單一影像自動化建立城市建築物場景模型之研究

Abstract

本論文提出一個從單一城市街景影像中自動化擷取建築物表面並重建其影像之場景之系統。我們使用基於投票機制的影像消失點預測方法與線段偵測與分析進一步建築物表面的透視幾何資訊，並依此作為透視修正的影像修補方法之基礎。自動化建築物外牆表面經由四個階段做擷取 ，分別為:前置處理階段、初步表面擷取階段、表面精製階段與前景影像修補階段。每階段依序分析影像中提供的建築物牆面之幾何資訊，並整合後擷取建築物的牆面範圍與其在立體空間的轉向，讓使用者可便利且快速地完成城市街景建模。

This paper presents an automatic method for extraction of building facades from a single street view image. These façades can be used to reconstruct a primitive 3D street scene from the image. The extraction process involves our voting- based vanishing points prediction method and automatic line segment detection. The estimated vanishing points in the image are used to group line segments. We extract individual building facades and use perspective inpainting to remove the obstacles in front of the buildings. The orientation of each building facade is automatically evaluated for 3D scene reconstruction.