An adaptive algorithm for generating 3D point clouds of the human body based on 4D millimeter-wave radar

Area of Science:

• * Millimeter-wave radar technology
• * 3D human pose estimation
• * Point cloud generation

Background:

• * Traditional 3D human point cloud algorithms struggle with multipath effects, leading to inaccurate data and manual labeling needs.
• * Electromagnetic multipath interference causes phantom targets and classification issues in existing methods.
• * Lack of accuracy and manual intervention limit the practical application of current 3D human sensing techniques.

Purpose of the Study:

• * To propose an adaptive method for generating accurate 3D human point clouds using 4D millimeter-wave radar.
• * To address challenges like phantom targets and noise interference caused by electromagnetic multipath effects.
• * To enhance point cloud density and reduce the need for manual labeling.

Main Methods:

• * Development of the Self-Adaptive mPoint (SA-mPoint) algorithm utilizing 4D millimeter-wave radar.
• * Integration of micro-motion and respiration characteristics with dynamic and static echo information.
• * Application of multi-frame dynamic fusion and adaptive density-based clustering to mitigate multipath noise and improve point cloud density.

Main Results:

• * The SA-mPoint algorithm achieved an average accuracy rate of 97.94% in generating 3D human point clouds.
• * Compared to TI-mPoint, SA-mPoint increased point cloud generation by 87.94% and accuracy by 78.3%.
• * The proposed method reduced running time by 11.41% while improving point cloud quality and density.

Conclusions:

• * The SA-mPoint algorithm effectively generates accurate and dense 3D human point clouds from 4D millimeter-wave radar data.
• * The method demonstrates significant improvements over existing algorithms, particularly in handling multipath effects and enhancing data quality.
• * SA-mPoint shows high practicality and promising applications in human sensing and pose estimation.