Passive Human Activity Recognition Based On LTE Signals

With the development of ubiquitous computing,radio frequency(RF)based human activity recognition has become one of the most popular research topics.However,traditional RF based human activity recognition systems still have many limitations,e.g.,small sensing range,high setup cost,and poor stability.In this thesis,we propose SpiderMon,a system that performs human activity recognition by using the LTE signals transmitted by commercial cellular base stations.Since the cellular base stations are deployed with a high density,our system can achieve large-scale,low-cost,and stable human activity recognition.We implement the system by using a software radio front-end.It captures the LTE signal reflections from surrounding human bodies and calculates CFR(Channel Frequency Response)without any active signal transmission.To reduce the phase noises in the CFR,we develop a two-step Carrier Frequency Offset(CFO)and Sampling Frequency Offset(SFO)estimator.After that,we leverage frequency compensation to mitigate the phase noise.We speed up signal processing algorithms by using multi-thread programming technology and Intel Math Kernel Library(MKL)so that one subframe of LTE data(with a duration of 1ms)can be processed in real-time on a workstation with a latency of 0.39ms.As a result,SpiderMon can extract the CFR from the LTE baseband in real-time by using normal work station.We validate that LTE signals can monitor user's^fine-grained motion distance and velocity by experimental studies.Then,we use the spectrogram of human activity to analyze the features on both time and frequency domain to distinguish the feature differences for different activities.Finally,SpiderMon applies short time Fourier transform(STFT)to extract the frequency features of different human activities and use support vector machine(SVM)to train and classify human activities.The main contribution of this thesis is that we are the first work that uses LTE signals to recognize human activities.We conduct different kinds of experiments to evaluate the performance of human activity recognition.Compared to previous activityrecognition systems that use Wi-Fi CSI,SpiderMon achieves longer operation distance from 5 meters to 20 meters.When the distance between the target and the receiving antenna is 20 meters,SpiderMon achieves 91.23%recognition accuracy for 6 different human activities on average.In the Non-line-of-sight(NLOS)scenario,SpiderMon achieves 89.69%recognition accuracy at a distance of 5 meters.