Model-driven segmentation of articulating humans in Laplacian Eigenspace

Area of Science:

• Computer Vision
• Medical Imaging
• Computational Anatomy

Background:

• Accurate segmentation of 3D human voxel data is crucial for various applications, including biomechanics and medical analysis.
• Existing methods struggle with complex poses and accurately delineating articulated body parts.

Purpose of the Study:

• To develop a general approach for segmenting 3D human voxel data into articulated chains.
• To enable automatic estimation of human body models from segmented data.
• To handle complex poses and non-rigid articulated structures.

Main Methods:

• Utilizing Laplacian Eigenmaps to map voxels into a high-dimensional Laplacian Eigenspace.
• Fitting 1D splines to identified articulated chains (limbs, head, trunk).
• Employing a top-down probabilistic approach for chain registration based on connectivity and properties.

Main Results:

• Laplacian Eigenspace effectively maps articulated chains to smooth 1D curves.
• Spline fit error successfully identifies junctions between body parts.
• The method accurately segments human body models in complex poses, including limb loops.
• Demonstrated effectiveness on both real and synthetic 3D voxel data.

Conclusions:

• The proposed method provides a robust and generalizable approach for segmenting articulated objects from voxel data.
• It facilitates automatic human body model estimation, even in challenging poses.
• The technique shows promise for applications beyond human body modeling, including robotics and animation.