AMAR: Attention-Based Multi-User Activity Recognition from Wi-Fi CSI
Wi-Fi-based human activity recognition (HAR) using channel state information (CSI) has been limited to single-user scenarios, but real-world deployments involve multiple users with overlapping CSI patterns. To address this, the paper introduces AMAR (Attention-Based Multi-User Activity Recognition), a transformer-based framework that formulates HAR as a set prediction problem. AMAR uses learnable query embeddings as specialized activity detectors to simultaneously identify multiple activities from composite CSI representations. It is designed in an edge-cloud split architecture for deployment efficiency. The work is published on arXiv with ID 2605.20649.
Key facts
- Wi-Fi-based HAR uses CSI from wireless transceivers
- Existing studies focus on single-user scenarios
- Multi-user settings cause overlapping CSI patterns
- AMAR formulates HAR as a set prediction problem
- Transformer architecture with learnable query embeddings
- Simultaneous identification of multiple activities
- Edge-cloud split architecture for deployment
- Published on arXiv: 2605.20649
Entities
Institutions
- arXiv