Automatic 3D-reconstruction of buildings using highly resolving video sequences

点击翻译按钮获取中文摘要

In the last period we successfully developed a prototype system to generate building models for an urban 3D GIS using terrestrial image data. One attribute of this fully automated system is its high redundancy of input and output data. This property has only been partially exploited so far. Therefore, we want to improve the egomotion recovery for extended camera excursions and the reconstruction quality using the integrated techniques of dense spatial matching and sparse temporal tracking. Several automatically selected representative points will be stored in an octree data structure and used as a control grid for the 3D scene. These points allow the system to determine optimal triangulation conditions for pose estimation, object reconstruction and error minimization. Furthermore, we want to extend our existing system with a quality measure for every 3D point. Weakly measured parts can reappear during the video sequence with better accuracy. These identical parts are detected by the representative points described above and enhance the global accuracy by selection of appropriate weights. The reconstructed fa?ade contains gaps from occluding objects like cars, trees and street lighting. Due to our large amount of redundant data we can fill up holes automatically. The critical task is to separate occluded parts and integrate point clouds from other reconstructions. Additionally, non-rigid objects must be detected using motion segmentation to prevent inconsistencies. Finally, we want to integrate several reconstructions with varying scales in an existing GIS database.