隨著無人機（Unmanned Aerial Vehicle， UAV）應用日益普及，即時監測與追蹤無人機逐漸成為重要議題。然而，遠距影像中的無人機通常尺寸小且特徵不明顯，若直接以完整影像進行深度學習推論，將增加邊緣平台之運算負擔。為此，本文提出一套輕量化 UAV 視覺辨識與追蹤系統，透過候選區域生成、輕量化多尺度 CNN 辨識模型、辨識追蹤兩模式切換與 PTZ 相機控制，使系統能在降低運算量的同時完成 UAV 目標搜尋、位置更新與連續追蹤。 模型設計上，本文導入離散式統一縮放規模設計，建立不同規模之模型版本，以分析辨識性能與推論成本之間的取捨，並選擇較適合邊緣部署之模型作為實際系統核心。實驗結果顯示，所提出方法可於 Arduino UNO Q、Jetson Orin NX 與 ROG-R9/3060 等不同運算能力平台上完成完整辨識與追蹤流程，並透過真實 PTZ 相機操作實測，驗證本文所提出之輕量化 UAV 視覺辨識與追蹤方法具備實際邊緣部署之可行性。;With the increasing use of Unmanned Aerial Vehicles (UAVs)， real-time UAV monitoring and tracking have become an important issue. However， UAVs in long-range images are usually small and less distinctive， while full-frame deep learning inference increases the computational load on edge platforms. To address this problem， this thesis proposes a lightweight UAV visual recognition and tracking system that integrates candidate-region selection， a lightweight multi-scale CNN model， dual recognition/tracking modes， and PTZ camera control. The proposed system reduces computational cost while performing UAV searching， position updating， and continuous tracking. For model design， a discrete unified scaling strategy is introduced to construct models of different scales and evaluate the trade-off between recognition performance and inference cost. The selected model is then deployed as the core of the practical system. Experimental results show that the proposed method can complete the full recognition and tracking process on Arduino UNO Q， Jetson Orin NX， and ROG-R9/3060 platforms with real-time or near-real-time performance. Real PTZ camera tests further verify the feasibility of the proposed lightweight UAV recognition and tracking method for practical edge deployment.
