本論文旨在提出一種新的無人機跑道辨識方法，藉由結合深度學習以及傳統電腦視覺技術，來彌補因尺寸以及操作成本限制而無法搭載完整儀器降落系統的缺陷。透過結合深度學習法和電腦視覺技術，剛好可發揮兩項技術的長處：深度學習對偵測複雜場景的穩健性，和電腦視覺對線段偵測的準確性。
具體來說，本研究使用YOLOv7搭配影像處理來實現跑道偵測以及中線偵測。在跑道偵測的部分首先需蒐集足夠多的數據來訓練YOLOv7模型，本研究使用了各種機場的跑道圖片以及實驗跑道的空拍照，在經過數據增強處理後，總共得到1068張訓練圖片，而訓練完的模型準確度則來到了0.95。而中線偵測的部分則使用HSV色彩空間轉換和Canny邊界偵測法，將跑道的兩邊邊界提取出來，並且對兩條邊界取平均即可得到跑道中線，在得到跑道中線後即可計算無人機與跑道的相對位置，並且將數據輸出，以利後續對姿態進行調整來達到降落的目的。而在研究最後，我們也對實驗所得到的測量值以及手動標記中線的真實值進行誤差測量，得到了偏移角度的誤差皆在6度以下，且偏移量的誤差皆在30個pixels之內的結果。

This thesis aims to propose a novel UAV runway recognition method by combining deep learning and traditional computer vision techniques. Due to the limitations of UAVs that cannot be equipped with a complete instrument landing system due to size and operational cost constraints. By integrating deep learning and computer vision, the method leverages the strengths of both technologies: the robustness of deep learning in detecting complex scenes and the accuracy of computer vision in line detection.
Specifically, this study employs YOLOv7 along with image processing to achieve runway detection and centerline detection. For runway detection, it is essential to collect sufficient data to train the YOLOv7 model. This study uses images of various airport runways and aerial photos of experimental runways. After data augmentation, a total of 1068 training images were obtained, and the trained model achieved an accuracy of 0.95. For centerline detection, HSV color space conversion and Canny edge detection are used to extract the edges of the runway. The average of the two edges is then taken to obtain the runway centerline. Once the centerline is obtained, the relative position of the UAV to the runway can be calculated, and the data can be output to facilitate subsequent attitude adjustments for landing. Finally, the study measures the errors between the experiment values and the true values of the manually marked centerline. The results show that the deviation angle error is within 6 degrees, and the displacement error is within 30 pixels.

第一章 緒論 1
1.1研究動機 1
1.2文獻回顧  2
1.3研究方法  3
第二章 跑道偵測演算法之簡介 4
2.1卷積神經網路  4
2.2 YOLOv7 7
2.3 YOLOv7損失函數 10
2.3.1 CIoU Loss(Complete Intersection Union Loss) 11
2.3.2 Focal Loss 12
2.4數據增強(Data Augmentation) 14
第三章 跑道中線檢測法 15
3.1色彩空間 15
3.1.1 RGB色彩空間 15
3.1.2 HSV色彩空間 16
3.1.3色彩空間轉換 17
3.2 Canny Edge Detector 18
3.3 Hough Transform 19
第四章 實驗架構與結果 21
4.1實驗載具 22
4.2建立資料集 23
4.3環境建置 24
4.4開始訓練 25
4.5 降落階段之中線偵測 31
4.5.1 影像處理 31
4.5.2 跑道中線偵測 33
4.5.3計算偏移角和偏移量 34
4.5.4誤差驗證 36
第五章 結論 40
參考文獻 41

圖目錄
圖1-1：ILS(Instrument Landing System) 1
圖1-2：跑道辨識流程 3
圖2-1卷積神經網路結構 4
圖2-2卷積層示意圖 5
圖2-3 max pooling示意圖 5
圖2-4 Fully Connected Layer示意圖 6
圖2-5 YOLO structure 7
圖2-6 Compound model scaling (複合模型縮放) 8
圖2-7 Extended efficient layer aggregation networks (擴展的高效層聚合網路) 8
圖2-8 Model re-parameterized (模型重參數化) 8
圖2-9 Dynamic label- assignment strategy (動態標籤分配策略) 9
圖2-10 YOLO series benchmark 9
圖2-11 IoU示意圖 11
圖2-12數據增強應用 14
圖3-1 RGB色彩空間模型 15
圖3-2 HSV色彩空間模型 16
圖3-3非極大值抑制 18
圖3-4 rho-theta示意圖 19
圖3-5坐標系轉換示意圖 20
圖3-6多點霍夫換示意圖 20
圖4-1：實驗架構流程圖 21
圖4-2：CMOS鏡頭 22
圖4-3：原跑道空照圖vs數據增強 23
圖4-4：labelme標註頁面 24
圖4-5：Anaconda環境示意圖 24
圖4-6：YOLOv7模型訓練曲線圖 25
圖4-7：TP/FP/FN/TN關係圖 27
圖4-8：YOLOv7預測結果 29
圖4-9：高度200公尺進場模擬 30
圖4-10：實驗跑道 31
圖4-11：HSV色彩空間轉換 31
圖4-12：提取出的跑道邊界 32
圖4-13：Canny edge detector示意圖 32
圖4-14：使用霍夫轉換提取出跑道邊界 33
圖4-15：藉由邊界平均求得中線 34
圖4-16：偏移角度關係圖 34
圖4-17：機身左偏示意圖 35
圖4-18：機身右偏示意圖 35
圖4-19：輸出數據 35
圖4-20：每幀畫面輸出 36
圖4-21：手動中線標記示意圖 36
圖4-22：角度誤差折線圖 39
圖4-23：偏移量誤差折線圖 39
表目錄
表4-1：實驗載具規格表 22
表4-2：設定參數 25
表4-3：YOLOv7模型訓練數據 26
表4-4：P、R-curve & PR-curve & F1-curve 27
表4-5：角度誤差表 37
表4-6：偏移量誤差表 38

