本研究旨在討論無人機在農業領域中實現自動巡視的應用，特別聚焦於路徑規劃的優化問題。論文提出一套整合性的系統，首先利用YOLO模型對空拍影像進行分析，以精確辨識果樹的位置資訊和果樹大小。接著，將辨識出的果樹位置作為目標點，運用基於果樹位置的TSPN(Traveling Salesman Problem with Neighborhoods)演算法進行路徑規劃。此TSPN演算法結合了ACO(Ant Colony Optimization)與2-OPT優化作為第一層處理，並使用混合式PSO(Particle Swarm Optimization)進行第二層優化，旨在求解高效且可行的巡視路徑。此外，論文還在各層中加上了不同的優化技術，包括考慮轉彎角度懲罰的改良型2-OPT演算法，以及基於連續邊界優化的方法，進一步提升路徑的平滑性並優化擊點位置，使其更符合無人機的飛行特性與任務需求。研究最後展示了本系統在不同案例的應用成果，驗證提出方法的有效性。為驗證所規劃路徑的可行性及無人機的自主控制能力，本研究更引入深度Q網路 (DQN) 模擬器。該模擬器利用DQN模型學習無人機在預定義路徑上的自主導航行為，並在虛擬環境中對自主控制策略進行快速迭代和評估，為實際無人機的部署提供理論與實驗基礎。 研究最後展示了本系統在不同案例的應用成果，驗證提出方法的有效性。

This research focuses on discussing the application of drones for automatic patrol in agriculture, with a particular emphasis on the problem of path planning optimization. The thesis proposes an integrated system that first utilizes the YOLO model to analyze aerial images, accurately identifying the location information and size of fruit trees. Subsequently, using the identified fruit tree locations as target points, a TSPN (Traveling Salesman Problem with Neighborhoods) algorithm based on fruit tree positions is applied for path planning. This TSPN algorithm combines ACO (Ant Colony Optimization) and 2-OPT optimization as the first layer of processing, and employs a hybrid PSO (Particle Swarm Optimization) for the second layer optimization, aiming to solve for efficient and feasible patrol paths. Furthermore, the thesis also incorporates different optimization techniques in each layer, including an improved 2-OPT algorithm that considers turning angle penalty, and a method based on continuous boundary optimization, to further enhance path smoothness and optimize inspection points, thereby better suiting drone flight characteristics and mission requirements. To validate the feasibility of the planned paths and the drone's autonomous control capabilities, this study also introduces a Deep Q-Network (DQN) simulator. This simulator utilizes the DQN model to learn the drone's autonomous navigation behavior along predefined paths, enabling rapid iteration and evaluation of autonomous control strategies in a virtual environment, providing theoretical and experimental foundations for actual drone deployment. Finally, the research presents the application results of this system in different case studies, validating the effectiveness of the proposed method.

目錄
第一章簡介. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
 1.1論文背景架構. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
 1.2現有技術與挑戰. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
 1.3研究目標與創新點. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
第二章文獻探討. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
 2.1 Duckiedrone. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  4
 2.1.1 Duckiedrone介紹. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  4
 2.2深度Q網路(DQN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
 2.2.1深度Q網路(DQN)介紹. . . . . . . . . . . . . . . . . . . . . . . . . . . . .  5
 2.3 UltralyticsYOLO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  6
 2.3.1 YOLOv8介紹. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  6
 2.3.2 YOLOv11介紹. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
 2.4演算法介紹. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  8
 2.4.1 Greedy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
 2.4.2 ParticleSwarmOptimization . . . . . . . . . . . . . . . . . . . . . . . .  9
 2.4.3 AntColonyOptimization . . . . . . . . . . . . . . . . . . . . . . . . . . 10
 2.4.4 GeneticAlgorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . .  11
 2.4.5 2-Opt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
 2.5路徑規劃. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .   13
 2.5.1 TSP介紹. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  13
 2.5.2 TSPN介紹. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
 2.5.3 TSPN應用. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
第三章系統實作. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
 3.1建構DuckieDrone21 . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  15
 3.1.1組裝及運行Duckiedrone . . . . . . . . . . . . . . . . . . . . . . . . . .   15
 3.1.2改寫Duckiedrone. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
 3.1.3 Duckiedrone硬體限制與起飛挑戰. . . . . . . . . . . . . . . . . . . . . . . 16
