Research On Key Techniques Of Activity Recognition Using Wireless Signals

The technology of wireless-based activity recognition has become a key research direction in field of human-computer interaction and pervasive computing.Narrow-band wireless signal is ubiquitous in space.When the signal encounters obstacles in propagation,it will be blocked,reflected or scattered.The features of these objects will be superimposed on the signal obtained by the receiver.If the obstacle is a moving human body,signals will carry the information of people's activities.Therefore,by abstracting the features of these signals and conducting classification,wireless-based activity recognition is achieved.First,we use wireless signals along with advanced machine learning techniques to implement the recognition of human activities.We exploit convolutional neutral networks to automatically extract key features hidden in signals to prevent the potential loss of accuracy caused by human factors.Also,this thesis presents a gradient-based method to detect the signal boundary of every single action,which effectively improves the recognition accuracy.Second,this is the first time that wireless signals are used to evaluate human's activity quality and the quantitative quality model of activities are proposed,through which fine-grained features(duration,speed,distance)are extracted from signals for quality recognition.Also,we propose an activity-based fusion policy.By detecting the difference between multi actions and their desired features,the quality of activities can be determined.In this thesis,an activity recognition and quality evaluation system is designed,naming WiQ,which utilizes narrow-band wireless signals for perception.In experiments,WiQ's accuracy of activity recognition reaches 98%.By using quality measuring results,WiQ performs an accuracy of 97% in recognizing three different physical and mental conditions of users.When used for the identification of 15 drivers,its average accuracy comes up to 88%.