本研究旨在講解、改善無序點雲降噪與銳化的步驟。為了後續的應用例如：應用於無人載具的同時定位與地圖建構(Simultaneous localization and mapping, SLAM) ，或是逆向工程、點雲重建、物體辨識、路徑規劃等，點雲必須有一定的清晰度，然而礙於環境複雜度以及掃描裝置的質量，常常使得點雲帶有雜訊、異常點，若是照樣執行後續的應用，將會導致定位不精確、重建模型誤差、無法準確辨識物體等錯誤，因此前處理成為必要的步驟。

除了清除噪點，強化點雲的特徵也是我們的目標，其中會使用到的法向量估計為本論文的一大重點，有精確的法向量即可以更準確的描述物體形狀。處理有序點雲時，我們可以透過已知的法向量使用Bilateral Filter 將點雲銳化；處理無序點雲時，就需要先估算法向量，而主成分分析(Principal Components Analysis, PCA) 為估算法向量以及曲率最通用的做法。對PCA 與Bilateral Filter 來說，影響兩者效果最大的原因為鄰域點的選擇， PCA 對異常值尤其敏感，在真實場景中，點雲可能因為深度誤差而產生殘影，使法向量估計變得更困難，因此除了消去異常值的影響，用甚麼策略去選取適當的鄰域點變得相當重要，選取過多鄰點可能會導致過度平滑使特徵消失，選取不足會導致法向量凌亂。

本論文將會介紹降噪與銳化相關演算法，演示無序點雲與有序點雲分別使用這些演算法的差異，分析各方法之優缺點並做改良，接著分析法向量估算較容易出錯的地方，例如較小的特徵、轉角以及邊緣。本論文最後提出了一個基於 PCA 的點雲降噪和銳化流程，該流程中使用的邊緣偵測方法在高雜訊情況下仍然表現出良好的強健性，能以不同的策略移動不同特性的點雲，有助於在銳化過程中保留特徵，並通過分別使用 RGBD 有序點雲和無序點雲來驗證了本研究方法的有效性。

Due to its higher accuracy and decreased costs, LIDAR applications have been increasing, such as in Simultaneous Localization and Mapping (SLAM) for UAVs, automobiles, and surface reconstruction, reverse-engineering, territory remote sensing. However,in complex environments or with low-accuracy devices, the resulting point cloudcan contain many noisy points or outliers, leading to a thicker point distribution and difficultyin object recognition. For subsequent application such as point cloud recostruction,3D printing, and reverse engineering , the pre-processing is necessary.

In this research, our goal is to strike a balance between noise reduction and feature preservation, avoiding over-smoothing or over-sharpening during the procedure. We focus on developing comprehensive processes for denoising and sharpening both organized and unorganized point clouds, analyzing the areas where normal estimation is more likely to be erroneous, such as smaller features, corners, and edges. This step incorporate edge detection technique. Finally, based on Principal Components Analysis (PCA) as the method for estimating normal vectors and Bilateral Filter as the option for sharpening. Each algorithm is discussed and improved, and strategies for adaptive scaling will be introduced. In the end of our research, we specifically focus on utilizing RGBD Kinect cameras to address these challenges and explore whether the known ray directions and the properties of 2D images can be leveraged to achieve effective denoising of point clouds.
We will evaluate the performance of these algorithms on both unorganized and organized RGBD point clouds, analyzing the strengths and weaknesses of each method and proposing our solutions. In the final part of this study, we propose a PCA-based approach for denoising and sharpening point clouds. The edge detection method employed in this approach demonstrates robustness even in the presence of high noise, allowing for the manipulation of point clouds with different characteristics using various strategies. This contributes to preserving features during the sharpening process. The effectiveness of our proposed method is validated through the use of both organized and unorganized RGBD point clouds.

[1] J. Sanchez, F. Denis, D. Coeurjolly, F. Dupont, L. Trassoudaine, and P. Checchin, “Robust normal vector estimation in 3D point clouds through iterative principal component analysis,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 163, pp. 18–35, 2020.

[2] X.-F. Han, J. S. Jin, M.-J. Wang, W. Jiang, L. Gao, and L. Xiao, “A review of algorithms for filtering the 3D point cloud,” Signal Processing: Image Communication, vol. 57, pp. 103–112, 2017.

[3] Y. Ioannou, B. Taati, R. Harrap, and M. Greenspan, “Difference of normals as a multi-scale operator in unorganized point clouds,” in 2012 Second International Conference on 3D Imaging, Modeling, Processing, Visualization & Transmission, pp. 501–508, 2012.

[4] S. Paris, “A gentle introduction to bilateral filtering and its applications,” in ACM SIGGRAPH 2007 courses, pp. 3–es, 2007.

[5] Y. Lipman, D. Cohen-Or, D. Levin, and H. Tal-Ezer, “Parameterization-free projection for geometry reconstruction,” ACM Transactions on Graphics (TOG), vol. 26, no. 3, pp. 22–es, 2007.

[6] L. Zhou, G. Sun, Y. Li, W. Li, and Z. Su, “Point cloud denoising review: from classical to deep learning-based approaches,” Graphical Models, vol. 121, p. 101140, 2022.

[7] X.-F. Han, J. S. Jin, M.-J. Wang, and W. Jiang, “Guided 3D point cloud filtering,” Multimedia Tools and Applications, vol. 77, pp. 17397–17411, 2018.

[8] S. Zhang, S. Cui, and Z. Ding, “Hypergraph spectral analysis and processing in 3D point cloud,” IEEE Transactions on Image Processing, vol. 30, pp. 1193–1206, 2020.

[9] E. Mattei and A. Castrodad, “Point cloud denoising via moving RPCA,” in Computer Graphics Forum, vol. 36, pp. 123–137, Wiley Online Library, 2017.

[10] C. C. Jia, C. J. Wang, T. Yang, B. H. Fan, and F. G. He, “A 3D point cloud filtering algorithm based on surface variation factor classification,” Procedia Computer Science, vol. 154, pp. 54–61, 2019.

[11] R. B. Rusu, “Semantic 3D object maps for everyday manipulation in human living environments,” KI-Künstliche Intelligenz, vol. 24, pp. 345–348, 2010.

[12] L. Du, Edge Detection in 3D Point Clouds for Industrial Applications. University of Toronto (Canada), 2020.

[13] J. Digne and C. De Franchis, “The bilateral filter for point clouds,” Image Processing On Line, vol. 7, pp. 278–287, 2017.

[14] G. Kurillo, E. Hemingway, M.-L. Cheng, and L. Cheng, “Evaluating the accuracy of the Azure Kinect and Kinect v2,” Sensors, vol. 22, no. 7, p. 2469, 2022.

[15] Z. Yu, T. Wang, T. Guo, H. Li, and J. Dong, “Robust point cloud normal estimation via neighborhood reconstruction,” Advances in Mechanical Engineering, vol. 11, no. 4, p. 1687814019836043, 2019.

[16] H. Chen and J. Shen, “Denoising of point cloud data for computer-aided design, engineering, and manufacturing,” Engineering with Computers, vol. 34, pp. 523–541, 2018.