請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/71288
標題: | 基於深度學習之強健性即時道路標記偵測系統 Deep Learning-based Robust Real-time Road Marking Detection System |
作者: | Xing-Yu Ye 葉興宇 |
指導教授: | 傅立成(Li-Chen Fu) |
關鍵字: | 深度學習,道路標記,即時物件偵測,物件分類, Deep Learning,Road Marking,Real-time Object Detection,Object Classification, |
出版年 : | 2018 |
學位: | 碩士 |
摘要: | 近年來,自動駕駛技術以及先進駕駛輔助系統變的越來越流行,其性能也越來越可靠。對於上述系統來説,最重要的就是對道路環境的偵測和理解。路面標記對於駕駛員以及行車輔助設備起到了非常重要的引導作用。但是路面標記會受到不同天氣、光照及視角的影響導致難以偵測。傳統的路面標記偵測方法通常使用固定的門檻值參數,因此無法應對現實狀況中各種各樣的情況。爲瞭解決這個難題,基於深度學習的即時物件偵測框架例如Single Shot Detector (SSD) 和You Only Look Once (YOLO)比較適合處理這個問題。但這些基於深度學習的方法都需要大量的資料來進行訓練,而目前網路上並沒有合適的路面標記資料庫可以訓練這些偵測系統。此外,這些偵測系統容易將高度扭曲的路面標記辨識成錯誤的類別。如何平衡準確率以及召回率對這些偵測系統來説也是一個難題。
本論文提出了一種包含兩個階段的深度學習系統來解決現有偵測架構中難以辨識高度扭曲的路面標記以及難以平衡召回率以及準確率的問題,本系統可在各種環境下即時準確的偵測地面標記。本論文還建立了一個新的路面標記偵測與分類的基準,收集的資料庫包含11800張高解析度影像。這些影像是在不同的時間和天氣狀況下拍攝於臺北的道路,並且手工標記出13種列別的候選框。實驗證明本論文提出的架構在路面標記偵測的任務中超過其他物件偵測架構。 In recent years, Autonomous Driving Systems (ADS) and Advanced Driver Assistance Systems (ADAS) become more and more popular and reliable. It is important for the above systems to understand the road environments. Road markings are important for drivers and driver assistance systems to better understand the road environment. But the detection of road markings will be influenced by various illumination, weather conditions and angles of view. Most traditional road marking detection methods use fixed threshold to detect the road marking, which is not robust enough to handle various situations in the real world. To solve this problem, deep learning-based real-time detection framework such as Single Shot Detector (SSD) and You Only Look Once (YOLO) is suitable for this task. However, these deep learning-based methods are data-driven but there is no suitable public road marking dataset for us to train the network. Besides, these detection frameworks usually struggle with classifying highly distorted road markings. Balancing the precision and recall is also a challenging task for these detection frameworks. In this thesis, we propose a two-stage deep learning-based system to tackle highly distorted road marking detection and to balance precision and recall of the detection framework. Our system can perform real-time road marking detection under diverse circumstances. We also create a new benchmark for road marking detection and classification tasks. The dataset consists of 11800 high resolution images captured at different times under various weather conditions in Taipei. The images are manually labeled with object bounding boxes and 13 classes. The experimental result shows that the proposed system outperforms other real-time detection framework and Faster R-CNN in road marking detection task. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/71288 |
DOI: | 10.6342/NTU201801616 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-107-1.pdf 目前未授權公開取用 | 4.17 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。