Structure from Stereo Vision using Optical Flow

Abstract

By incorporating principles of optical flow and scene flow for camera egomotion tracking, and stereopsis for calculating depth data, it is possible to generate a three-dimensional model of the camera’s surroundings. In this paper we describe our research into combining these two techniques, to generate a digital three-dimensional model of an environment from a video stream in real-time. Reconstructing the three-dimensional shape of a scene from its two-dimensional images is a problem that has attracted a great deal of research. The focus has mainly been on stereopsis, as the underlying geometry is well understood. By taking images from two cameras with a known positional relationship, we match points to determine disparity, and therefore depth. Optical flow is the two-dimensional motion field of point features in a sequence of images. This optical flow can represent the motion of an object within the scene, or the motion of the camera in a static scene. We therefore use optical flow to estimate the egomotion of a camera through a scene in real-time. Stereopsis, optical flow and scene flow were combined in this research to create a three-dimensional map of an unrestricted scene.