Dynamic 3D Scene Reconstruction in Outdoor Environments

A number of systems have been developed for dynamic 3D reconstruction from multiple view videos over the past decade. In this paper we present a system for multiple view reconstruction of dynamic outdoor scenes transferring studio technology to uncontrolled environments. A synchronised portable multiple camera system is composed of off-the-shelf HD cameras for dynamic scene capture. For foreground extraction, we propose a multi-view trimap propagation method which is robust against dynamic changes in appearance between views and over time. This allows us to apply state-of-the-art natural image matting algorithms for multi-view sequences with minimal interaction. Optimal 3D surface of the foreground models are reconstructed by integrating multi-view shape cues and features. For background modelling, we use a line scan camera with a fish eye lens to capture a full environment with high resolution. The environment model is reconstructed from a spherical stereo image pair with sub-pixel correspondence. Finally the foreground and background models are merged into a 3D world coordinate and the composite model is rendered from arbitrary viewpoints. We show that the proposed system generates high quality scene images with dynamic virtual camera actions