應用多視角影像於UAV航拍遮蔽區之地形重建

The Application of Topography Reconstruction from Multi-View Images on the Tree Coverage of UAV Aerial Photogrammetry.

翁婕晞(Chieh-Hsi Weng)

指導教授 ： 黃金聰

國立臺北大學/公共事務學院/不動產與城鄉環境學系/碩士(2013年)

摘要

以航空攝影方式獲取地形三維資訊時，地面部分被植被覆蓋（例如：行道樹及路邊植栽等）或建物周圍與樹木相鄰與重疊時造成的遮蔽情況時常而見，以致於無法獲取遮蔽區（Tree coverage）底下的空間資訊，不論在三維建模方面或求取地形資訊時都會造成影響，因此本研究以非量測型數位相機所拍攝的多視角影像，經由Microsoft○R PhotosynthTM運用尺度不變特徵轉換（Scale-Invariant Feature Transform, SIFT）與移動中獲取空間結構（Structure from Motion, SfM）演算法，產製點雲資料（Point Cloud），並套合航空攝影測量資料，對遮蔽區下進行地形補償與重建，獲取完整的地形三維資訊。
本研究先採用UAV航拍影像，以航空攝影測量方法測繪數值地表模型（Digital Surface Model, DSM）與正射影像，並以物件導向影像分類方式（Object-based Classification），偵測正射影像中因樹木產生的遮蔽範圍，以此範圍濾除樹高後測製數值高程模型（Digital Elevation Model, DEM），並保留建築物高度，作為評估多視角影像重建地形之精度基準面。航空攝影測量空三的精度，航高自地面起算為540 m，檢核點三軸方向精度分別為 。對於遮蔽區底下地形，以非量測型數位相機拍攝的多視角影像，產生具有三維及色彩資訊之點雲，經坐標轉換套合至大地坐標系統中，評估精度得知點雲與基準面間的均方根誤差值（Root Mean Square Error, RMSE）為 0.277 m。最後，套合航測方式產製的DEM與SketchUP建構的建物資料，重建遮蔽區底下的完整地形資訊。

關鍵字

尺度不變特徵轉換 ； 移動中獲取空間結構 ； 點雲 ； 資料套合

並列摘要

Both the trees and plantes are occasionally coverd on the ground or partially overlap the building cause the “tree coverage” in the aerial images. The shade and shelter of trees and plantes prevent us from obtaining spatial information of topography. No matter when building 3D model or getting the information of topography is affected, especially measureing digital model. Therefore, we used the multiple-view images that shoot under the tree coverage by non-matric camera and procedured the point cloud through Microsoft○R PhotosynthTM which combining scale-invariant feature transform（SIFT）and structure from motion（SfM）algorithm, and make registration to the data from aerial photogrammetry. As result, we could obtain the completely 3D information of topography.
Our study took aerial photos of study area via unmanned aerial vehicles（UAV）, and generated the digital surface model（DSM）and ortho-image. Simultaneously, we used the tree coverage that detected by ortho-image and object-based classification to delete the high of trees, create digital elevation model（DEM）as well as keep the building. The DEM was used to be the basis of evaluating accuracy. The (x,y,z) accuracy of checking points in aerial trangulation was when flying height above ground was 540 meters. Under the tree coverage, we used the multiple-view images that shoot by non-metric camera to reconstruct geometry and procedure point cloud recording 3D and color information, and make registration to the coordinate of aerial photogrammetry after transformation. The root mean square error (RMSE) of the difference between the basis and the point cloud was 0.277 m. This value was calculated to check the accuracy of the experiement. Finally, the result was registrated to the DEM generated by aerial photogrammetry, constructed digital building model in SketchUP as well. In this way, we got the reconstruction of the complete 3D information of topography without the tree coverage.

並列關鍵字

Scale-Invariant Feature Transform（SIFT） ； Structure from Motion（SfM） ； Point cloud ； Registration

參考文獻

鄭智仁，2010，應用空載光達與正射影像以物件導向影像分析法進行自動化崩塌地偵測， 國立成功大學衛星資訊暨地球環境研究所碩士論文。

林耿帆，2012，以物件為基礎之光達點雲分類，國立臺灣大學工學院土木工程學系碩士論文。

蔡展榮、吳建億、李明軒，2010，測量車雙相機同步攝影測量，地籍測量，第19卷，第1期，頁13-25。

李育華、江凱偉、饒見有，2009，車載移動式測繪系統及直接地理定位控制場之設置，地籍測量，第4期第28卷，17-33。

黃韋凱、林銘郎、陳良健、林彥亨、簫震洋，2010，物件導向分析方法應用於遙測影像之分區及崩塌地與人工設施分類，航測及遙測學刊，第十五卷，第1期，頁29-49。