AMAR: Lightweight Attention-Based Multi-User Activity Recognition from Wi-Fi CSI

2026-05-19

AI Summary

This study addresses the challenge of signal overlap in Wi-Fi channel state information (CSI)-based human activity recognition caused by concurrent multi-user activities. To this end, the work proposes a lightweight edge–cloud collaborative framework that, for the first time, integrates set prediction and attention mechanisms into this domain. At the edge, a compact convolutional network extracts features and compresses data using residual vector quantization. In the cloud, a Transformer-based attention module with learnable query embeddings enables simultaneous recognition of multiple activities. The approach accommodates an arbitrary number of users and achieves substantially improved performance: in real-world settings, it nearly doubles the complete recognition rate of concurrent activities compared to the best baseline, attains an F₁-score of 53.4% (surpassing the baseline’s 45.6%), reduces user-count estimation error by 74%, and significantly lowers transmission bandwidth requirements.

Abstract

Wi-Fi-based human activity recognition (HAR) has emerged as a promising approach for contactless sensing, leveraging channel state information (CSI) collected from wireless transceivers. While existing studies have primarily concentrated on single-user scenarios, real-world deployments often involve multi-user settings where concurrent users' movements induce overlapping CSI patterns that challenge conventional classification methods. To address this limitation, this paper introduces an attention-based multi-user activity recognition (AMAR) framework that formulates HAR as a set prediction problem. The transformer-based architecture in AMAR leverages learnable query embeddings acting as specialized activity detectors, enabling the simultaneous identification of multiple activities from composite CSI representations. Moreover, to address deployment constraints, AMAR is designed in an edge-cloud split architecture form where lightweight convolutional networks on edge devices perform initial feature extraction, followed by residual vector quantization that achieves substantial bandwidth reduction while preserving activity-discriminative information. The cloud component performs final activity prediction through attention-based set matching, enabling the system to handle varying occupancy levels. Across classroom, meeting-room, and empty-room environments, on average AMAR nearly doubles the rate of perfectly predicting all concurrent activities compared to the best baseline. Moreover, it achieves an $F_1$-score of 53.4% compared to 45.6% for the best benchmark, and reduces occupancy estimation error by 74%, while minimizing bandwidth substantially.

Problem

• multi-user activity recognition
• Wi-Fi CSI
• overlapping CSI patterns
• human activity recognition
• concurrent activities

Innovation

• attention mechanism
• multi-user activity recognition
• Wi-Fi CSI
• edge-cloud architecture
• set prediction