 3.1.4平台選擇的調整與考量. . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
 3.2基於深度Q網路(DQN)的自動巡視. . . . . . . . . . . . . . . . . . . . . . . . . 17
 3.2.1 DQN模型在無人機巡視任務中的應用. . . . . . . . . . . . . . . . . . . . . .  17
 3.2.2 DQN模型架構. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  19
 3.2.3目標點的提供. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  19
 3.2.4總結. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  19
 3.3訓練YOLO從空拍圖中辨識果樹. . . . . . . . . . . . . . . . . . . . . . . . . . 20
 3.3.1資料集的選用. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  20
 3.3.2訓練YOLO. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
 3.3.3 YOLOv8n. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
 3.3.4 YOLOv8s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  25
 3.3.5 YOLOv11n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
 3.3.6 YOLO訓練結論. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
 3.4基於果樹位置的TSP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  26
 3.4.1演算法的比較. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
 3.5基於果樹位置的TSPN. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
 3.5.1 TSPN實作方向. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
 3.5.2第二層PSO. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
 3.5.3結果分析. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  31
 3.5.4第二層hybrid_pso的微調. . . . . . . . . . . . . . . . . . . . . . . . . .  33
 3.6基於路徑規劃的再優化. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
 3.6.1第一層2-OPT的再優化. . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
 3.6.2第二層PSO的再優化. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
第四章結論. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  37
 4.1路徑規劃成果展示. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  37
 4.2 DQN模擬飛行路線. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  41
 4.3未來展望. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  45
參考文獻. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  47


圖目錄
1 20240520在台灣樹梅派中的價格爲新台幣18500(不含税). . . . . . . . . . . . . . . . 4
 2 YOLOv8訓練結果比較圖[9] . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
 3 YOLOv11訓練結果比較圖[10] . . . . . . . . . . . . . . . . . . . . . . . . . . 7
 4系統架構圖. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
 5 DuckieDrone21架構圖[1] . . . . . . . . . . . . . . . . . . . . . . . . . .   16
 6兩種不同分布情況的果園. . . . . . . . . . . . . . . . . . . . . . . . . . . .  20
 7使用YOLOv8n進行訓練和辨識. . . . . . . . . . . . . . . . . . . . . . . . . . . 21
 8使用YOLOv8s進行訓練和辨識. . . . . . . . . . . . . . . . . . . . . . . . . . . 22
 9使用YOLOv11n進行訓練和辨識. . . . . . . . . . . . . . . . . . . . . . . . . .  23
 10使用YOLOv11n重新進行訓練和辨識. . . . . . . . . . . . . . . . . . . . . . . . 24
 11平均圖. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  27
 12兩種不同分布情況的果園. . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
 13分別使用100、300、500個粒子對(圖12a)進行優化. . . . . . . . . . . . . . . . .  30
 14分別使用100、300、500個粒子對(圖12b)進行優化. . . . . . . . . . . . . . . . .  32
 15使用150個粒子對三種hybrid版本PSO進行優化. . . . . . . . . . . . . . . . . . .  34
 16使用新版hybrid進行分析. . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
 17優化示意圖. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  36
 18 DJI_0084優化結果. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
 19 loose優化結果. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  38
 20 loose3X2優化結果. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
 21 loose4X3優化結果. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
 22 OIP優化結果. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  39
 23 OIP2優化結果. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
 24 OIP3優化結果. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
 25 DJI_0084飛行路線圖. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
 26 loose系列果園的飛行路線圖. . . . . . . . . . . . . . . . . . . . . . . . . . 43
 27 OIP系列果園的飛行路線圖. . . . . . . . . . . . . . . . . . . . . . . . . . . 44


表目錄
1三種模型在各圖片上的樹木計數結果與辨識準確率比較. . . . . . . . . . . . . . . . .  25
 2演算法比較結果. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .  27
 3在(圖12a)情況下的演算法比較結果. . . . . . . . . . . . . . . . . . . . . . . .  31
 4在(圖12b)情況下的演算法比較結果. . . . . . . . . . . . . . . . . . . . . . . .  31

 [1] Duckietown. Duckietown Operations Manual. Accessed: Jan. 17, 2025. [Online].
