3-D Data Acquisition of Scene by Motion Vision and its Understanding by Spatial Reasoning

通过解读不同视点拍摄的图像来理解场景的空间结构是计算机视觉的核心问题之一。为了对这些图像进行连贯的解释，重要的是建立图像之间的特征对应关系。我们发现了建立三摄像机立体特征对应关系的强几何约束，我们称之为三视约束。在这项研究中，我们扩展了三视约束，使它们也可以应用于运动视觉。我们利用边段作为对应的单位。建立对应关系后，利用三角剖分原理得到边缘段的三维坐标，建立场景的三维模型。然而，该模型是一种只知道边段三维坐标的线框模型。从这个线框模型中提取表面，通过分析这些表面之间的关系来识别场景的空间结构。在这里，我们采用基于知识的方法，其中知识库包含在三维世界中普遍存在的规则。

It is one of central problems of computer vision to understand the spatial structure of scene by interpreting images taken at various viewpoints.In order to interpret those images coherently, it is important to establish correspondences of features among the images. We have found strong geometrical constraints to establish correspondences of features for three camera stereo which we call as trinocular constraints.In this study, we have extended the trinocular constraints so that they can also be applied to motion vision. We have utilized edge-segments as units for correspondences. When the correspondences have been established, we obtain the 3-D coordinates of the edge-segments by using the triangulatin principle and make the 3-D model of the scene. This model is, however,a kind of wire-framed model where only the 3-D coordinates of edge-segments are known. From this wire-frame model, we next extract surfaces and recognize spatial structure of the scene by analyzing relations of these surfaces. Here, we took knowledge-based approach in which knowledge base contains rules which hold generally in the 3-D world.