立體視覺技術（Stereo Vision）被廣泛應用於各種領域如3D電影等，而深度圖（Depth Map）則是產生立體視覺的重要資訊。深度圖可經由立體匹配（Stereo Matching）而得，但在即時性方面，雖然區域性演算法複雜度較低，可達即時性的要求，但產生的深度品質較不佳。而能產生深度品質較佳的全域性演算法，卻因複雜度較高，以軟體實作方面較不容易達到即時性的要求。本論文提出以全域性演算法中的可信度傳遞法（Belief Propagation，BP）為基礎之硬體設計，來加速立體匹配產生深度資訊的運算。可信度傳遞法主要分成三部分：匹配代價計算（Data Cost Calculation）、訊息更新（Message Update）與可信度計算（Compute Belief），其中以訊息更新為可信度傳遞法之核心，也是時間成本與面積成本的主要差異處。所以模組化訊息更新之硬體（PE），透過置放不同數量的訊息更新硬體來分析時間成本與硬體面積成本之間的關係，找出可達即時性下之最小硬體面積成本。

Stereo vision is widely used in 3D graphics applications. Depth map, an important information in stereo vision, can be generated using stereo matching. In general, there are two categories of stereo matching methods: local and global. Local stereo matching with low computation complexity can achieve real-time processing, but the quality of generated depth map is not satisfactory. On the other hand, global stereo matching can generate high-quality depth map but requires much higher computation complexity and thus cannot achieve real time processing using pure software implementation. In this thesis, we present hardware implementation of a global stereo matching method based on belief propagation (BP) in order to generate depth map from two images of different view angles. The BP-based stereo matching has three parts: matching cost computation, message update, and belief computation where message update is the critical part. This thesis presents several versions of designs for message update using various number of processing elements (PEs) by considering the trade-off of area cost and computation speed.

第1章 概論（Introduction） 1
1.1 研究動機 1
1.2 本文大鋼 1
第2章 研究背景與相關研究 3
2.1 立體匹配 3
2.1.1 立體視覺幾何（Stereo Geometry） 3
2.1.2 極線幾何與極線幾何限制（Epipolar Geometry） 5
2.2 立體匹配演算法 6
2.2.1 匹配代價計算（Matching Cost Computation） 6
2.2.2 代價函數聚合（Cost Aggregation） 7
2.2.3 視差計算（Disparity Computation） 8
2.2.4 視差優化（Disparity Refinement） 13
第3章 演算法流程 15
3.1 色彩影像轉亮度影像（RGB to Intensity） 15
3.2 匹配代價計算（Data Cost Calculation） 15
3.3 訊息更新（Message Update） 16
3.3.1 線性平滑函式（Linear Model） 18
3.3.2 截斷式線性平滑函式（Truncated Linear Model） 21
3.4 可信度計算（Compute Belief） 23
第4章 硬體架構設計與實現 25
4.1 色彩影像轉亮度影像（RGB to Intensity） 26
4.2 匹配代價計算（Data Cost Calculation） 26
4.3 訊息更新 （Message Update） 28
4.3.1 Linear Model硬體設計 29
4.3.2 Parallel Linear Model硬體設計 30
4.3.3 Sequential Linear Model硬體設計 33
4.4 可信度計算 （Compute Belief） 36
4.4.1 Compute Belief硬體設計 36
4.4.2 Sequential Compute Belief硬體設計 37
4.5 可信度傳遞法之硬體系統設計 38
4.5.1 全畫面處理單元(Full-Screen PE)之訊息更新架構設計 38
4.5.2 單一處理單元(Single PE)之訊息更新架構設計 41
4.5.3 多重處理單元(Multiple PE)之訊息更新架構設 44
4.5.4 連續影像之管線化系統架構設計 50
第5章 實驗結果分析與比較 53
5.1 週期數分析 53
5.2 硬體面積分析 57
5.3 邏輯合成數據 62
5.4 效能評比 65
5.5 測試影像 69
第6章 結論與未來展望 71
6.1 結論 71
6.2 未來展望 71
參考文獻（References） 73

[1] C. Fehn, “A 3D-TV system based on video plus depth information,” in Proc. of Asilomar Conference on Signals, Systems and Computers, vol. 2, pp. 1529-1533, 2003
[2] D. Scharstein, “View Synthesis Using Stereo Vision”, Dissertation of Cornell University PHD, 1997
[3] D. Scharstein, and R. Szeliski, “A taxonomy and evaluation of dense two-frame stereo correspondence algorithms,” Int. J. Comput. Vision, vol. 47, pp. 7–42, 2002.
[4] M. Gong, et al., “A Performance Study on Different Cost Aggregation Approaches Used in Real-Time Stereo Matching,” Int. J. Comput. Vision, vol. 75, pp. 283–296, 2007.
[5] T. Tao, “A fast block matching algorthim for stereo correspondence,” in IEEE Conference on Cybernetics and Intelligent System, vol. 21-24, pp.38-41, 2008
[6] Ingemar J. Cox, et al., “A Maximum Likelihood Stereo Algorithm_Computer Vision and Image Understanding,” Computer Vision and Image Understanding, vol. 33, pp.542-567, 1996
[7] J. Sun, N.-N. Zheng, and H.-Y. Shum, “Stereo matching using belief propagation,” IEEE Transaction on Pattern Analysis and Machine Intelligence, vol. 25 pp.787-800, 2003
[8] A. Delong, and Y. Boykov, “A scalable graph-cut algorithm for N-D grids,” in IEEE Conference on Computer Vision and Pattern Recognition, pp.1-8, 2008
[9] W. Wang, and C. Zhang, “Local Disparity Refinement with Disparity Inheritance,” Photonics and Optoelectronics, vol. 21-23, pp.1-4, 2012
[10] Q. Yang, et al., “Stereo Matching with Color-Weighted Correlation, Hierarchical Belief Propagation, and Occlusion Handling,” IEEE Transaction on Pattern Analysis and Machine Intelligence, vol.31, pp.492-504
[11] P. F. Felzenszwalb, and D. P. Huttenlocher, “Efficient belief propagation for early vision,” Int. J. Comput. Vision, vol.70, pp.41-54, 2006
[12] W.-C. Kao, et al., “Real-Time Depth Map Estimation for Stereoscopic Displays with GPU,” IEEE 15th International Symposium on Consumer Electronics, pp.228-231 2011
[13] Y.-K. Choi, W. Williem, and I. K. Park, “Memory-efficient belief propagation in stereo matching on GPU,” Signal & Information Processing Association Annual Summit and Conference, vol. 3-6 pp.1-4 2012
[14] Y.-C. Tseng, and T.-S. Chang, “Architecture Design of Belief Propagation for Real-Time Disparity Estimation,” IEEE Transaction on Circuits and System for Video Technology, vol.20 pp.1555-1564 2010
[15] W.-L. Wang, “Hardware Design for Disparity Estimation Using Dynamic Programming,” Master's thesis, University of NSYSU, 2012