A Transformer-Based Network for Human Pose Estimation using Millimeter Wave Radar Data

This paper proposes a human pose estimation method based on multi-angle millimeter wave radar images. The multi-angle images imply the 3D modeling of humans that can be used to recognize the pose. However, existing methods combine multi-angle features relying on local receptive fields, which misses the global information and has a poor precision of human pose reconstruction. A new network structure based on a transformer module is proposed in this paper to extract global information from multi-angle data and obtain an accurate human pose. In the proposed method, the transformer module is added between the encoder network and the decoder network. Then, a confidence refinement network is used to improve the position precision of human keypoints. Finally, a cross-modal supervision framework is utilized to train the network. Experimental results demonstrate an average OKS value of 0.716 in the AP75 evaluation metric, representing a 10% improvement over traditional networks.