Approaches to acquisition of 3D facial data include laser scanners, structuredlight devices and (passive) stereo vision. The laser scanner and structured lightmethods allow accurate reconstruction of the 3D surface but strong light is projectedon the faces of subjects. Passive stereo vision based approaches do not require stronglight to be projected, however, it is hard to obtain comparable accuracy and robustnessof the surface reconstruction. In this paper a passive multiple view approach using5 cameras in a ’+’ configuration is proposed that significantly increases robustnessand accuracy relative to traditional stereo vision approaches. The normalised crosscorrelations of all 5 views are combined using direct projection of points instead ofthe traditionally used rectified images. Also, errors caused by different perspectivedeformation of the surface in the different views are reduced by using an iterative reconstructiontechnique where the depth estimation of the previous iteration is used towarp the windows of the normalised cross correlation for the different views.
展开▼