Multi-pose Volume Reconstruction Across Arbitrary Trajectory from Multiple Fisheye Cameras

Abstract

Volume reconstruction from silhouettes is a known subject in the case of projective cameras. The use of dioptric omni-directional cameras (fisheye) that exhibit semispheric Field of View (FoV) allows simultaneous imaging of the whole available space. In this work we employ two fisheye cameras in order to acquire silhouettes of humans that move within the imaged indoor space without any kind of restriction. We present the basic algorithm for reconstructing the volumetric model of a human in the case of using two images of its silhouette, acquired by the fisheye cameras. Then we extend this algorithm in the case of a human moving along any trajectory a) assuming no pose change and b) considering pose change during motion. Quantitative results from synthetic data indicate that volumes can be reconstructed with high accuracy (error of few cm), from a small number of positions using only two fisheye cameras. The proposed algorithm achieves the same level of accuracy in the case of recovering the volume of multiple poses under same conditions.