A Sensing Dataset Protocol for Benchmarking and Multi-Task Wireless Sensing

TL;DR

The paper tackles fragmentation in wireless sensing data and benchmarks by introducing the Sensing Dataset Protocol (SDP), which defines a canonical tensor layout and lightweight synchronization to unify heterogeneous measurements. A CP-ALS pooling step yields a task-agnostic descriptor $\mathbf{h}$ that feeds a shared encoder and small task heads for detection, recognition, and vital-sign estimation, enabling a scalable, multi-task benchmark. Experiments on cross-user splits across Widar and Gait demonstrate substantially reduced variance (approx. $88\%$) while maintaining competitive accuracy, validating SDP as a reproducible foundation for multi-modal sensing research. The work advances cross-device reproducibility and fair benchmarking by standardizing preprocessing, pooling, and training/inference pipelines grounded in physically meaningful representations of wireless channels.

Abstract

Wireless sensing has become a fundamental enabler for intelligent environments, supporting applications such as human detection, activity recognition, localization, and vital sign monitoring. Despite rapid advances, existing datasets and pipelines remain fragmented across sensing modalities, hindering fair comparison, transfer, and reproducibility. We propose the Sensing Dataset Protocol (SDP), a protocol-level specification and benchmark framework for large-scale wireless sensing. SDP defines how heterogeneous wireless signals are mapped into a unified perception data-block schema through lightweight synchronization, frequency-time alignment, and resampling, while a Canonical Polyadic-Alternating Least Squares (CP-ALS) pooling stage provides a task-agnostic representation that preserves multipath, spectral, and temporal structures. Built upon this protocol, a unified benchmark is established for detection, recognition, and vital-sign estimation with consistent preprocessing, training, and evaluation. Experiments under the cross-user split demonstrate that SDP significantly reduces variance (approximately 88%) across seeds while maintaining competitive accuracy and latency, confirming its value as a reproducible foundation for multi-modal and multitask sensing research.

Figures (8)

• Figure 1: Protocol-defined data organization in SDP. Each packet-level CSIFrame contains timestamped per-subcarrier matrices, which are aggregated through a sliding window into CSIData samples and tensorized into the canonical (A, K, T) layout used by all tasks.
• Figure 2: Network architecture of the SDP benchmark model. Each window of CSI data is reshaped and processed by 2D spatial convolutions and a temporal Transformer module to generate task-specific features for detection, recognition, and vital-sign estimation. Dimensions are shown for a representative configuration.
• Figure 3: Comparison of two recognition tasks. Bars show mean performance over 500 runs (five fixed seeds, 100 repeats each); error bars indicate 95% confidence intervals (Student-$t$). On Gait, the SDP model significantly outperforms the baseline ($p<0.05$). Across both tasks, the SDP model exhibits narrower intervals, indicating improved stability.
• Figure 4: Row-normalized confusion matrices averaged over 500 runs.
• Figure 5: Across-seed variability maps over 500 runs. Variance concentrates on sparse off-diagonal cells that match residual confusions in Fig. \ref{['fig:confusion_matrices']}; diagonals remain low, indicating stable per-class recall.