Research Article
WiMTAR: A Contactless Multi-target Activity Recognition Model
@INPROCEEDINGS{10.1007/978-3-030-94763-7_13, author={Pengsong Duan and Chen Li and Chenfei Jiao and Wenning Zhang and Jinsheng Kong}, title={WiMTAR: A Contactless Multi-target Activity Recognition Model}, proceedings={Mobile Networks and Management. 11th EAI International Conference, MONAMI 2021, Virtual Event, October 27-29, 2021, Proceedings}, proceedings_a={MONAMI}, year={2022}, month={1}, keywords={Wi-Fi sensing Multi-target Blind source separation Deep learning}, doi={10.1007/978-3-030-94763-7_13} }- Pengsong Duan
Chen Li
Chenfei Jiao
Wenning Zhang
Jinsheng Kong
Year: 2022
WiMTAR: A Contactless Multi-target Activity Recognition Model
MONAMI
Springer
DOI: 10.1007/978-3-030-94763-7_13
Abstract
At present, most Wi-Fi based sensing researches aim at single target scene, due to the difficulties in separation of mixed signals. In this paper, a Wi-Fi based model for multi-target activity recognition is proposed. A diverse dataset of sufficient volume for multi-target activity recognition is first collected in our paper. After blind source separation algorithm (FastICA) processing, the dataset is input to the proposed signal sort algorithm named CC-ICA for efficient and accurate signal sort according to CSI correlation coefficient. Experimental results show that CC-ICA algorithm can effectively solve the problem of random order caused by FastICA. Separated CSI data is input into a neural network consisting of ABiGRU and TCN for training and multi-target recognition evaluation. The experiments demonstrate that accuracy of WiMTAR is improved by 26% after CSI data is processed by CC-ICA for multi-target recognition, and accuracy of WiMTAR is also more than 2.6% higher than that of other single target recognition schemes.
