3D human model reconstruction is an evolving research area that is attracting significant attention due to its immense potential applications. The reconstructed models are widely adopted in various industries, such as anthropometry, ergonomics, motion analysis and garment fitting. Various techniques have been proposed to overcome the issues such as shape definition, shape matching and feature point extraction....[ Read more ]

3D human model reconstruction is an evolving research area that is attracting significant attention due to its immense potential applications. The reconstructed models are widely adopted in various industries, such as anthropometry, ergonomics, motion analysis and garment fitting. Various techniques have been proposed to overcome the issues such as shape definition, shape matching and feature point extraction.

This thesis presents a more natural and cost effective approach for 3D human model reconstruction. Monocular image sequences of natural human walking motion are captured by a typical digital video camera. Feature-based technique is adopted and a parametric human model template is created for the model reconstruction process. In the parametric rule generation of the feature template, a neural-based technique is employed. To support the training process, a sample set of 500 laser scanned human data are acquired and analyzed. The constructed parametric rules represent the unique 3D shape of human models across different anthropometric measurements. In the reconstruction, process, the captured parametric human model template is adopted.

In the model reconstruction, feature points representing the human landmark are extracted from the image sequence for the reconstruction process. A new signed curvature-based shape descriptor, Shape Profile, is proposed to cater for the extraction process. Shape similarity matching is firstly performed to select the 'best' matched template for the initial frame of the image sequences. Feature extraction is then carried out to extract the feature points from each selected frame of the image sequences using the cross correlation method. Compared with other feature extraction techniques, the proposed shape descriptor is more flexible, robust and efficient.

Applying the parametric feature-based model template, 3D human models are reconstructed and the anthropometry measurements are extracted from the image sequences. Experimental results indicate the proposed methodology reconstructs the human model with a higher accuracy than other known image based techniques.