Research On Device-free Wireless Localization And Activity Recognition

Device-free wireless localization and activity recognition(DFLAR)is an emerging technology in recent years,which evolves traditional communication network into an intelligent network with the ability to sense location and activity of a human while not requires the target to be equipped with any device.This makes DFLAR system a promising technique for many smart applications,such as,smart space,smart city and smart home.DFLAR is a pattern recognition problem,thus,features extracted from the RSS is crucial for the system.Traditional method utilizes time domain or frequency domain features to characterize the shadowing effect,which provides relatively simple information.In order to further extract features that reflect the impact on the wireless link,the paper explores to utilize of wavelet decomposition to decompose the data into different frequency bands,and extract zero-crossing points,variance and energy of two low-frequency bands,which could provide robust information.Compared with traditional methods,the DFLAR system based on wavelet features could improve the performance significantly.Whether traditional features or wavelet features,they all need to manually design handcraft features for specific scenario,which are not universal for the changing of target state or application scenario.In order to settle the problem,we adopt deep learning model to extract complete features from raw signal to realize DFLAR.In particular,we utilize a sparse autoencoder network to extract features automatically,and merge the softmax regression model to fine tune the whole network to realize DFLAR.Due to the fine description ability of deep features,this paper explores the shadowing mode of more fine-grained activities based on the extracted deep features,which not only realizes localization and activity recognition,but also identifies the gesture,thus,enabling the wireless networks with more power.The paper performs experiments with two hardware testbeds,the results show that the wavelet features could characterize the link measurement in both time and frequency domains,so that it could achieve better localization and activity recognition performance and stronger adaptability than traditional features;the features automatically extracted from deep learning model could not only save time and manpower,but have stronger adaptability to different scenarios,which can achieve satisfactory results even in a challenging environment.