李辨識：寵物行為實時辨識於嵌入式系統

Abstract

隨著家庭中寵物數量的增加，它們的健康和福利已成為現代社會的重要關注點。實時行為識別技術的發展，特別是其在嵌入式系統中的應用，為監控和理解寵物行為提供了新的途徑。本文探討了如何通過生成技術和模型優化來實現在嵌入式系統中的實時寵物行為辨識，並研究了這項技術在寵物照護中的有效應用。

本文首先通過使用Stable Diffusion技術生成各種貓行為的高品質影像，以解決訓練數據集不足的問題。接著，選用YOLOv7-tiny模型來進行貓行為辨識，並使用修剪技術進一步優化模型，減少其大小和計算需求。最後，將優化後的模型部署到Realtek AMB82-mini晶片上，實現高效的實時推斷。

本研究展示了利用生成技術擴充數據並修剪網路來優化模型以適應資源受限環境的有效方法，並驗證了其在實時貓行為辨識中的實用性和高效性。

關鍵字：寵物行為識別、嵌入式系統、生成模型、Stable Diffusion、YOLOv7-tiny、模型剪枝

With the rising number of pets in households, their health and welfare have become significant concerns in modern society. The development of real-time behavior recognition technology, especially its application in embedded systems, offers new ways to monitor and understand pet behaviors. This study explores how to implement real-time pet behavior recognition on embedded systems through generative techniques, network pruning, and model optimization and investigates the effective application of this technology in pet care.

First, high-quality images of various cat behaviors were generated using Stable Diffusion to address the issue of insufficient training datasets. Next, the YOLOv7-tiny model was employed for cat behavior recognition, further optimized using pruning techniques to reduce its size and computational requirements. Finally, the optimized model was deployed on the Realtek AMB82-mini chip, achieving efficient real-time inference.

This research demonstrates the effectiveness of using generative techniques to augment datasets and optimize models for resource-constrained environments, validating its practicality and efficiency in real-time cat behavior recognition.

Keywords: pet behavior recognition, embedded systems, generative models, Stable Diffusion, YOLOv7-tiny, model pruning