in 8 scenarios.Experimental results demonstrate that AutoProfile can achieve overall recognition accuracies of 91.4%and 90.6%when using Random Forest and k-nearest Neighbors classifiers,respectively.2.MotionParser is mainly realized based on the correlation between the user’s keystroke position,the user’s force and the resulting change of direction sensor data.Through the evaluation on the smartphones,the experimental results show that the inference accuracy of words can successfully reach more than 70%in a specific scene(sitting at the desk and interacting).Compared with the existing methods,we propose to use letter group as the inference of letters,which greatly improves the recognition accuracy of smartphone keypad.3.AFall can detect the falls of the elderly indoors by sensing the angle of arrival(AoA)changes in different dimensions of the user.By evaluating the performance of AFall in five different scenarios,the experimental results demonstrate that AFall achieves 83.74%,83.64%,82.61%,85.36%and 85.91%,respectively.Compared with existing pattern-based work,AFall can be achieved without scenario-related training efforts and does not require recalibration to adapt to the changes in the living environment of the elderly people.Besides,even in the face of more complex environmental changes,our system can achieve good performance with a simple calibration.