以多重疊近景影像萃取牆面三維線段之研究

Abstract

建物模型為數碼城市中重要的元素之一，建物模型在細節上，可由粗略的方型模型，逐漸以屋頂、牆面、室內結構加以細化，以提升將建物模型的細節。細緻的建物模型，除視覺上更貼近其真實樣貌，並能使用於較精細的應用，以利後續決策程序。為了將建物的細節等級提升，本研究發展一自動化牆面線型結構重建程序。主要項目為自動化多重疊近景影像的方位重建、多影像的牆面線型結構匹配與線段重建。自動化方位重建中，利用加速強健特徵點演算法，於多影像中萃取大量特徵點，配合少量地面控制點，以光束法平差方式重建影像外方位。匹配時，以建物中的線型特徵做為匹配目標，利用物空間匹配搭配多視窗，降低近景影像中的尺度、旋轉以及高差移位的影像。完成匹配後，物空間線型結構點群以RANSAC方式，計算三維線段參數，重建三維線型結構。成果顯示，自動化方位重建方面，檢核點差值為3~5公分；物空間匹配方面，精度平均為2公分；而線段重建方面，精度平均為10公分。

Building model is one of the important elements in cyber city. The detail of building model can be distinguished into wired frame, roof structure, facade structure and indoor structure. The more detailed model is not only more similar to its true appearance, but also can be applied to more delicate aspect, which may facilitate decision making procedure. This research, in order to raise level of detail of building model, has developed an automatic facade linear structure extraction procedure, including orientation reconstruction, multiple image matching and line fitting. A large number of conjugate points in multiple images are generated by Speeded Up Robust Features (SURF). With a few control points, orientation reconstruction can then be done by bundle adjustment. Aiming on linear features, object-based matching combined with multiple windows matching is applied to decrease the effect of image difference caused by scale, rotation and relief displacement. Finally, 3D line fitting is done by Random Sample Consensus (RANSAC). The experimental results indicate that the precision of orientation reconstruction is about 3 to 5 cm. The precision of object-based matching is 2 cm in average and the precision of line fitting is 10 cm in average.