WiFi sensing driven multi-person pose estimation via query-based set prediction

WiFi channel state information (CSI) has emerged as a privacy-preserving alternative to camera-based systems for human pose estimation. However, CSI-based pose estimation in multi-person scenarios remains an open challenge since CSI responses from multiple individuals are mixed through the wireless channel. As such, predicting the poses of all individuals in this scenario results in an unordered set of unknown quantities and ordering. To address this, we propose WiQPose, a query-based set prediction framework for multi-person pose estimation via WiFi CSI. A set of learnable query embeddings, each representing a candidate human instance slot, enables our model to directly produce decoupled pose predictions without explicit person detection. CSI measurements are first passed into a 3D convolutional encoder to extract spatio-temporal features. Through self-attention among the queries and cross-attention against the encoded CSI features, each query can aggregate instance-specific information and produce the corresponding keypoint heatmaps. We utilize the Hungarian algorithm during training to establish permutation-invariant supervision between predicted and ground-truth pose sets. Experiments demonstrate that WiQPose achieves an overall PCK@50 of 97.04%, while reaching 98.36% on the single-person benchmark, outperforming other representative methods.