本篇論文運用 OpenCL - 平行化程式的語言來達到通用圖形處理器的目的,使運用引導影像濾波器的立體匹配演算法透過通用圖形處理器盡可能的平行加速。目前多數的立體匹配演算法依然著重在準確度,但立體匹配的應用上,對於速度也有所需求,如自動駕駛、模型掃描等。我們的演算法是局部式演算法,作為相當經典的基礎演算法,藉由這基礎的演算法來展示通用圖形處理器搭配平行化計算的加速潛力,其步驟如同經典論文 A Taxonomy and Evaluation of Dense Two-Frame Stereo Correspondence Algorithms 此論文所介紹的步驟,分別是匹配代價 –使用普查轉換以及索貝爾算子處理後,利用絕對誤差和計算出代價,代價聚合 -則使用引導影像濾波器來處理各個像差圖片(Disparity Image)(像差圖片空間Disparity Image Space),最後最佳化的方式則是贏家全拿,如同一般的局部式方 法,以上步驟經過 OpenCL 的平行化後根據硬體的不同,我們整體加速至少一倍以上。這篇主要貢獻是以 OpenCL 來實作平行化運用 guided filter 的立體匹配演算法,並提供目前在新版本的平台,中央處理器、通用圖形處理器 – OpenCL 1.2 、通用圖形處理器– OpenCL 2.0 的速度比較,OpenCL 2.0 版本的特性使用解釋,以及當下最新硬體 AMD R9-390X 的紀錄。
In this thesis we use OpenCL which is a programming language for GPGPU to parallelize the Stereo Matching Algorithm Using Guided Filter(GF) on cost aggregation. Today most algorithms of Stereo Matching focus on accuracy but not the processing time, there are certain applications that require real-time processing. For example - auto-driving. Our algorithm is a basic method of stereo matching which can be categorized as a local method. Its steps are similar to the thesis “A Taxonomy and Evaluation of Dense Two-Frame Stereo Correspondence Algorithms”. The first step is Matching-Cost. We will use Census Transform (CT), Sobel Operator (SO) and Sum of Absolute Difference (SAD). And we use Guided Filter in the second Steps: Cost Aggregation. Finally, like other local method, we apply winner-take-all. All the steps we previously describe speed up at least two times after parallelize by OpenCL. Our contribution is on the parallelization of Stereo Matching with Guided Filter and record of processing time with CPU, GPU -OpenCL 1.2 and OpenCL 2.0 on recent hardware – AMD R9 390X