最終未見測試資料集包含 672 個片段，涵蓋一般室內辦公光源與低光源兩種環境。最佳模型達到 99.26% 的辨識準確率與 99.27% 的巨集 F1 分數。在較高效能硬體上，圖形處理器加速使所有評估之時間模型變體皆能滿足 495 毫秒的即時門檻；而在較低功耗硬體上，最小模型變體則具有最實際的延遲表現。實驗結果顯示，當模型複雜度、推論後端與硬體能力一併納入考量時，事件相機手勢辨識系統能同時達成高辨識率與實際即時部署能力。;Hand gesture recognition is an important component of natural Human-Computer Interaction because it enables users to communicate commands without physical input devices. However， many RGB-based systems are affected by motion blur， illumination variation， background redundancy， and high computational cost， which can limit real-time deployment on resource-constrained hardware.
This thesis proposes an end-to-end real-time hand gesture recognition pipeline using an event camera and lightweight temporal modeling. Asynchronous event streams are converted into polarity-based frame representations， temporally sampled with a stride of 3， and grouped into 30-frame clips representing approximately 2.97 seconds of motion context. The clips are classified into four classes: rock， paper， scissors， and others. Three lightweight temporal model variants were evaluated， and supplementary spatial-temporal baselines were included for comparison. The trained models were deployed using an optimized runtime backend with CPU and GPU execution on different commodity hardware platforms.
The final unseen test set contained 672 clips recorded under office-lighting and low-light conditions. The best model achieved 99.26% accuracy and 99.27% macro F1-score. GPU acceleration enabled all evaluated temporal variants to satisfy the 495 ms real-time threshold on stronger hardware， while the smallest variant provided the most practical latency profile on lower-power hardware. These results show that event-camera-based gesture recognition can achieve high accuracy and practical real-time deployability when model complexity， runtime backend， and hardware capability are considered together.
