3D Reconstruction of Human Motion from Monocular Image Sequences

Area of Science:

• Computer Vision
• 3D Reconstruction
• Human Motion Analysis

Background:

• Estimating non-rigid human 3D shape and motion from uncalibrated images is challenging.
• Existing methods often require significant camera motion for accurate 3D reconstruction.
• Arbitrary camera motion and occlusions limit current 3D human pose estimation techniques.

Purpose of the Study:

• To develop a robust method for estimating 3D human shape and motion from image sequences with uncalibrated cameras.
• To overcome limitations of existing methods by handling arbitrary camera motion.
• To achieve accurate 3D human reconstruction without predefined skeletons or anthropometric constraints.

Main Methods:

• Factorization of 2D observations into camera parameters, base poses, and mixing coefficients.
• Utilizing a-priorly trained base poses to enable reconstruction from arbitrary camera motion.
• Employing periodic assumptions for efficient estimation of periodic motions (e.g., walking).
• Introducing a novel regularization term based on temporal bone length constancy for non-periodic motion.

Main Results:

• The proposed method achieves convincing 3D reconstructions even with arbitrary camera motion.
• Demonstrated stability and accuracy through experiments using a 3D error metric.
• Successfully handles noisy data and occlusions.
• Outperforms state-of-the-art methods in 3D human shape and motion estimation.

Conclusions:

• The developed algorithm provides a significant improvement for 3D human shape and motion estimation from uncalibrated image sequences.
• The method is robust to challenging conditions like arbitrary camera motion, noise, and occlusions.
• It offers a flexible approach without relying on predefined skeletons or anthropometric priors.