Research Article
Design and Implementation of Non-intrusive Stationary Occupancy Count in Elevator with WiFi
@INPROCEEDINGS{10.1007/978-3-030-36442-7_1, author={Wei Shi and Umer Tahir and Hui Zhang and Jizhong Zhao}, title={Design and Implementation of Non-intrusive Stationary Occupancy Count in Elevator with WiFi}, proceedings={Broadband Communications, Networks, and Systems. 10th EAI International Conference, Broadnets 2019, Xi’an, China, October 27-28, 2019, Proceedings}, proceedings_a={BROADNETS}, year={2019}, month={12}, keywords={Wi-Fi Sensing CSI CWT CNN}, doi={10.1007/978-3-030-36442-7_1} }- Wei Shi
Umer Tahir
Hui Zhang
Jizhong Zhao
Year: 2019
Design and Implementation of Non-intrusive Stationary Occupancy Count in Elevator with WiFi
BROADNETS
Springer
DOI: 10.1007/978-3-030-36442-7_1
Abstract
Wi-Fi Sensing has shown huge progress in last few years. Multiple Input and Multiple Output (MIMO) has opened a gateway of new generation of sensing capabilities. This can also be used as a passive surveillance technology which is non-intrusive meaning it is not a nuisance as it is not need the subjects to carry any dedicated device. In this thesis, we present a way to count crowd in the elevator non-intrusively with 5 GHz Wi-Fi signals. For this purpose, Channel State Information (CSI) is collected from the commercially available off-the-shelf (COTS) Wi-Fi devices setup in an elevator. Our goal is to Analyze the CSI of every subcarrier frequency and then count the occupancy in it with the help of Convolutional Neural Network (CNN). After CSI data collection, we normalize the data with Savitzky Golay method. Each CSI subcarrier data of all the samples is made mean centered and then outliers are removed by applying Hampel Filter. The resultant wave is decimated and divided into 5 equal length segments representing the human presence recorded in 5 s. Continuous wavelet frequency representations are generated for all segments of every CSI sub-carrier frequency waves. These frequency pattern images are then fed to the CNN model to generalize and classify what category of crowd they belong to. After training, the model can achieve the test accuracy of more than 90%.
