M3Pose: Multi-Person 3D Pose Estimation Using Sparse Millimeter-Wave Radar Point Clouds

Yuexuan Feng, Songchen Dai, Qifei Zhang, Zhao Wang, Xianmin Zhang, Yulin Zhou

Abstract: Existing multi-person three-dimensional (3D) human pose estimation methods currently rely heavily on cameras, which are inevitably affected by smoke, haze, dust, poor lighting conditions, and privacy breaches in real-world scenarios. Due to the utilization of actively emitted frequency-modulated continuous waves for sensing instead of visible light, millimeter-wave (mmWave) radar possesses clear advantages in privacy protection and penetrative human pose estimation. To effectively estimate human pose, many researchers are turning towards the utilization of lightweight point clouds, which can be obtained from commercial mmWave radar units. However, the sparse nature of point clouds derived from mmWave radar, stemming from its low spatial resolution, poses challenges for accurate multi-person pose estimation. In this paper, a multi-person pose estimation (M3Pose) method is proposed using sparse mmWave radar point clouds. Firstly, improve the density-based spatial clustering with the applied noise (DBSCAN) method, which preprocesses multi-frame fused point clouds for multi-person separation and noise point removal. Then, inspired by PointNeXt, we further form a grouped point feature extractor (GPFE) module to obtain global action features and perform pose estimation by fusing local and global features of stacked GPFE outputs. M3Pose is the first work to use mmWave point clouds for multi-person 3D pose estimation. On the collected data set, comprehensive experiments illustrate the superiority of M3Pose compared to existing methods.