本篇論文提出一套結合改良式孿生網路與深度強化學習之 PTZ相機無人機視覺追蹤系統，旨在應對無人機於高速飛行及不穩定軌跡下所產生的追蹤困難。於追蹤模組部分，本研究以 SiamFC++ 為基礎架構，導入結合 AlexNet 與 Coordinate Attention 、Global Context Attention 機制之特徵提取模組，以增強模型對空間與通道資訊的感知能力，進而提升追蹤準確性與穩定性。 在控制策略方面，採用 Deep Q Network 演算法設計 PTZ 相機之動態角度調整機制，使系統能根據目標於畫面中的偏移自動修正雲台角度，以保持目標穩定位於畫面中心。為驗證所提方法之效能，本研究於理想路徑、軌跡擾動與速度變化等三種模擬場景中進行實驗比較。實驗結果度，能有效應用於無人機即時視覺追蹤任務。顯示，所提出之追蹤系統在方位角、俯仰角與視覺中心偏移距離等指標上，皆優於傳統控制方法，證實本系統具備良好的追蹤穩定性與控制精度。;This study proposes a PTZ camera-based UAV visual tracking system that integrates an improved Siamese network with deep reinforcement learning, aiming to address the challenges of tracking UAVs during high-speed flight and unstable trajectories. The system builds on the SiamFC++ architecture and incorporates a feature extractor combining AlexNet with the Coordinate Attention mechanism. This enhances the model＇s perception of spatial and channel information, improving tracking accuracy and stability.
For control, the system uses the Deep Q Network algorithm to dynamically adjust PTZ camera angles, allowing automatic correction based on the target＇s displacement in the frame and keeping it centered. To evaluate the system＇s effectiveness, experiments were conducted under three simulated conditions: ideal trajectory, trajectory perturbation, and variable speed. Results show the proposed system outperforms traditional control methods in azimuth, elevation, and visual offset, demonstrating strong tracking stability and control for real-time UAV tasks.
