MUC2I: A Low-Redundancy CSI Preprocessing Method for Multi-User Wireless Human Activity Recognition
Year: 2026
Type: conference-paper
Abstract: Wireless Multi-User Human Activity Recognition (WMUHAR) enables non-invasive, device-free motion monitoring. As a key Integrated Sensing and Communication application in smart homes and healthcare, it faces deployment challenges in edge-cloud environments due to the high dimensionality of Channel State Information (CSI) data. Existing deep-learning WMUHAR research prioritizes accuracy optimization and model lightweighting but overlooks compression of raw CSI signals, leading to excessive communication, storage, and computational costs. This paper introduces MUC2I, a low-redundancy CSI preprocessing method that reduces data volume while maintaining high recognition accuracy. MUC2I first applies scale normalization to raw CSI to mitigate hardware and environmental interference, then converts the normalized data into pseudo three-channel images for efficient representation, enhanced by bicubic interpolation for richer perceptual details. Experiments on the WiMANS dataset demonstrate that WiGRUNT with MUC2I achieves a 9% average accuracy improvement over existing methods in the 5GHz band, while reducing preprocessed CSI file size by 83.3%, showing excellent performance in both recognition accuracy and data compression for complex multiuser scenarios.11This work is supported by Macao Polytechnic University (RP/FCA-03/2026 and No.fca.66f4.9e39.a). The corresponding author is Yue Liu.
Authors: Yangjing Zhou, Yue Liu, Chuan Liu
Institution: Macao Polytechnic University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7224824080