UniTHP: Advancing Unified Representation Learning for Through-Wall Human-Centric Perception

Rui Zhang, Yadong Li, Ruiyuan Song, Hanqin Gong, Ruixu Geng, Binquan Wang, Dongheng Zhang, Yang Hu, Yan ChenIEEE Trans. Multimedia 2026 · DOI

Abstract

Recent years have witnessed increasing interest in sensing humans through walls with radio frequency (RF) signals, which enables critical applications such as surveillance and rescue. However, existing RF-based through-wall sensing methods either fail to consider the amplitude-phase characteristics of RF signals or cannot handle various perception tasks through an end-to-end approach. To address these challenges, we propose UniTHP - an RF-tailored single-stage multi-task framework specifically designed for through-wall human-centric perception. In particular, since RF signals contain both amplitude and phase information, utilizing both is critical for effective sensing. Therefore, we introduce a scale-mix attention block that models RF signals in their complex form while effectively preserving their signal properties. Additionally, this block implements multi-scale attention by splitting multiple attention heads into several groups, and each group focusing on a specific granularity of attention for different tasks. Building upon the scale-mix attention block, we carefully design the scale-mix encoder and task-specific decoder to develop our UniTHP framework. To evaluate the performance of UniTHP, we collect the first large-scale through-wall human perception dataset containing 491,000 pairs of optical images and radar frames and will make it public to the research community. Extensive experimental results demonstrate that UniTHP achieves new state-of-the-art results throughout all three downstream tasks, including pose estimation, action recognition, and person re-identification.
BibTeX
@article{zhang2026unithp,
  author = {Rui Zhang and Yadong Li and Ruiyuan Song and Hanqin Gong and Ruixu Geng and Binquan Wang and Dongheng Zhang and Yang Hu and Yan Chen},
  title = {{UniTHP: Advancing Unified Representation Learning for Through-Wall Human-Centric Perception}},
  journal = {IEEE Transactions on Multimedia},
  year = {2026},
  doi = {10.1109/tmm.2026.3675430},
}

Loading related content...

Figure