以視訊式環景影像產製三維點雲

Abstract

環景影像(Panorama image)為一涵蓋水平方向360度之影像，產製環景影像方法可分為：(1)單相機非同步取樣和(2)多相機同步取樣。單相機受限於非同步取樣的問題，本研究採用GoPro Hero4五台相機及Nikon KeyMission360單相機雙魚眼鏡頭同步取樣，由於同步取樣有些微時間秒差的問題，必須將各相機影像先進行時間同步，再進行環景影像拼接。本研究目的為透過多相機及單相機雙鏡頭以視訊式移動取樣，用影像拼接技術將多組視訊式影像拼接，再以不同的環景影像投影方式產製環景影像，最後結合多測站環景影像、地面控制點進行影像方位三維重建，利用影像匹配產製三維點雲(3D point clouds)。 研究方法分為四個部分，分別為：(1)環景影像產製、(2)影像方位重建、(3)密匹配產製三維點雲及(4)建立建築資訊模型(Building Information Modeling, BIM)。其中，環景影像產製是將各相機影像進行特徵萃取及影像匹配，利用匹配點進行影像拼接；影像方位重建以從運動回復結構(Structure from Motion, SfM)演算法進行方位求解；密匹配產製三維點雲是匹配出像空間中之共軛點，以共線條件式解算點位坐標，產製高密度三維點雲；建立建築資訊模型是以三維點雲為參考依據數化三維建物模型。 實驗分析項目包含：(1) GoPro Hero4五台相機不同測站數及不同投影方式產製三維點雲成果分析、(2) KeyMission360單相機雙魚眼鏡頭產製三維點雲成果分析、(3)點雲成果精確度比較分析、(4)產製BIM模型成果及(5)產製三維點雲與FARO地面光達點雲套合比較分析。實驗成果顯示，五台GoPro Hero4相機及Nikon KeyMission360相機之環景影像所產製之三維點雲，經過點雲精確度比較分析，點雲中線段長度與實驗區真實線段長度計算相對誤差，絕對值均小於3%；以點雲為參考依據建立BIM模型，模型之長寬高尺寸之相對誤差絕對值均小於1.01%。

Panorama image is a 360-degree image covering the horizontal direction. The generation of the panorama image can be divided into (1) single-camera non-synchronous taking images, and (2) multi-camera synchronous taking images. Since the single camera is limited by the problem of non-synchronous taking images, in this study, five GoPro Hero4 cameras and one Nikon KeyMission360 camera with dual fisheye lenses were used for synchronous sampling. Because of the slight time lag problem in synchronous taking images, each camera implements time synchronous before stitching panorama images. The purpose of this study is to use multi-cameras and single-camera with dual lenses to take images in a video mode, then stitch each image to panorama images in the different projection modes. The last, implement image orientation recovery by combining multi-station panorama images and ground control points to generate 3D point clouds using image matching technique. The methodology includes four major parts, (1) panorama images generation, (2) image orientation recovery, (3) 3D point clouds generation with dense matching, and (4) Building Information Modeling (BIM) construction. First, panorama images generation is to extract tie-points from every overlapped image pair, so as to stitch to panorama images with the extracted tie-points. Second, image orientation recovery uses the Structure from Motion (SfM) algorithm. Third, 3D point clouds generation with dense matching is to dense match the tie points in the image space, and then calculate 3D points coordinates with the collinearity condition equation. The last, Building Information Modeling (BIM) construction is to construct modeling based on the generated 3D point clouds. This experiment analysis includes five steps, (1) the 3D point clouds from five GoPro Hero4 cameras in the number of different stations and the different projection modes. (2) the 3D point clouds from a Nikon KeyMission360 camera with dual fisheye lens. (3) the comparison of 3D point clouds accuracy. (4) the analysis of BIM construction, and (5) the co-registration or image-based 3D point clouds and the FARO terrestrial LiDAR point clouds. The experiments show that through the comparison of the 3D point clouds accuracy between five GoPro Hero4 cameras and one Nikon KeyMission360 camera, the relative error from a length in 3D point clouds and actual line is less than 3%. Moreover, the relative error of 3D point clouds-based BIM model in length, width and height are all less than 1.01%.