Abstract
Comparison to optical sensors, millimeter wave (mmWave) radar has advantages in scenarios with poor visibility or in the cases where privacy needs to be protected. Thus, it has been widely applied in human posture estimation for health monitoring or autonomous driving. However, the posture estimation using a single mmWave radar faces challenges such as low resolution, specularity, and noise artifacts in information extraction. To address these issues, we propose a novel human posture estimation method based on a diffusion framework, which is called as Diff-HPE. The Diff-HPE method employs a neural network denoising approach to reduce the noise in point clouds, and a multi-frame joint enhancement strategy based on motion sequence continuity to mitigate the sparsity and improve the stability. Furthermore, a diffusion model is used to extract the spatiotemporal details, enabling precise estimation of key human joints in indoor environments over long distance. Our comprehensive experiments demonstrate that the system can achieve efficient human pose estimation with an average localization error of 5.2 cm, exhibiting high accuracy and robustness even in the scenarios where the visual pathways are completely obstructed.
Funding Information
  • China Postdoctoral Science Foundation (2023M730554, BX20220065)
  • Natural Science Foundation of Jiangsu Province of China (BK20230816)
  • National Natural Science Foundation of China (62288101, 62301146)