應用深度學習演算法萃取移動式測繪系統影像之道路標線;Deep Learning for Extracting Road Markings from Mobile Mapping System Images

NCU Institutional Repository > 研究中心 > 遙測科技碩士學位學程 > 博碩士論文 > Item 987654321/85083

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/85083

題名:	應用深度學習演算法萃取移動式測繪系統影像之道路標線;Deep Learning for Extracting Road Markings from Mobile Mapping System Images
作者:	張致維;Chang, Chih-Wei
貢獻者:	遙測科技碩士學位學程
關鍵詞:	深度學習;移動式測繪系統;道路標線辨識;高精地圖;Deep Learning;Mobile Mapping System (MMS);Lane Marking Recognition;HD Map
日期:	2020-12-24
上傳時間:	2021-03-18 17:36:53 (UTC+8)
出版者:	國立中央大學
摘要:	隨著電腦運算能力的進步與空間資訊領域的發展，政府與企業相繼投入自動駕駛系統的相關研究及開發。自動駕駛車（以下簡稱自駕車）技術往往仰賴多種車載感測器進行即時運算，然而若能利用預先建置的高精地圖提供自駕車周遭環境的空間資訊作為定位定向輔助，便能大幅減低車載電腦的運算負擔。因此，如何高效率地產製高精地圖的向量圖資是開發自駕車應用中至關重要的一個環節。本研究之目的為應用開源測繪車影像資料集與深度學習演算法，萃取移動式測繪系統影像中的道路標線，自動化地產製可供高精地圖使用的向量圖資。主要研究內容分成三部份：(1)影像語意分割(2)直接地理定位(3)道路標線修正。影像語意分割的部份，本研究使用開源測繪車影像資料集ApolloScape訓練深度學習模型，為使得訓練成果符合測試資料的環境條件，以提升影像辨識的準確度，本研究切除訓練資料中車頂入鏡部份，並移除部份逆光與雜訊之影像，再以移動式測繪系統影像測試模型的訓練成果。接著，在直接地理定位的部份，為了求取道路標線的物空間坐標資訊，本研究結合攝影測量技術與移動式測繪系統提供之相機內、外方位參數，將辨識出的道路標線標籤投影至三維空間坐標。最後，依據《道路交通標誌標線號誌設置規則》分別修正不同類別的道路標線投影的成果，產出道路標線的向量圖資。研究成果顯示，本研究提出的深度學習演算法可以有效地辨識出移動式測繪系統影像中的道路標線，而且後續的道路標線修正步驟中亦能修正錯誤辨識與缺漏，產出正確有效的道路標線向量圖資。 ;With the growing computing capacity and the development in the field of Geoinformation, governments and enterprises have actively invested in the research and development of autonomous driving systems. Most autonomous driving technologies rely on multiple on-board sensors for real-time operation. However, the implementation of pre-built High-Definition map (HD Map) can provide detail information of the surrounding environment for self-driving cars. For instance, positioning and orientation, which can significantly reduce the computational burden of on-board computers. Therefore, efficiently producing data layers for HD maps is an important step in the development of self-driving car technology. The purpose of this research is to apply the open source street scene image dataset and deep learning algorithms to extract lane markings from the Mobile Mapping System (MMS) images, and to automatically create the data layer for HD maps. The proposed scheme consists of three parts: (1) semantic segmentation, (2) direct georeferencing, (3) lane marking correction. In the semantic segmentation part, this study uses the open source self-driving car image dataset ApolloScape to train the deep learning model. In order to increase the accuracy of image segmentation, this study crop the training images and remove the surveying car-roof area which is not shown in the test data. Also, it removes parts of training images with high illumination and noise. Afterward, in order to get the coordinates of lane markings, this study uses photogrammetry method and orientation parameters of camera to derived the positional information of the lane markings. Finally, produce the HD Map data layer according to the regulations issued by Ministry of Transportation and Communications, Taiwan. The experimental results show that the proposed method can effectively identify the road markings in the MMS images. In addition, the developed method can further correct the mis-classified and missing parts in the subsequent lane marking correction steps, and produce reliable road marking data layer for High -Definition Map.
顯示於類別:	[遙測科技碩士學位學程] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	184	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....