Vision-Based 3D Scene Modeling Research And Implementation

With the development of information technology, there are more and more picturesand videos in people's life and the Internet. Nowadays virtual reality technology and freeview video technology are stimulating people's interest in watching 3D digital scenes. Itbecomes necessary and prospective to organize existing pictures and videos to 3D scenemodels. Moreover, computer vision theory has been signi?cantly developed in recentdecades, and laid the foundation of vision-based 3D scene modeling.3D scene modeling is important in the interdisciplinary research of computer visionand computer graphics. For vision part, 3D scene modeling is to process pictures andvideos and get the 3D scene model stored in memory or disk. Graphics technology, mainlyas post-processing, to render and represent the 3D scene model. We focus on the visionpart in this paper. The main contribution of this paper is as follows:Firstly, we summarize the most important theory and algorithms for 3D scene mod-eling with detail, and improve the bundler algorithm, getting a more ef?cient algorithm.Secondly, we propose and implement a robust two-frame stereo algorithm undervariant illumination and devices. We combine SIFT descriptor and NCC method to geta matching cost function: SIFT-NCC, and re?ne the result using second-order smoothconstraint under MRF framework. It is proved to be more accurate than other algorithmsfor illumination and device variation.Finally, we propose and implement a vision-based 3D scene modeling method fromlarge uncalibrated image collection. We develop an effective image collecting strategyto get a subset, only high quality images included. The image collecting strategy beingintegrated, we implement a self-calibration algorithm based on sparse bundle adjustment.Furthermore, we propose MFF-MVS(Multiple Features Fusion Multi-View Stereo) mod-eling algorithm to get the dense 3D point data and 3D surface model.