使用影像導引圖片分割之立體匹配演算法

Stereo Matching Algorithm Using Image-guided Graph-cut

葉騏銘(Chi-Ming Yeh)

指導教授 ： 盧奕璋

國立臺灣大學/電機資訊學院/電子工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202102122

摘要

立體匹配演算法的原理是利用不同視角的鏡頭拍攝照片，再藉由場景中各個物體之間不同的偏移量計算出場景的深度資訊。這樣的演算法在目前已經有多種應用，例如機器人的立體視覺以及自動駕駛汽車。
本篇論文主要目的是利用演算法求得場景的視差值，演算法總共可分為以下幾個步驟。首先，設計一組成本公式計算出場景中每一個點的成本，並利用影像導引濾波器進行成本聚合。利用視差值具有空間連續性的特性，設計出一組成本函數同時考慮資料項與平滑項之總和最小。最後，將此問題對應到一個圖，藉此將最小化成本函數問題對應成圖的最小割問題。為了不受限於正面平行假設，再將視差值變數對應到一組三維的標籤，並利用所設計的離散型三維圖割演算法求得最小化成本函數的解。本文演算法採用Middlebury網站以及KITTI網站的測試資料進行驗證，使用的平台為Matlab R2020b，運作之處理器為Intel Core i7-9700K CPU @ 3.6GHz，記憶體容量為32GB。

關鍵字

立體匹配 ； 圖割 ； 三維標籤法

並列摘要

Stereo matching problem is a popular issue nowadays since there are plenty of applications such as stereo vision for robots and auto-pilot vehicles. In this thesis, we construct an algorithm flow to solve the stereo matching problem.
First of all, we design a pixel-wise function to calculate the cost volume according to the scene pairs. Then we use guided filter to aggregate the cost volume as the data term. Because the disparity is spatially continuous in the scene, we need to consider the data term and smooth term together, and find the minimum of the cost function. Since the cost function can be mapped to a graph, the minimum-solving problem is transformed into a min-cut problem. In the graph, all of the vertices represent the pixels in the scene image, and the weights of the edges represent the costs. In addition, in order to avoid problems caused by frontal-parallel assumption, we also use 3D labeling, which has 3 parameters to represent a single disparity variable. In the last, we designed a discrete 3D graph cut, which is improved from traditional graph cut algorithm to solve the 3D graph cut problem.
We verified the correctness of the designed algorithm flow with the dataset provided by Middlebury website and KITTI website. The algorithm is coded in Matlab R2020b with Intel Core i7-9700K CPU @ 3.6GHz and 32GB RAM.

並列關鍵字

stereo matching ； Graph cut ； 3D labeling

參考文獻

[1] Y. Zhan, Y. Gu, K. Huang, C. Zhang and K. Hu, “Accurate Image-Guided Stereo Matching With Efficient Matching Cost and Disparity Refinement,” IEEE Transactions on Circuits and Systems for Video Technology, September 2016.

[2] X. Mei, X. Sun, W. Dong, H. Wang, X. Zhang, “Segment-Tree based Cost Aggregation for Stereo Matching,” IEEE Computer Society Conference on Computer Vision and Pattern, June 2013.

[3] Y. Fu, K. Lai, W. Chen, and Y. Xiang, “A pixel pair based encoding pattern for stereo matching via an adaptively weighted cost,” IET Image Processing, October 2019.

[4] L. Wang and R. Yang, “Global stereo matching leveraged by sparse ground control points,” Proc. of IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), p. 3033–3040, 2011.

[5] S. N. Sinha, D. Scharstein and R. Szeliski, “ Efficient High-Resolution Stereo Matching Using Local Plane Sweeps,” 2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus, OH, pp. 1582-1589, 2014.