 Available: https://docs.duckietown.com/daffy/opmanual-duckiedrone/intro.html
 [2] M. Refaai, et al., “[Retracted] Application of IoT-Based Drones in Precision Agri
culture for Pest Control,”Advances in Materials Science and Engineering, vol. 2022,
 2022, doi: 10.1155/2022/1160258.
 [3] K. Ye, et al., “Key Intelligent Pesticide Prescription Spraying Technologies for the
 Control of Pests, Diseases, and Weeds: A Review,”Agriculture, vol. 15, no. 1, Art.
 no. 81, Jan. 2025, doi: 10.3390/agriculture15010081.
 [4] DC. Tsouros, S. Bibi, and PG. Sarigiannidis, “A review on UAV-based applications
 for precision agriculture,”Information, vol. 10, no. 11, Art. no. 349, Nov. 2019, doi:
 10.3390/info10110349.
 [5] S. Kurkute, et al., “Drones for smart agriculture: A technical report,”International
 Journal for Research in Applied Science and Engineering Technology, vol. 6, no. 4,
 pp. 341–346, 2018.
 [6] GP. Borikar, C. Gharat, and SR. Deshmukh, “Application of drone systems for
 spraying pesticides in advanced agriculture: A Review,”in Proc. IOP Conference
 Series: Materials Science and Engineering, 2022, vol. 1259, no. 1, Art. no. 012015,
 doi:10.1088/1757-899X/1259/1/012015.
 [7] 深圳市鹏锦科技有限公司.“AI 识别病害，果树全生命周期管理｜大疆机场2进果
园.”谷瀑环保. Accessed: Mar. 5, 2025. [Online]. Available: https://nyj2010.cn.goepe.
 com/apollo/jsshow-474870.html
 [8] 林吉洋. “02 黑珍珠傳奇破滅，光電蠶食鯨吞蓮霧園.”上下游 News and Market.
 Accessed: Mar. 5, 2025. [Online]. Available: https://www.newsmarket.com.tw/blog/
 179541/
 [9] G. Jocher, A. Chaurasia, and J. Qiu. Ultralytics YOLOv8. GitHub repository. (2023).
 Accessed: Mar. 5, 2025. [Online]. Available: https://github.com/ultralytics/ultralytics
 [10] G. Jocher and J. Qiu. Ultralytics YOLO11. GitHub repository. (2024). Accessed:
 Mar. 5, 2025. [Online]. Available: https://github.com/ultralytics/ultralytics
 [11] A. Ammar, A. Koubaa, and B. Benjdira, “Deep-learning-based automated palm tree
 counting and geolocation in large farms from aerial geotagged images,”Agronomy,
 vol. 11, no. 8, Art. no. 1458, 2021, doi:10.3390/agronomy11081458.
 [12] X. Xu, J. Li, and M. Zhou, “Bi-Objective Colored Traveling Salesman Problems,”
 IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 7, pp. 6326
6337, Jul. 2022, doi:10.1109/TITS.2021.3086625.
 [13] P. He and J. Hao, “Hybrid search with neighborhood reduction for the multiple
 traveling salesman problem,”Computers and Operations Research, vol. 142, no. 2,
 Art. no. 105726, Feb. 2022, doi:10.1016/j.cor.2022.105726.
 [14] Y. Zhou, et al., “Multi-Neighborhood Simulated Annealing-Based Iterated Lo
cal Search for Colored Traveling Salesman Problems,”IEEE Transactions on In
telligent Transportation Systems, vol. 23, no. 9, pp. 16072-16082, Sep. 2022,
 doi:10.1109/TITS.2022.3147924.
