隨著物聯網技術的持續發展，邊緣運算裝置因其低耗能與容易部署的特點，逐漸成為商業化應用的重要推動力。本研究主要探討並比較不同版本的輕量化YOLO (You Only Look Once) 演算法，具體為各模型在視訊鏡頭車輛追蹤系統的效能表現，其中包含YOLOv8n、YOLO11n及YOLO12n。本研究亦深入評估邊緣運算設備與傳統高效能設備在能耗、部署便利性上的差異，測試平台包括傳統高效能運算設備 (RTX 4060 Laptop) 以及邊緣運算設備 (Raspberry Pi 5與NVIDIA Jetson Orin Nano Super)。透過專注於邊緣運算優化的YOLO輕量版本分析，期望為未來智慧交通及智慧停車引導系統的發展提供有價值的參考依據。

With the continuous development of Internet of Things (IoT) technology, edge computing devices have gradually become a key driving force for commercial applications due to their low power consumption and ease of deployment. This study primarily investigates and compares various versions of the lightweight YOLO (You Only Look Once) algorithm, focusing on the performance of each model in a video-based vehicle tracking system, including YOLOv8n, YOLO11n, and YOLO12n. Furthermore, the study conducts an in-depth evaluation of the differences in energy consumption and deployment convenience between edge computing devices and traditional high-performance computing platforms. The test platforms include a conventional high-performance computing device (RTX 4060 Laptop) as well as edge computing devices (Raspberry Pi 5 and NVIDIA Jetson Orin Nano Super). By analyzing the YOLO lightweight versions optimized for edge computing, this research aims to provide valuable insights for the future development of intelligent transportation and smart parking guidance systems.

第一章	緒論	1
1.1	研究背景與動機	1
1.2	研究目的	2
1.3	論文架構	4
第二章	背景技術與相關研究	6
2.1	邊緣運算技術	6
2.1.1	邊緣運算的定義與核心概念	6
2.1.2	邊緣運算與傳統雲端運算	7
2.1.3	邊緣運算技術架構	8
2.1.4	邊緣運算之效能優化	9
2.1.5	邊緣運算安全性與隱私保護	9
2.1.6	邊緣運算未來研究	10
2.2	YOLO 輕量級模型分析與比較	10
2.2.1	模型選擇策略	10
2.2.2	模型精確度分析	11
2.2.3	推論速度效能比較	12
2.2.4	模型大小及計算複雜度分析	13
2.2.5	模型架構特色比較	13
2.3	邊緣運算裝置效能優化策略	14
2.3.1	TensorRT硬體加速	14
2.3.2	剪枝	15
2.3.3	其他優化策略	15
2.4	邊緣運算裝置選用	16
2.5	相關研究與回顧	19
第三章	系統架構	21
3.1	相關硬體介紹	21
3.2	實驗流程	25
3.3	模型效能優化與比較流程	27
第四章	系統實作與優化結果展示	31
4.1	實驗設備與場域介紹	31
4.2	各版本 YOLO 模型初始性能分析	33
4.3	TensorRT 硬體加速評估	37
4.3.1	原始模型執行測試	37
4.3.2	TensorRT硬體加速測試	42
4.4	綜合實驗結果與效能比較	44
4.5	問題與討論	49
第五章	結論與後續研究	53
5.1	結論	53
5.2	後續研究	54
參考文獻	56

圖目錄
圖 2.1 1邊緣運算架構	8
圖 3.1 1 Raspberry Pi 4 & Pi 5	21
圖 3.1 2 NVIDIA Jetson Orin Nano Super	22
圖 3.1 3 功耗檢測儀	23
圖 3.2 1實驗流程	25
圖 3.3 1模型效能優化流程	27
圖 4.1 1電流檢測儀連接方式	32
圖 4.1 2模擬小型停車場實驗場域	32
圖 4.2 1 Ultralytics版本	33
圖 4.2 2 NVIDIA Jetson裝置版本資訊	34
圖 4.3 1 YOLOv8n實際執行兩輛模型車測試系統畫面	39
圖 4.3 2 YOLO11m實際執行兩輛模型車測試系統畫面	39
圖 4.3 3 YOLO11n實際執行兩輛模型車測試系統畫面	40
圖 4.3 4 Raspberry Pi 5實際執行運算之CPU使用率	41
圖 4.3 5 PyTorch 模型至 TensorRT 引擎檔案轉換流程 	43
圖 4.3 6 Ultralytics Docs Library轉換方式	43
圖 4.4 1 NVIDIA Jetson & RTX 4060功耗	45
圖 4.4 2 NVIDIA Jetson Orin Nano Super最終硬體設定	46
圖 4.5 1 多個鏡頭實際連接方式	50
圖 4.5 2 鏡頭啟動Error Message	51