[1]
周威宇， 機器視覺於跑道辨識以引導無人機近場與降落》碩士論文。 台南：國立成功大學航空太空工程學系碩士班，2015年。
[2]
張展維， 無人飛機影像辨識輔助降落系統》碩士論文。桃園：元智大學機械工程學系碩士班，2012年。
[3]
黃佩瑩， 以影像識別方法進行跑道偵測於固定翼無人飛機之實現》碩士論文。台南：國立成功大學航空太空工程學系碩士班，2009年。
[4]
R. A. Amit and C. K. Mohan, “A Robust Airport Runway Detection Network Based on R-CNN Using Remote Sensing Images,” IEEE Aerosp. Electron. Syst. Mag., vol. 36, no. 11, pp. 4–20, Nov. 2021, doi: 10.1109/MAES.2021.3088477.
[5]
H. Mirhajianmoghadam and B. B. Haghighi, “EYNet: Extended YOLO for Airport Detection in Remote Sensing Images,” Mar. 26, 2022, arXiv: arXiv:2203.14007. Accessed: May 15, 2024. [Online]. Available: http://arxiv.org/abs/2203.14007
[6]
J. Du, “Understanding of Object Detection Based on CNN Family and YOLO,” J. Phys.: Conf. Ser., vol. 1004, p. 012029, Apr. 2018, doi: 10.1088/1742-6596/1004/1/012029.
[7]
吳宥霖， 基於YOLOv7與深度影像於樓梯間跌倒事件的偵測》碩士論文。高雄：國立高雄科技大學資訊管理系碩士班，2023年。
[8]
C.-Y. Wang, A. Bochkovskiy, and H.-Y. M. Liao, “YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” Jul. 06, 2022, arXiv: arXiv:2207.02696. Accessed: Jul. 30, 2024. [Online]. Available: http://arxiv.org/abs/2207.02696
[9]
J. Redmon, S. Divvala, R. Girshick, and A. Farhadi, “You Only Look Once: Unified, Real-Time Object Detection,” May 09, 2016, arXiv: arXiv:1506.02640. Accessed: May 07, 2024. [Online]. Available: http://arxiv.org/abs/1506.02640
[10]
L. G. Delphina and V. Naidu, “Detection of Airport Runway Edges using Line Detection Techniques”.
[11]
Z. Zheng, P. Wang, W. Liu, J. Li, R. Ye, and D. Ren, “Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression,” Nov. 19, 2019, arXiv: arXiv:1911.08287. Accessed: May 17, 2024. [Online]. Available: http://arxiv.org/abs/1911.08287
[12]
T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal Loss for Dense Object Detection,” Feb. 07, 2018, arXiv: arXiv:1708.02002. Accessed: May 17, 2024. [Online]. Available: http://arxiv.org/abs/1708.02002
[13]
“How Instrument Landing System (ILS) works | Instrument landing system, Aviation education, Student pilot,” Pinterest. Accessed: Jun. 15, 2024. [Online]. Available: https://www.pinterest.co.uk/pin/instrument-landing-system-ils-ils-is-the-only-navigational-aid-for-landing-and-final-approach--681521356082239542/
[14]
D. Kalita, “Basics of CNN in Deep Learning,” Analytics Vidhya. Accessed: Jun. 15, 2024. [Online]. Available: https://www.analyticsvidhya.com/blog/2022/03/basics-of-cnn-in-deep-learning/
[15]
Y. James, “[資料分析&機器學習] 第5.1講：卷積神經網絡介紹(Convolutional Neural Network),” JamesLearningNote. Accessed: Jun. 15, 2024. [Online].Available:https://medium.com/jameslearningnote/%E8%B3%87%E6%96%99%E5%88%86%E6%9E%90-%E6%A9%9F%E5%99%A8%E5%AD%B8%E7%BF%92-%E7%AC%AC5-1%E8%AC%9B-%E5%8D%B7%E7%A9%8D%E7%A5%9E%E7%B6%93%E7%B6%B2%E7%B5%A1%E4%BB%8B%E7%B4%B9-convolutional-neural-network-4f8249d65d4f
[16]
J. Chou, “Distance-IOU Loss : Faster and Better Learning for Box Regression,” Medium. Accessed: Jun. 15, 2024. [Online]. Available: https://medium.com/@jacksonchou/distance-iou-loss-faster-and-better-learning-for-box-regression-9df8fc627e8
[17]
V. Popov, M. Ostarek, and C. Tenison, “Practices and pitfalls in inferring neural representations,” NeuroImage, vol. 174, pp. 340–351, Jul. 2018, doi: 10.1016/j.neuroimage.2018.03.041.
[18]
“OpenCV: Canny Edge Detection.” Accessed: Jun. 15, 2024. [Online]. Available: https://docs.opencv.org/4.x/da/d22/tutorial_py_canny.html
[19]
天道酬勤, “Hough Transform,” Medium. Accessed: Jun. 15, 2024. [Online]. Available: https://medium.com/@bob800530/hough-transform-cf6cb8337eac
[20]
A. Idris, “Confusion Matrix,” Medium. Accessed: Jun. 15, 2024. [Online]. Available: https://medium.com/@awabmohammedomer/confusion-matrix-b504b8f8e1d1