RPDiff: Multi-hypothesis Human Pose Estimation based on mmWave Radar

Compared to optical sensors, mmWave radar could provide 3D point clouds about the scene where has poor visibility and privacy needs to be protected, thus has been used in human posture estimation (HPE) in health monitoring and autonomous driving. However, the accuracy of HPE is limited by the sparsity of mmWave radar point clouds, resulting in a significant gap compared to visionbased solutions. Here, we introduce the RPDiff, a diffusion-based framework for human pose estimation (HPE) based on point clouds from mmWave radar. RPDiff combines PointNet ++ and selfattention mechanism into a denoising diffusion probabilistic model to balance accuracy and efficiency. Pose estimation is refined through forward and reverse diffusion processes of RPDiff with filtered radar point clouds as input. Testing on both public and proprietary datasets shows that RPDiff excels in reducing pose estimation errors and decreasing computational demands, making it suitable for real-world applications where optical sensors are disabled, demonstrating its strong potential and practicality.