Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/8462
Title: | 應用三維深度學習於光達點雲資料之語義分類 Semantic Classification of LiDAR Point Cloud with 3D Deep Learning |
Authors: | Zong-Yi Zhuang 莊宗易 |
Advisor: | 徐百輝(Pai-Hui Hsu) |
Keyword: | 光達,點雲資料,三維深度學習,語義分類, LiDAR,point cloud,3D deep learning,semantic classification, |
Publication Year : | 2020 |
Degree: | 碩士 |
Abstract: | 為提升光達點雲分類之效能及準確性,本研究利用三個表現良好的深度學習模型PointNet、PointNet++及KPConv於自行標註的光達資料進行點雲語義分類的實驗。在分類程序上,不同於過去需先分類出地面/非地面點的處理模式,也不需由人工設計特徵以分類器進行點雲分類,而是由模型自動進行特徵的萃取及分類。此外,本研究透過分析資料的正規化方式及加入幾何特徵的資料組合,進行模型最佳化參數的評估及調整,找出一個最佳的策略以利後續應用。 為驗證不同輸入特徵組合的深度學習模型於各式光達資料之適用性,實驗中分別以ALS及MLS資料進行自動化地物分類,除了以各項準確度指標評估及可視化分析外,同時也與傳統機器學習演算法隨機森林與商用軟體LiDAR360的分類成果進作比較。實驗成果顯示,在兩類型的資料中,三維深度學習的方法皆優於傳統方法及商用軟體。其中幾何特徵及強度資訊的加入對PointNet的分類成效有顯著的提升,對採用階層式架構以鄰域球萃取局部特徵的PointNet++也有些許幫助;KPConv則透過點卷積的方式於原始點雲坐標學習到更具細節的局部特徵。而對於各模型來說,不管是對ALS還是MLS的場景應用,高程扮演著一個重要的特徵,尤其在ALS場景中更是明顯。最終由KPConv分別在ALS及MLS資料中獲得最佳的分類成果,但其模型較複雜且計算時間長;PointNet++則在準確度與計算效率上都在有不錯的表現;而PointNet透過加入幾何特徵,達到不亞於PointNet++的表現,且其模型更簡單、計算更快速。 To improve the efficiency and accuracy of LiDAR point cloud classification, three crucial deep learning models, PointNet, PointNet++ and KPConv are utilized in this study to experiment on semantic classification of point cloud for LiDAR data labeled manually. In the classification procedure, the processing mode is different from the past that needs to classify ground/non-ground points first, or design features manually to classify the point cloud. However, our model can extract and classify the features automatically. In addition, the research finds an optimal strategy for subsequent application, through the analysis of the data normalization, the combination of additional geometric features, and the model parameters optimization and adjustment. In order to verify the applicability of the used deep learning model with different input feature combinations to various types of LiDAR data, airborne laser scanning and mobile laser scanning data sets were applied to automatic land feature classification in this study. Besides the classification indicators evaluation and visual analysis, the classification results were compared with regular machine learning algorithms, random forest and commercial software, LiDAR360. The results show that in the two types of data, the result of 3D deep learning are superior to traditional methods and commercial software, in addition, the addition of geometric features and intensity information has significantly improved the classification results of PointNet, and it is also a little help for PointNet++ that uses a hierarchical structure to extract local features with neighborhood balls; KPConv uses point convolution to learns more detailed local features from the original point cloud coordinates. In addition, for each model, elevation plays an important role, especially in ALS data. Finally, KPConv obtained the best classification results in both ALS and MLS data, but it is more complicated and time consuming; PointNet++ has a good performance both in accuracy and efficiency; and PointNet achieves performance comparable to PointNet++ by adding geometric features, and it is simpler and more efficient. Different models and strategies can be adopted according to different application levels. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/8462 |
DOI: | 10.6342/NTU202001331 |
Fulltext Rights: | 同意授權(全球公開) |
metadata.dc.date.embargo-lift: | 2025-07-09 |
Appears in Collections: | 土木工程學系 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
U0001-0607202009292100.pdf Until 2025-07-09 | 9.43 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.