.png){kind=icon}
.png){kind=icon}
Return
Human Skeletal Pose Estimation Based on Multiscale Spatial and Temporal Block Features Using Millimeter-Wave Radar
.png){kind=icon}
PRE
AI
C
H
Z
P
J
DOI：10.1109/jiot.2026.3704301.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Human skeletal pose estimation using millimeter-wave radar has gained increasing attention for its privacy-preserving and light-independent characteristics. However, the sparse nature of radar point clouds makes it difficult to robustly handle different motion states using a single-scale clutter suppression and feature extraction strategy. This article proposes a human skeletal pose estimation method using mmWave radar, based on multiscale spatial and temporal block features. First, a multiscale point cloud estimation strategy is adopted to capture transient short-term estimation, which extracts highly time-sensitive point clouds from single-frame radar data, while long-term estimation fuses multiple consecutive frames along the Doppler dimension to improve Doppler resolution and motion continuity. On this basis, multiscale point cloud blocks are constructed to represent different body segments, and K-means clustering compresses redundant points within each block and extracts high-velocity keypoints from subregions to highlight distinct motion features. Finally, a lightweight dual-branch neural network is designed to map the extracted features to human skeletal poses. By learning multiscale spatio–temporal block features with a joint weighting loss guided by motion velocity priors, the model effectively captures nuanced torso motion and dynamic velocity patterns. Experimental results demonstrate that the proposed method significantly reduces skeletal joint estimation errors and outperforms the existing related methods.
Keywords:
Human skeletal pose estimation
millimeter-wave radar
multiscale point clouds
spatio–temporal block features
Journal
IF:
8.9
Papers:
1.4W
Citations:
7.8W
Organization
Cited Papers