WiFi-InstancePose: Multi-Person Counting and 3D Human Pose Estimation from WiFi Signals

Abstract

WiFi-based human perception has recently emerged as a promising sensing modality due to its advantages in privacy preservation, illumination robustness, and deployment flexibility. However, existing methods remain limited to single-person or task-specific settings and cannot jointly infer human count and 3D pose, largely because of the lack of an instance-aware WiFi representation capable of modelling multi-person dynamics. To address the lack of instance-aware representations, we propose WiFi-InstancePose, an end-to-end framework capable of jointly performing human counting and multi-person 3D pose estimation directly from raw channel state information (CSI). Specifically, our framework incorporates a masked autoencoder-based WiFi backbone and an instance-aware query decoder, formulating multi-person 3D pose estimation as a direct set prediction task. By leveraging auxiliary supervision for person counting, the proposed method jointly estimates crowd size and 3D human poses in a unified and scalable manner. Evaluations on the Person-in-WiFi-3D Dataset (1-3 persons) demonstrate that WiFi-InstancePose achieves a human counting accuracy of 91.13%. Furthermore, it substantially outperforms the Person-in-WiFi-3D baseline in multi-person pose estimation, reducing the mean per joint position error (MPJPE) by $\mathbf{5. 5 6 \%}$ and $\mathbf{4. 1 6 \%}$ for 1-person and 2-person scenarios, respectively.