Towards Fast 3D Reconstruction Using Silhouettes and Sparse Motion

Abstract

An efficient new 3D reconstruction algorithm designed for an industrial vision system is presented. The algorithm generates 3D models and motion estimates of rotating objects moving along a conveyor past a set of calibrated cameras. For our application the objects of interest (natural produce) have relatively simple surface geometries and this feature can be exploited in the reconstruction process. The proposed method combines shape-from-silhouette concepts with sparse motion tracking and is potentially fast enough to extend to real-time industrial applications. In addition to being robust and extremely computationally efficient, key differentiators for this work include (a) full exploitation of a priori model knowledge, (b) handling of highly dynamic and unpredictable object motions, and (c) support for objects containing relatively little shape and texture definition. The method is demonstrated and evaluated on a collection of synthetic example image sequences, and surface errors have been quantified as being less than 0.2mm from the ground-truth, providing confidence in the solution.