由於三維感測器如光達雷達、立體視覺 等感測器逐漸的普及，點雲資料的獲取更加方便，並開始利用於各種領域中。自走車領域會使用立體感測器來獲取路面三維資訊應 用在式內定位或避障系統。逆向工程中可以將無 CAD模型的複雜 實體 曲面可以高 精度三維感測器量測並還原曲面的幾何資訊。 本研究希望以立體視覺的方法對應用於橋梁裂縫的檢測。
橋樑檢測前以目視檢測為主，危險性高且非常耗時，隨著無人機的研究發展，了解到無人機搭配影像的技術非常適合用於輔助橋樑檢測。但是單純使用影像在量測上會有需多限制，因此本研究希望以有深度的影像感測器應用於橋樑檢測上。研究中使用雙目視覺深度像機產生點雲資料和影像資料，以點雲濾波初步處理點雲資料，後計算點雲表面法向量最後並使用表面法向量為特徵以隨機採樣一致性的方法分割裂縫平面，整個點雲資料處理的目的在得到裂縫所在的平面的三維資訊。在以影像強化、影像形態學、和區域成長法等影像處理方法分離裂縫影像素，最後合併平面和裂縫影像的資訊獲得裂縫點雲，並以最小平方法來估計裂縫的尺寸。在實驗數據顯示當拍攝歪斜角度於 40度 時，轉正後誤差在 6%以內，說明若立體視覺量測可在一定程度上改善單純影像 量測時的誤差，且立體視覺因為有距離資訊在量測時不需要以另外的參考物來推估實際尺寸。

As three-dimensional sensors such as lidar and stereo vision have gradually become popular, the acquisition of point cloud data has become more convenient and has begun to use in various fields. In the field of automatic mobile vehicles, three-dimensional sensors are used to obtain the information of the road surface, which can be used in the indoor positioning or obstacle avoidance system. In reverse engineering , complex surfaces without CAD models can measured with high-precision three-dimensional sensors so the geometric information of the surface can be restored. This research hopes to used a stereo vision base method to detect and measure the bridge crack.
The main method of bridge inspection is visual by human’s eyes, which is dangerous and time-consuming. The technology of drones with images is very suitable for bridge inspection. However, simply using the image will have many restrictions on the measurement. This research uses a stereo vision to generate point cloud data and image data. First use point cloud filtering to process data. Then calculate the surface normal vector of the point cloud, and using Random Sample Consensus (RANSAC) method to divide the crack plane. To obtain the three-dimensional information of the plane where the crack is located. To Separate the crack image pixel, using image processing methods such as image enhancement, image morphology, and region growth method. Combined the information of crack image and crack plane and used least square method to estimate the size of the crack. Finally, the experimental data shows that when the sensor and crack surface is less than 40 degree, the error after correction is within 6%. It shows that using stereo vision measurement can improve compare to only using image sensors. And because of the distance information, the stereo vision is no need a reference object when estimate the measurement.

[1] 交通部全球資訊網 . (2021). 道路長度及橋梁座數 -中華民國交通部 . Available: https://www.motc.gov.tw/ch/home.jsp?id=64&parentpath=0,6
[2] Stereolabs. (2021). Developer Center | Stereolabs. Available: https://www.stereolabs.com/developers/
[3] OpenCV. (2021). Home - OpenCV. Available: https://opencv.org/
[4] PointCloudLibrary. (2021). Point Cloud Library | The Point Cloud Library (PCL) is a standalone, large scale, open project for 2D/3D image and point cloud processing. Available: https://pointclouds.org/
[5] PointCloudLibrary. (2021). The PCD (Point Cloud Data) file format — Point Cloud Library 0.0 documentation. Available: https://pcl.readthedocs.io/projects/tutorials/en/latest/pcd_file_format.html#pcd-file-format
[6] PointCloudLibrary. (2021). Writing Point Cloud data to PCD files — Point Cloud Library 0.0 documentation. Available: https://pcl.readthedocs.io/projects/tutorials/en/latest/writing_pcd.html#writing-pcd
[7] 吳載根 , "線上自動光學檢測系統 ," 碩士 , 機械工程 系, 國立臺灣科技大學 , 台北 市, 2019.
[8] 郭浩 , 點雲庫 PCL從入門到精通 . 機械工業出版 , 2020.
[9] PointCloudLibrary. (2021). Removing outliers using a Conditional or RadiusOutlier removal — Point Cloud Library 0.0 documentation. Available: https://pcl.readthedocs.io/projects/tutorials/en/latest/remove_outliers.html#remove-outliers
94
[10] PointCloudLibrary. (2021). Downsampling a PointCloud using a VoxelGrid filter — Point Cloud Library 0.0 documentation. Available: https://pcl.readthedocs.io/projects/tutorials/en/latest/voxel_grid.html#voxelgrid
[11] PointCloudLibrary. (2021). How to use a KdTree to search — Point Cloud Library 0.0 documentation. Available: https://pcl.readthedocs.io/projects/tutorials/en/latest/kdtree_search.html#kdtree-search
[12] YuCheng. (2021). KD-Tree 演算法筆記演算法筆記 | YuCheng. Available: https://blog.yucheng.me/post/kd-tree/
[13] G. Strang, Linear algebra and learning from data / Gilbert Strang. Wellesley, MA : Wellesley-Cambridge Press, c2019., 2019.
[14] PointCloudLibrary. (2021). Estimating Surface Normals in a PointCloud — Point Cloud Library 0.0 documentation. Available: https://pcl.readthedocs.io/projects/tutorials/en/latest/normal_estimation.html#normal-estimation
[15] K. G. Derpanis, "Overview of the RANSAC Algorithm," vol. 4, no. 1, pp. 2-3, 2010.
[16] PointCloudLibrary. (2021). How to use Random Sample Consensus model — Point Cloud Library 0.0 documentation. Available: https://pcl.readthedocs.io/projects/tutorials/en/latest/random_sample_consensus.html#random-sample-consensus
[17] MathWorks. (2021). Image Processing Toolbox - MATLAB. Available: https://www.mathworks.com/products/image.html
[18] R. C. 岡薩雷斯岡薩雷斯 Gonzalez, 數位影像處理數位影像處理. 臺灣培生教育出版 2009.
95
[19] MathWorks. (2021). Contrast-limited adaptive histogram equalization (CLAHE) - MATLAB adapthisteq. Available: https://www.mathworks.com/help/images/ref/adapthisteq.html
[20] 蕭凱文 , "利用多軸無人機影像萃取橋梁劣化區三維空間資訊 ," 碩士 , 測量及空間 資訊學系 , 國立成功大學 , 台南市 , 2017.
[21] MathWorks. (2021). Morphologically close image - MATLAB imclose. Available: https://www.mathworks.com/help/images/ref/imclose.html
[22] MathWorks. (2021). Adaptive image threshold using local first-order statistics - MATLAB adaptthresh. Available: https://www.mathworks.com/help/images/ref/adaptthresh.html
[23] MathWorks. (2021). Measure properties of image regions - MATLAB regionprops. Available: https://www.mathworks.com/help/images/ref/regionprops.html
[24] R. M. Murray, Z. Li, and S. S. Sastry, A mathematical introduction to robotic manipulation. CRC press, 2017.
[25] G. Strang, Introduction to linear algebra. Wellesley-Cambridge Press Wellesley, MA, 1993.