Joint Loss-Optimized End-to-End Network for Through-Wall Multi-Person Pose Estimation with Portable UWB Radar

With advancements in through-wall radar (TWR), perceiving the fine-grained human poses obscured behind walls has become possible. Although imaging-based pose estimation methods have been explored, they heavily rely on image quality, which is influenced by sensor aperture. Due to the limited aperture and bandwidth of the portable TWR sensors, effectively utilizing them for multi-person pose estimation remains an ongoing challenge. This paper proposes an end-to-end neural network that directly processes TWR echoes for simultaneous multi-person detection and pose estimation using a portable low-frequency multiple-input multiple-output (MIMO) ultra-wideband (UWB) TWR. The proposed network architecture comprises three key components: a ConvNeXt V2 backbone for hierarchical feature extraction, a physics-informed mapping module that projects features into the 3D space, and dual output heads generating detection and pose estimation heatmaps across the entire scene. We introduce a jointly constrained loss function that integrates detection and regression results, effectively guiding network training toward accurate convergence using stereo camera-derived poses as supervision. Experiments conducted on both simulated and real-measured datasets demonstrate the method’s effectiveness in both line-of-sight (LOS) and through-wall scenarios, with ablation studies quantitatively validating the contributions of the proposed loss function and network design.