A High-Fidelity mmWave Radar Dataset for Privacy-Sensitive Human Pose Estimation

Area of Science:

• Computer Vision
• Sensor Technology
• Human-Computer Interaction

Background:

• Privacy-sensitive environments require non-visual sensing for human pose estimation (HPE).
• Millimeter-wave (mmWave) radar offers a promising alternative, but lacks high-fidelity, annotated datasets.
• Existing methods often rely on visual sensors, posing privacy risks.

Purpose of the Study:

• Introduce mmFree-Pose, the first mmWave radar dataset for privacy-preserving HPE.
• Provide a high-quality, visually-free dataset for advancing RF sensing in sensitive areas.
• Facilitate research in non-visual human activity recognition and monitoring.

Main Methods:

• Developed a novel visual-free framework synchronizing mmWave radar with VDSuit-Full motion capture.
• Collected data covering 10+ actions, including gestures and falls, without visual sensors.
• Ensured privacy by design, eliminating visual data leakage while maintaining annotation fidelity.

Main Results:

• The mmFree-Pose dataset includes raw 3D mmWave radar point clouds (Doppler velocity, intensity).
• Features precise 23-joint skeletal annotations and full-body motion sequences.
• Incorporates challenging scenarios: occlusions, varied viewing angles, and multiple subjects.

Conclusions:

• mmFree-Pose addresses the critical need for privacy-compliant datasets in RF sensing for home monitoring.
• The dataset bridges the gap between mmWave radar capabilities and real-world privacy concerns.
• Enables robust and privacy-preserving human pose estimation in sensitive environments.