Low-cost Robust Device-free Activity Recognition

Activity Sensing is emerging as a vital role in the health care,smart home,fitness tracking applications.Since falling can lead to almost 400000 people,indoor activity recognition has been gradually paid close attention by academia and industry.In recent years,both device-based and device-free activity recognition technologies have been proposed.Device-based solutions rely on carrying wearable devices at all times,which is very difficult and unpractical for every person,however,device-free activity recognition without any device attached to human becomes attractive and popular.Nevertheless,pervasive device-free solutions encounter several challenges.Particularly,its applicability is extremely limited by several drawbacks including poor robustness of fingerprints due to different multipath conditions,recognition errors caused by different activity speeds and orientations,etc.This dissertation targets at the above problems and studies RFID-based device-free activity recognition by designing a low-cost robust solution.In summary,the major results and core innovations are as follows:(1)Robust activity fingerprints extraction from low-cost and low sampling rate devices.Leveraging the theory of spatial spectrum estimation in MUSIC algorithm,we design a time domain based spectrum estimation method.In this method,we extract communication path length change rate caused by human activities(Doppler shifts),and build Doppler fingerprints which are not sensitive to multipath.Therefore,we can effectively avoid to rebuild the fingerprints for all activities when the monitoring environment changes,which means the human effort is reduced.(2)An effective activity matching method for different activity speeds and orientations.During the real-world experiments,we discover that different activity speeds and orientations can cause different time length fluctuations and amplitude fluctuations in activity fingerprints,which can lead to large errors when activity matching is processing.For the time length fluctuation problem,we employ DTW algorithm because of Dynamic programming technique.For the amplitude fluctuation problem,we optimize DTW algorithm by dynamically modify the proportion of variations in peaks and valleys of fingerprints to the dissimilarities of fingerprints.Thus,we can accurately recognize activities with building reference fingerprints only once.Finally,we make a prototype system for the idea only using cheap RFID hardware.Results from three typical experimental environments including a laboratory,a lodging house and a library,show that our design can identify human activities with the corresponding recognition accuracy of 92%,93%and 91%,which demonstrate the robustness of our methods.