表目錄
表 2.2 1輕量化版本各項數據對比	11
表 3.1 1所使用設備規格	23
表 4.2 1各模型於PyTorch中測試結果	34
表 4.2 2 各模型於TorchScript中測試結果	35
表 4.2 3 各模型於ONNX中測試結果	35
表 4.2 4 各模型於TensorRT中測試結果	35
表 4.2 5 各模型於NCNN中測試結果	35
表 4.3 1 NVIDIA Jetson實際場域運作偵測一輛模型車	37
表 4.3 2 NVIDIA Jetson實際場域運作偵測兩輛模型車	38
表 4.3 3 Raspberry Pi 5 實際場域運作偵測	41
表 4.3 4 YOLO TensorRT模型實際場域運作偵測兩輛車	43
表 4.4 1 各模型能效比	47

[1]
Li, S., Xu, L.D. & Zhao, S. The internet of things: a survey. Inf Syst Front 17, 243–259 (2015). https://doi.org/10.1007/s10796-014-9492-7
[2]
F. Liu, G. Tang, Y. Li, Z. Cai, X. Zhang and T. Zhou, "A Survey on Edge Computing Systems and Tools," in Proceedings of the IEEE, vol. 107, no. 8, pp. 1537-1562, Aug. 2019, doi: 10.1109/JPROC.2019.2920341.
[3] 
林晉嶽, "基於多攝影機交遞之智慧停車車輛追蹤系統," 淡江大學資訊工程學系碩士論文初稿 , 2025.
[4]
B. Varghese, N. Wang, S. Barbhuiya, P. Kilpatrick and D. S. Nikolopoulos, "Challenges and Opportunities in Edge Computing," 2016 IEEE International Conference on Smart Cloud (SmartCloud), New York, NY, USA, 2016, pp. 20-26, doi: 10.1109/SmartCloud.2016.18.
[5]
W. Shi, J. Cao, Q. Zhang, Y. Li and L. Xu, "Edge Computing: Vision and Challenges," in IEEE Internet of Things Journal, vol. 3, no. 5, pp. 637-646, Oct. 2016, doi: 10.1109/JIOT.2016.2579198 .
[6] 
Ultralytics YOLO Official Documentation, Available at:
https://docs.ultralytics.com/

[7] 
NVIDIA TensorRT Documentation, Available at:
https://docs.NVIDIA.com/deeplearning/tensorrt/archives/tensorrt-1070/index.html

[8]
Bailey J. Eccles, Philip Rodgers, Peter Kilpatrick, Ivor Spence, and Blesson Varghese. 2024. DNNShifter: An efficient DNN pruning system for edge computing. Future Gener. Comput. Syst. 152, C (Mar 2024), 43–54. https://doi.org/10.1016/j.future.2023.09.025

[9]
什麼是AIoT智慧聯網？AIoT的發展與趨勢, Available at:
https://college.itri.org.tw/Info/InfoData/2c3c2efb-04bd-4ad5-8607-1494de2f4022