 [15] J. Schmidt and S. Irnich, “New neighborhoods and an iterated local search
 algorithm for the generalized traveling salesman problem,”EURO Journal on
 Computational Optimization, vol. 10, no.1, Art. no. 100029, May. 2022,
 doi:10.1016/j.ejco.2022.100029.
 [16] X. Xu, et al., “Precedence-Constrained Colored Traveling Salesman Problem: An
 Augmented Variable Neighborhood Search Approach,”IEEE Transactions on Cyber
netics, vol. 52, no. 9, pp. 9797–9808, Sep. 2022, doi:10.1109/TCYB.2021.3070143.
 [17] Y. Tsai, “A Two-level Particle Swarm Optimization for Traveling Salesman Problem
 with Neighborhood,”Dept. of CSIE, Tamkang Univ., Tamsui, New Taipei, Taiwan,
 Tech. Rep., 2011.
 [18] I. Bercea, “Improved Bounds for the Traveling Salesman Problem with Neighbor
hoods on Uniform Disks,”2018, arXiv:1809.07159.
 [19] A. Antoniadis, et al., “A PTAS for Euclidean TSP with Hyperplane Neighborhoods,”
 2018, arXiv:1804.03953.
 [20] A. Dumitrescu and C. Tóth, “The traveling salesman problem for lines, balls and
 planes,”2013, arXiv:1303.6659.
 [21] A. Dumitrescu and J. Mitchell, “Approximation algorithms for TSP with neighbor
hoods in the plane,”2017, arXiv:1703.01640.
 [22] K. Vazquez-Santiago, et al., “Optimization Framework for Global Path Planning and
 Local Motion Planning for Robotic Welding of Multiple Large Industrial Parts,”in
 Proc. 2023 IEEE 19th Int. Conf. Automat. Sci. Eng. (CASE), 2023, pp. 1–8, doi:
 10.1109/CASE56687.2023.10260659.
 [23] C. Gao, N. Wei, and J. Walteros, “An Exact Approach for Solving Pickup-and
Delivery Traveling Salesman Problems with Neighborhoods,”Transportation Science,
 vol. 57, no. 6, pp. 1560–1580, Aug. 2023, doi:10.1287/trsc.2022.0138.
 [24] F. B. Sorbelli, S. Ghobadi, and C. M. Pinotti, “Single- and Multi-Depot Optimization
 for UAV-Based IoT Data Collection in Neighborhoods,”ACM Transactions on Sensor
 Networks, vol. 21, no. 1, Art. no. 3, pp. 1–30, 2024, doi:10.1145/3704810.
 [25] C. Peng, M. Wei, and V. Isler, “Stochastic Traveling Salesperson Problem with
 Neighborhoods for Object Detection,”in Proc. 2023 IEEE Int. Conf. Robot. Automat.
 (ICRA), 2023, pp. 3607–3613, doi:10.1109/ICRA48891.2023.10161120.
 [26] J. Mikula and M. Kulich, “Towards a Continuous Solution of the d-Visibility Watch
man Route Problem in a Polygon With Holes,”IEEE Robotics and Automation Let
ters, vol. 7, no. 3, pp. 5934-5941, Jul. 2022, doi:10.1109/LRA.2022.3159824.
 [27] J. Deckerová, J. Faigl, and V. Krátký, “Traveling Salesman Problem with
 neighborhoods on a sphere in reflectance transformation imaging scenarios,”Expert Systems With Applications, vol. 198, no. 3, Art. no. 116814, Mar. 2022,
 doi:10.1016/j.eswa.2022.116814.
[28] M. Kulich, J. Vidašič, and J. Mikula, “On the Travelling Salesman Problem with
 Neighborhoods in a Polygonal World,”in Robotics in Natural Settings, J. Cascalho et
 al., Eds. Cham: Springer, 2022, pp. 334–345, doi:10.1007/978-3-031-15226-9.
 [29] S. Mara, et al., “An adaptive large neighborhood search heuristic for the flying
 sidekick traveling salesman problem with multiple drops,”Expert Systems With Ap
plications, vol. 205, Art. no. 117647, 2022, doi:10.1016/j.eswa.2022.117647.