Reconstruction of a Three-Dimensional Scene from its Projections in Computer Vision Systems
D. V. Gadasin, A. V. Shvedov, I. A. Kuzin
Abstract
The problem of reconstructing a three-dimensional scene from its projections is one of the most urgent and studied among the problems solved by computer vision methods. Most often, such systems are implemented by creating geographically distributed complexes that require the transfer of a large amount of data between their components, which creates a large load on the transmission lines. Scene reconstruction using projections obtained from several observation points is a special case and is not always possible, especially when used in unmanned aerial vehicles or autonomous vehicles. To resolve these limitations, the possibility of using a computer stereoscopic vision system and a three-dimensional convolutional neural network capable of reconstructing a complete scene from a single image containing subjective characteristics of scene objects, such as color and depth, is considered in this work, creating its three-dimensional representation. The article describes the architecture and principle of operation of a neural network, and also builds a model of a computer stereo vision stand for image registration. An image formation algorithm is proposed for this model.