[10]
Jingke Tu, Lei Yang, and Jiannong Cao. 2025. Distributed Machine Learning in Edge Computing: Challenges, Solutions and Future Directions. ACM Comput. Surv. 57, 5, Article 132 (May 2025), 37 pages. https://doi.org/10.1145/3708495
[11]
Grzesik, P.; Mrozek, D. Combining Machine Learning and Edge Computing: Opportunities, Challenges, Platforms, Frameworks, and Use Cases. Electronics 2024, 13, 640. https://doi.org/10.3390/electronics13030640
[12]
R. A. Amin, M. Hasan, V. Wiese and R. Obermaisser, "FPGA-Based Real-Time Object Detection and Classification System Using YOLO for Edge Computing," in IEEE Access, vol. 12, pp. 73268-73278, 2024, doi: 10.1109/ACCESS.2024.3404623.
[13]
H. Feng, G. Mu, S. Zhong, P. Zhang and T. Yuan, "Benchmark Analysis of YOLO Performance on Edge Intelligence Devices," 2021 Cross Strait Radio Science and Wireless Technology Conference (CSRSWTC), Shenzhen, China, 2021, pp. 319-321, doi: 10.1109/CSRSWTC52801.2021.9631594.
[14]
Shin, D.-J.; Kim, J.-J. A Deep Learning Framework Performance Evaluation to Use YOLO in Nvidia Jetson Platform. Appl. Sci. 2022, 12, 3734. https://doi.org/10.3390/app12083734
[15] 
Hakim, Ade & Juanara, Elmo & Rispandi, Rispandi. (2023). Mask Detection System with Computer Vision-Based on CNN and YOLO Method Using Nvidia Jetson Nano. Journal of Information System Exploration and Research. 1. 10.52465/joiser.v1i2.175. 
[16]
Y. Liu, Z. Qin and G. Y. Li, "Energy-Efficient Distributed Spiking Neural Network for Wireless Edge Intelligence," in IEEE Transactions on Wireless Communications, vol. 23, no. 9, pp. 10683-10697, Sept. 2024, doi: 10.1109/TWC.2024.3374549.
[17]
Chen, R., Wang, P., Lin, B. et al. An optimized lightweight real-time detection network model for IoT embedded devices. Sci Rep 15, 3839 (2025). https://doi.org/10.1038/s41598-025-88439-w
[18]
R. C. Camara de M. Santos, M. Coelho and R. Oliveira, "Real-time Object Detection Performance Analysis Using YOLOv7 on Edge Devices," in IEEE Latin America Transactions, vol. 22, no. 10, pp. 799-805, Oct. 2024, doi: 10.1109/TLA.2024.10705971.
[19] 
RS-Online shop, Available at:
https://twcn.rs-online.com/web/

[20] 
K. Cao, Y. Liu, G. Meng and Q. Sun, "An Overview on Edge Computing Research," in IEEE Access, vol. 8, pp. 85714-85728, 2020, doi: 10.1109/ACCESS.2020.2991734 . 
[21]
A. Hazarika, N. Choudhury, M. M. Nasralla, S. B. A. Khattak and I. U. Rehman, "Edge ML Technique for Smart Traffic Management in Intelligent Transportation Systems," in IEEE Access, vol. 12, pp. 25443-25458, 2024, doi: 10.1109/ACCESS.2024.3365930.
[22]
A. Bandyopadhyay et al., "EdgeMatch: A Smart Approach for Scheduling IoT-Edge Tasks With Multiple Criteria Using Game Theory," in IEEE Access, vol. 12, pp. 7609-7623, 2024, doi: 10.1109/ACCESS.2024.3350556.
[23]
TK282PLUS Digital Power Meter Wattmeter Energy Consumption Tester Voltage Current Electricity, Available at:
https://www.tekcoplus.com/products/tk282plus?srsltid=AfmBOooeywo1WJFXX7iRRPhZrDoHUdfyV6cSDYM1ZPwR9ZTIkndYiOeN

[24]
MJPG-Streamer, Available at:
https://github.com/jacksonliam/mjpg-streamer

[25] 
NVIDIA JetPack 6.2 Super Mode, Available at:
https://developer.NVIDIA.com/blog/NVIDIA-jetpack-6-2-brings-super-mode-to-NVIDIA-jetson-orin-nano-and-jetson-orin-nx-modules/

[26]
ONNX Github, Available at:
https://github.com/onnx/onnx

[27]
台灣電力公司官網, Available at:
https://www.taipower.com.tw/?checkLang=N