以立體視覺為基礎的工業檢測或是機器視覺技術需要一連串高複雜度的影像處理流程，因此以軟體實現立體視覺通常需要高效能的處理器以應付複雜的演算法計算。而相對於軟體，硬體資源受限的嵌入式系統則受到成本與技術的限制，使得立體視覺於嵌入式系統開發難以實現。本研究藉由MIAT嵌入式硬體設計方法論提出一個立體視覺嵌入式硬體架構，我們先計算比對成本，然後利用樹狀結構進行視差計算，接著得到深度影像後再進行本研究所提出的視差平滑化演算法，可大幅減少物件深度因為光線與陰影造成得錯誤率，實驗結果顯示能夠有效提升偵測物件深度的準確率，減低光線與陰影的影響，在動態影像中使系統偵測的物件深度能夠穩定，不受短暫光源變化影響而變動。最後則整合成立體視覺晶片，應用在各種嵌入式系統。;Industrial inspection system or machine vision system are based on stereo vision technology, those systems requires a lot of highly complex image processing to implement their algorithms. Therefore, implementing stereo vision system with software usually needs high-efficiency processor to cope with calculating complex algorithms. And with respect to the software, embedded systems are subject to restrictions of cost and technology due to restriction on hardware resources. For this reasons, making the development of stereo vision in embedded systems is difficult to achieve. In this paper, we proposed a stereo vision embedded hardware architecture using MIAT embedded hardware design methodology. We first calculate the matching cost, and then use the minimum spanning tree algorithm to calculate disparity. After getting the depth of the image, we do disparity smoothing to significantly reduce the error rate of the object depth, which is caused by the light and shadow. The experimental results show that our system can effectively improve the accuracy of detection of the object depth, and reduce the impact of light and shadow. Our system detects moving objects’ depth in image can be stable, and won’t be affects by short-term changes of light. At last, we integrate all algorithms into a stereo vision chip, and applied to various embedded systems.