Robust Preprocessing and Deep Learning for Device-free Multi-Person Interaction Recognition Using Wi-Fi CSI Data

Abstract

Device-free Human Activity Recognition (HAR) using Wi-Fi Channel State Information (CSI) has emerged as a promising approach for intelligent indoor sensing applications due to its low cost and privacy-preserving capability. However, robust recognition of multi-person interactions remains challenging because existing studies often suffer from environmental noise, unstable signal variations, and limited investigation of effective segmentation strategies for spatial-temporal feature learning. This paper presents a comprehensive Wi-Fi CSI-based interaction recognition framework that investigates the effectiveness of sample-based and time-based segmentation approaches for improving classification performance. In the proposed framework, both CSI amplitude and phase information are combined to enhance feature representation. The raw CSI signals are preprocessed using min-max normalization and low-pass filtering to reduce noise and stabilize CSI fluctuations. Moreover, two window-based methods are evaluated, namely, by sample-based segmentation by 256 packets and time-based segmentation by 1-second windows. And then, these data are classified by using Support Vector Machine (SVM), and Convolutional Neural Network (CNN) models to analyze them. The experimental results are demonstrated, and the comparative analysis is presented based on recognition performance. The proposed framework contributes toward the development of robust and generalized Wi-Fi CSI-based multi-person interaction recognition systems for smart indoor environments.