Model-Based Vision System for Reconstructing 3D Objects from Multiple Images

Abstract

In this paper, we propose a model-based 3D reconstruction method that is suitable for feature-less objects. Typical objects with very few features are usually seen in industrial applications, such as heap of stones/minerals whose 3D reconstruction is very essential in automated storage and transportation. The shape of the heap is usually similar to a cone; hence, we use a parametric cone as the basic geometric model. At the beginning, we use background subtraction to detect the silhouette of the object for each image. The silhouettes are then used to estimate the parameters of the model. We adjust the size, position, and orientation of the estimated model by minimizing the overlapping ratio of its back-projections and detected silhouettes in images. At last, the model is voxelized to facilitate redundant area removal, and surface coloring is accomplished with visibility testing. For scenes with multiple objects, occlusion problem is formulated as a registration problem and solved by iterative closet point algorithm. Experimental results show that the proposed method is usually fast and the quality of the recovered objects is satisfactory.