计算机科学
标杆管理
协议(科学)
计算机网络
分布式计算
嵌入式系统
钥匙(锁)
实时计算
移动计算
通信协议
服务器
路由协议
作者
Di Zhang,Jiawei Huang,Yuanhao Cui,Xiaowen Cao,Tony Xiao Han,Xiaojun Jing,Christos Masouros
标识
DOI:10.1109/tmc.2026.3723025
摘要
Learning-based Wi-Fi sensing has made rapid progress, yet reproducible evaluation remains difficult because Channel State Information (CSI) measurements are strongly affected by hardware-dependent acquisition factors, preprocessing choices, and heterogeneous subcarrier layouts. This paper proposes the Sensing Data Protocol (SDP), a protocol-level data abstraction and benchmark interface for reproducible OFDM-based Wi-Fi CSI sensing. SDP specifies deterministic signal sanitization, profile-defined baseband frequency-axis projection, canonical tensor construction, and standardized evaluation reporting, mapping heterogeneous CSI measurements into a Canonical CSI Tensor. The protocol is designed to reduce avoidable measurement-side inconsistencies while retaining task-relevant physical variations, with an explicit trade-off between representation stability and information resolution. We evaluate SDP across representative Wi-Fi sensing tasks, including gesture recognition, gait-based identification, fine-grained activity recognition, and activity–location recognition. Controlled analyses compare native-resolution input, principal component analysis (PCA)-based reduction, uniform reduction, and SDP projection under matched splits and seeds. The results show that metadata-driven canonical projection and frequency-dimensionality control both affect sensing performance and computational cost, and thatK=30 provides a practical performance–cost trade-off in the evaluated Wi-Fi settings. Additional analyses on phase preservation, representation-level diagnostics, backbone compatibility, and calibrated residential stream inference further show that SDP provides a controlled and reproducible input interface rather than a task-specific accuracy optimizer. Overall, SDP provides a reproducible data and evaluation interface for Wi-Fi sensing, while leaving model architecture and task-specific optimization as independent design choices.
科研通智能强力驱动
Strongly Powered by AbleSci AI