摘 要
隨著無人駕駛技術的普及，其應用範圍不僅限於道路交通，還可擴展至工廠無人搬運車、果園採果搬運車、校園文件配送車等領域。為了促進相關應用的開發，本論文提出一套基於模仿學習與物件辨識技術的智慧載具系統，設計並實現了一個小型智慧載具的自動駕駛系統，目的在透過影像辨識技術提升載具在虛擬道路環境中的自駕能力。系統硬體採用Raspberry Pi 4 和 NVIDIA Jetson Nano 平台，並結合 Donkey Car 系統和Tiny-Yolo V7物件辨識技術，能夠有效識別道路分隔線、交通燈號、交通標誌及行人車輛等物體，並根據辨識結果做出相應的駕駛決策。為了提升駕駛安全性，系統還在車前劃定了危險區域，當物體進入危險區時，系統會立即啟動緊急煞車動作。除此之外，系統可以根據交通燈號的顏色變換、交通標誌的指示等信息調整車輛行駛狀態。測試結果顯示，該系統能夠準確識別交通燈號、交通標誌，並根據辨識到的物體進行適當的停車或啟動操作，達到基本的自駕行駛需求。本研究成功展示了基於低成本嵌入式硬體平台的自動駕駛解決方案，並為未來進一步拓展至真實交通環境提供了有價值的參考。未來可針對物件識別精度、系統運算效能及多場景應用進行進一步優化。
關鍵詞：模仿學習、自動駕駛、無人駕駛、物件辨識、嵌入式系統 Abstract
With the widespread adoption of autonomous driving technology, its application scope extends beyond road traffic to areas such as unmanned transport vehicles in factories, fruit-picking transport vehicles in orchards, and campus document delivery vehicles. To promote the development of related applications, this study proposes an intelligent vehicle system based on imitation learning and object recognition technology. A small intelligent vehicle’s autonomous driving system was designed and implemented to improve the vehicle’s self-driving capability in a virtual road environment using image recognition technology. The system hardware is built on the Raspberry Pi 4 and NVIDIA Jetson Nano platforms, integrating the Donkey Car system with Tiny-Yolo V7 object recognition technology. It can effectively identify objects such as road dividers, traffic lights, traffic signs, pedestrians, and vehicles, making corresponding driving decisions based on the recognition results. To enhance driving safety, the system defines a danger zone in front of the vehicle. When an object enters the danger zone, the system immediately activates an emergency brake. Furthermore, the system adjusts the vehicle's driving state based on traffic light color changes and traffic sign instructions. Test results show that the system can accurately identify traffic lights and traffic signs, perform appropriate stop or start operations based on the recognized objects, and meet the basic self-driving requirements. This research successfully demonstrates an autonomous driving solution based on a low-cost embedded hardware platform, providing a valuable reference for future expansion into real-world traffic environments. Future work can further optimize object recognition accuracy, system computational performance, and multi-scenario applications.
Keywords: Imitation learning, Autonomous driving, Autonomous driving, Object recognition, Embedded systems