Human model adaptation for multiview markerless motion capture

Abstract

An approach to automatic modeling of individual human bodies using complex shape and pose information. The aim is to address the need for human shape and pose model generation for markerless motion capture. With multi-view markerless motion capture, three-dimensional morphable models are learned from an existing database of registered body scans in different shapes and poses. We estimate the body skeleton and pose parameters from the visual hull mesh reconstructed from multiple human silhouettes. Pose variation of body shapes is implemented by the defined underlying skeleton. The shape parameters are estimated by fitting the morphable model to the silhouettes. It is done relying on extracted silhouettes only. An error function is defined to measure how well the human model fits the input data, and minimize it to get the good estimate result. Further, experiments on some data show the robustness of the method, where the body shape and the initial pose can be obtained automatically. © 2013 Dianyong Zhang et al.