臺灣博碩士論文加值系統

傳統影像辨識是透過顏色判斷，先將RGB色彩模型轉換成HSB色彩模型，辨識其影像中的色彩後再辨識其形狀，最終才判斷號誌。而傳統的影像辨識方式會因為影像來源的解析度不足，或是夜晚眩光、抑或是有物體遮擋等因素造成辨識率下降。因此，本研究旨在提升因上述原因而導致的辨識率下降問題。本研究提出的方法為蒐集行車紀錄器的行車影片與其他車輛行進中所錄製之影片，透過轉檔程式將影片轉為圖片，再來使用繪圖工具將部分號誌進行遮擋，透過不同方位的遮擋與原圖進行訓練後再與單只有以原圖做訓練的模型做比對。本研究所提出交通標誌辨識系統主要以交通號誌燈、速限標誌、警告標誌、以及測速照相，當心兒童這幾種在路上須謹慎注意之交通號誌。實現該系統的方法，首先在Ubuntu18.04系統上以特徵框選，YOLOv4損失函數與邊框預測演算法模擬驗證，並以原圖、原圖與不完全圖片這兩套資料作為訓練，再以訓練後的模型進行結果驗證，可得知增加了不同方位的遮擋所訓練的模型相較只以原圖訓練的模型在一般原圖辨識的情況下高出10.13%的辨識率，而在不完全圖片的情形下則高出14.51%的辨識率。

Traditional image recognition converts the RGB color model to the HSB color model to identify colors and recognize shapes, enabling the system to determine traffic signals. However, this conventional approach is prone to decreased accuracy due to factors like low-resolution sources, nighttime glare, and obstructions. This research aims to address these challenges and improve recognition rates.
Our study proposes an innovative method involving dashcam footage and recordings from other vehicles, which are converted to images. To enhance recognition, specific sections of traffic signals are intentionally obscured using advanced drawing tools. Training utilizes a combination of these obscured images and original unobstructed images from different angles.
The presented traffic sign recognition system focuses on critical elements like signal lights, speed limit signs, warnings, speed cameras, and signs for caution in areas with children. Implementation involves a comprehensive methodology with feature selection, the YOLOv4 loss function, and bounding box prediction algorithms. Experimental validation is conducted on an Ubuntu 18.04 system using two sets of training data: original unobstructed images and a mix of unobstructed and partially obscured images. Results show significant improvements, with the model trained on diverse occluded images outperforming the solely unobstructed model by 10.13% in standard scenarios and 14.51% in dealing with partial obstructions.

摘要.............i
Abstract.............ii
誌謝.............iii
目錄.............iv
表目錄.............vi
圖目錄.............vii
第一章 緒論............. 1
1.1 研究背景.............1
1.2 文獻回顧.............2
1.3 研究目的.............2
1.4 研究方法.............3
1.5 章節概要.............3
第二章 理論與技術.............4
2.1 深度學習理論.............4
.2.2.1 CNN介紹.............4
.2.1.2 R-CNN介紹.............5
.2.1.3 Fast R-CNN與Faster R-CNN介紹............. 6
2.2 軟體介紹.............7
.2.2.1 LabelImg軟體.........7
.2.2.2 OpenCV軟體...........7
.2.2.3 Darknet.............7
2.3 YOLO系統.............8
.2.3.1 YOLOv1.............8
.2.3.2 YOLOv2.............8
.2.3.3 YOLOv3.............8
.2.3.4 YOLOv4.............9
第三章 系統架構.............10
3.1 系統架構與流程.............10
3.2 使用繪圖工具對交通號誌進行特徵進行遮擋.............13
3.3 使用LabelImg對交通標誌特徵框選與標籤命名.............18
3.4 訓練環境設定.............28
3.6 learning_rate (學習率) 決定.............31
第四章 系統模擬與分析.............36
4.1 交通號誌場景介紹.............36
4.2 辨識模擬結果.............37
4.3 不同閾值之辨識率.............47
4.4 討論與分析.............51
第五章 結論............. 52
5.1 總結............. 52
5.2 未來研究方向.............52
參考文獻.............53
附錄一........................55
Extended Abstract............61

[1]內政部警政署交通事故統計，2023，“道路交通事故發生件數”。線上檢索日期：2023年5月19日。
網址：https://www.npa.gov.tw/ch/app/data/list?module=wg073&id=2299
[2]林永鑫，2018，“植基於卷積神經網路之高效能車牌辨識系統” 國立師範大學電機工程學系研究所碩士論文。
[3]林伯臻，2010，“交通號誌燈之快速偵測與辨識” 國立中正大學電機工程研究所碩士論文
[4]蔡凱勛，2018，“基於Python與樹莓派之交通標誌辨識系統研究” 國立虎尾科技大學電機工程研究所碩士論文
[5]馬聖珉，2006，“一個監督式學習與非監督式學習技術應用於多國語言文件探勘之比較研究” 國立高雄科技大學電機工程研究所碩士論文
[6]林育新，2019，“基於深度學習的樹莓派系統之實現-以即時交通號誌辨識系統為例” 國立虎尾科技大學電機工程研究所碩士論文
[7]R. Girshick, J. Donahue, T. Darrell and J. Malik, "Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation," 2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus, OH, 2014, pp. 580-587
[8]J. Redmon, S. Divvala, R. Girshick and A. Farhadi, "You Only Look Once: Unified, Real-Time Object Detection," 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, 2016, pp. 779-788.
[9]J. Redmon and A. Farhadi, "YOLO9000: Better, Faster, Stronger," 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, 2017, pp. 6517-6525.
[10]Ian Goodfellow，2019 "Deep Learning" 陳仁和譯 碁峰出版社
[11]Z. He, F. Nan, X. Li, S. Lee and Y. Yang, "Traffic sign recognition by combining global and local features based on semi-supervised classification," in IET Intelligent Transport Systems, vol. 14, no. 5, pp. 323-330, 5 2020.
[12]H. Zhang et al., "Real-Time Detection Method for Small Traffic Signs Based on Yolov3," in IEEE Access, vol. 8, pp. 64145-64156, 2020.
[13]Fast R-CNN，
https://notesonai.com/FastRCNN
[14]LabelImg，
(https://github.com/heartexlabs/labelImg)
[15]OpenCV，
(https://opencv.org/)
[16]Darknet，
(https://github.com/pjreddie/darknet)
[17]Alexey Bochkovskiy, "YOLOv4: Optimal Speed and Accuracy of Object Detection," 2020 in arxiv