以物件群聚為基底之半全域視差匹配深度圖生成演算法

Object Clustering based Semi-Global Depth Estimation Algorithm

傅煜翔(Yu-Siang Fu)

指導教授 ： 李佩君

國立暨南國際大學/科技學院/電機工程學系/碩士(2016年)

摘要

隨著3D視訊技術日漸普及，為了因應大眾對於立體視訊需求，可藉由市售雙鏡頭影像設備取得雙視角影像產生立體視訊。立體視訊影像的產生，可利用雙視角深度預估演算法產生深度圖，進行深度影像繪圖(depth image-based rendering, DIBR)技術產生多個新視角組成的立體影像。
現行雙視角深度圖估測方法可分為全域視差匹配(Global Matching)及區域視差匹配(Local Matching)。全域視差匹配利用兩張完整畫面資訊估算出之間的視差，使其預估出的深度圖較為準確，然而其缺點具有較高的計算複雜度；而區域視差匹配利用匹配點附近的區域資訊進行視差計算，藉以降低計算時間，但因可匹配之資訊量較少，預估出來的深度圖準確性不如全域視差匹配方式精準。因此半全域視差匹配(SGM)是結合區域視差匹配的代價計算方法及全域視差匹配的代價聚合(Cost Aggregation)發展出的演算法。
本論文應用區域視差匹配的Census Transform & Hamming Distance進行匹配代價值(Matching Cost Value)計算，接著利用全域視差匹配進行代價聚合(Cost Aggregation)，計算出的匹配代價值(Matching Cost Value)進行視差估計(Disparity Calculation)，並將估算出來的視差資訊轉換為深度影像。為了改善區域匹配代價資訊搜索不足與視差造成的遮蔽(occlusion)現象的錯誤匹配視差問題，本論文提出利用物件群聚(Object Clustering)方法改善錯誤的匹配視差資訊。

關鍵字

立體匹配 ； 物件群聚 ； 深度圖產生 ； 半全域視差匹配 ； 普查式轉換 ； 深度圖影像繪圖法

並列摘要

As the 3D video technology popular and the demands of the three-dimensional video, the three-dimensional video can be generated by using the images which captured by the multi camera devices. And then these two images are used to estimate the depth map as called binocular depth generation. The depth image-based rendering (DIBR) is used to generate the virtual views for 3D display by using the color image and its corresponding depth map.
The existing binocular depth map estimation method includes global disparity matching (GDM) and local disparity matching (LDM) methods. The GDM uses two entire images information to estimate the disparity between two view images. The advantage of the GDM methods is high accurately depth map estimation. However, the computational time is very high. Therefore, the LDM uses the local information near by the matching point to calculate the disparity for reducing computational time. But the accuracy of the depth map by using LDM generation is not good as the depth map by using GDM generation. Therefore, Semi-Global Disparity Matching (SGM) is an algorithm that is combined by the cost calculation method of LDM and cost aggregation of GDM.
This thesis adopts Census Transform & Hamming Distance of local disparity to calculate matching cost value. Then, the proposed algorithm adopts matching cost value to calculate cost aggregation for disparity estimation. In order to improve the mismatching in disqualification of local matching cost information and occlusion in stereo matching processing, this thesis proposes object clustering method to improve the accurately cost aggregation.

並列關鍵字

Stereo Matching ； Object Clustering ； Depth Map Generation ； Semi-Global Matching ； Census Transform ； DIBR

參考文獻

[1] R. Lagendijk, R. Franich and E. Hendriks, “Stereoscopic Image Processing,”http://mmc.tudelft.nl/sites/default/files/Stereoscopic%20Image%20Processing.pdf

[2]賭城天天報新聞:(1.1) http://www.lvcdn.com/news/fashion/20160111/6194.html Jan 2016.

[3]C. Fehn, “A 3D-TV system based on video plus depth information” Conference Record of the Thirty-Seventh Asilomar Conference on Signals System and Computers, vol. 2,pp. 1529-1533, Jun. 2003.

[4] C. Fehn, “Depth-Image-Based-Rendering (DIBR), Compression and Transmission for a New Approach on 3D-TV,” in Proceeding of International Society for Optical Engineering Conf. Stereoscopic Displays and Virtual Reality Systems XI, vol. 5291, San Jose, CA, USA, pp. 93-104, Jan. 2004.

[5]D. Scharstein and R. Szeliski, “A taxonomy and evaluation of dense two-frame stereo correspondence algorithms,” International Journal of Computer Vision, vol. 47, pp.63-77 ,April-June 2002.