請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/77590
標題: | 基於卷積式類神經網路之台灣交通號誌自動偵測與辨識系統設計 Automatic Taiwan Traffic Sign Detection and Recognition System Design Based on Convolutional Neural Networks |
作者: | Yi-Hsuan Liu 劉懿萱 |
指導教授: | 闕志達 |
關鍵字: | 卷積式類神經網路,深度學習,自動駕駛車,機器視覺,目標偵測, Convolutional Neural Network,Deep-learning,Self-driving cars,Computer Vision,Object Detection, |
出版年 : | 2018 |
學位: | 碩士 |
摘要: | 隨著市場對於節省人力、自動化在工業、居家環境、交通載具、醫療照護、娛樂等的需求量增加,使卷積式類神經網路近幾年來的發展也突飛猛進。物體辨識與目標偵測領域吸引了無數學者的目光,也獲得了廣泛的應用。而令人最期待的應用非自駕車領域莫屬,許多學術研究單位以及私人企業紛紛投入研究。
目標偵測在自駕車領域的應用相當廣泛,交通號誌偵測與辨識是其中一項很重要的部分。並非路面上的每一個交通號誌都在電子地圖上被記載,許多交通號誌會因為路況的改變而臨時搭建或拆除,因此電子地圖無法完全替代自駕車上的目標偵測系統。 本文利用目標偵測卷積式類神經網路在用於台灣交通號誌自動偵測與辨識系統。我們建置並標記台灣交通號誌目標偵測資料庫 ;為了能夠進一步提升目標偵測對於小目標的正確率,本文系統基於Faster RCNN的Caffe平台上,提出兩種新型特徵提取架構,分別基於VGG-16與ZF net 。除外,為了能夠減少重新標記圖片新座標位置的時間與人力,我們提出能夠找出使用變形數據擴增後圖像上目標的新座標位置的演算法。本文發展一個完整的交通號誌自動偵測與辨識系統,並建置一個圖形化使用者介面。 In the past several years, image classification and object detection using Convolutional Neural Networks (CNN) has attracted enormous attention and found numerous applications. A recent exciting emerging field is autonomous vehicles, in which many organizations from academia and industry have made great strides. An important task in self-deriving vehicles is to automatically detect and recognize traffic signs, as oftentimes the temporary signs are not registered even in the most sophisticated electronic maps. Many traffic signs are temporarily built or removed due to the changes of road conditions. Therefore, the electronic maps system cannot replace the object detection system on the self-driving cars. Our work aims to apply CNN model to achieve automatic detection and recognition of traffic signs on Taiwanese road scenes. In this thesis, we built and labeled a Taiwan traffic sign object detection database. Our proposed approach is based on the Caffe package of the Faster RCNN model. In order to improve the accuracy of detecting small objects, we proposed two new feature extraction architectures based on VGG-16 and ZF-net. Furthermore, in order to save extra time and efforts to relabel new coordinates of images after using deformable data augmentation on object detection, we proposed an algorithm to find new coordinates using deformable data augmentation. Finally, we also developed a Graphic User Interface (GUI) for the proposed automatic traffic sign detection and recognition system. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/77590 |
DOI: | 10.6342/NTU201802291 |
全文授權: | 未授權 |
顯示於系所單位: | 電子工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-107-R04943142-1.pdf 目前未授權公開取用 | 7.39 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。