body2vec: 3D Point Cloud Reconstruction for Precise Anthropometry with Handheld Devices

Area of Science:

• Computer Vision
• Medical Imaging
• Biometrics

Background:

• Photogrammetry-based point cloud extraction often yields spurious detections, hindering 3D mesh reconstruction and accurate measurements.
• Existing noise removal techniques for point clouds are computationally intensive and struggle with semantic noise.

Purpose of the Study:

• To introduce body2vec, a model-based body segmentation tool utilizing a specialized Neural Network architecture.
• To enable high-quality human body point cloud reconstruction and anthropometric measurements from videos captured by handheld devices.
• To mitigate spurious point generation common in photogrammetric reconstruction through an effective background removal process.

Main Methods:

• Development and application of a specifically trained Neural Network architecture for body segmentation.
• Utilizing videos from handheld devices (smartphones/tablets) for point cloud reconstruction.
• Implementing a background removal step to eliminate spurious points.
• Comparison against LiDAR-based 3D meshes and expert anthropometric measurements.

Main Results:

• Body2vec achieved high-quality anthropometric measurements, outperforming traditional methods.
• The quality of point cloud reconstruction was significantly enhanced compared to LiDAR-based meshes.
• The method demonstrated a background removal capability, preventing usual spurious point generation.

Conclusions:

• Body2vec offers a robust solution for accurate 3D human body reconstruction and measurement using accessible technology.
• The tool achieves results comparable to LiDAR reconstruction while being more efficient.
• This approach significantly improves the quality of point cloud data and anthropometric measurements derived from consumer devices.