Three-dimensional scene reconstruction from a two-dimensional image

摘要

We propose and simulate a method of reconstructing a three-dimensional scene from a two-dimensional image. The method is useful for developing and augmenting world models or maps for autonomous navigation. Our solution to this ill posed problem is the construction of a 3D scene from its 2D projected image. The inverse problem of the mapping from a 2D plane of pixels into a 3D space of voxels requires prior knowledge in order to infer the voxel states. Three main approaches can be taken in addressing this problem; surface based, model based, and volumetric based. Our approach is model and volumetric based encompassing object recognition, segmentation, 3D pose estimation, and 3D scene construction. It is an extension of the Perspective-n-Point (PnP) method which uses a sampling of the 3D scene, 2D image point parings, and Random Sampling Consensus (RANSAC) to infer the pose of the object and produce a 3D mesh of the original scene. In this research, we extend the PnP method on a scene of 3D objects and with an eye to implementation on embeddable hardware. The computations required by these methods are processor intensive but are parallelizable enough to implement on GPU hardware approaching the performance of a real-time system. One of the main objectives of this research is to utilize 3D scene construction on an embedded system where it may be used for navigation. The final solution will be deployed on the NVIDIA Tegra platform.