穩健的影像及影片比對演算法及其應用於視野結合

Robust Image and Video Matching and Its Application to View Integration

陳心怡(Hsin-I Chen)

指導教授 ： 陳炳宇

國立臺灣大學/電機資訊學院/資訊工程學研究所/博士(2015年)

https://doi.org/10.6342/NTU.2015.00178

摘要

建立影像間的特徵點對應關係是許多影像分析課題中，一個十分基礎與核心的資料分析技術。
盡管其是許多影像相關應用的基本要素, 在複雜場景下仍普遍存在兩個問題:
低準確率 (precision) 與低檢出率 (recall)。
在另外一方面, 如何識別影片之間像素的對應關係，仍是一個具有很大挑戰但尚未被被廣泛討論的問題。
在本論文中，我們發展以投票為基礎之特徵點配對演算法來進行影像的匹配，並研究影片之間像素稠密配對的技術。
首先，我們提出了一個植基於霍夫轉換 (Hough transform) 的特徵點對應演算法來大幅提高幾何驗證的速度，並發展了反霍夫轉換 (inverted Hough transform)，透過交替疊代霍夫轉換與反霍夫轉換，所提出的方法可同時提昇特徵比對之準確率 (precision) 與檢出率 (recall)。
接著我們更進一步結合影像共分割 (image co-segmentation) 和描述子 (descriptor) 的不變性，探索如何搜尋和增加更多的可能的候選特徵點達成稠密的對應效果。
最後，我們提出一個創新的視頻映射技術來建立兩段重疊區域極小的影片間像素稠密對應關係，
我們將此技術應用在行車紀錄器拍攝到的影片中，並對前車拍攝到的影像做區域性的形變，使其視角及輪廓邊緣能夠與後車被遮擋部分對應並能無縫的接合在一起，創造出前方車輛似乎變得半透明的效果以達成改善駕駛者可見度的目的之應用。

關鍵字

電腦視覺 ； 特徵點對應 ； 影像共分割 ； 影片比對 ； 影像形變

並列摘要

Establishing feature correspondences is a fundamental problem in many
image analysis tasks, and is required for a wide range of applications. Despite
the great applicability, two main difficulties hinder the advance in establishing
the correspondences of high quality: (1) low precision and (2) low recall. In
addition, how to establish dense mapping between videos is a more challenging
but less addressed in the community. In this dissertation, we introduce
a voting-based algorithm for image matching, and describe an inter-video
mapping framework to establish dense mapping between partial overlapping
videos. First, we propose an algorithm that is based on the Hough transform
to establish feature correspondences, which leads to speed-up in geometric
checking. We also develop an inverted Hough transform, and through an iterative
optimization process, we can enhance the quality of matching in both
precision and recall. Second, we integrate image co-segmentation into feature
matching and combine different descriptors, which can yield more accurate
and dense correspondences. Finally, we present a novel inter-video mapping
approach to align videos with small overlapping regions, and apply it to video
footages from two different dashcams installed on back vehicle. We show that
with our technique, it is able to locally adjust the shape of the unobstructed
view in the preceding vehicle so that its perspective and boundary could be
matched to that of the occluded region in its following vehicle, creating an impression as if the preceding vehicle is transparent thus increases drivers' visibilities.

並列關鍵字

computer vision ； feature correspondence ； co-segmentation ； video matching ； image warping

參考文獻

[32] P. E. R. Gomes, F. Vieira, and M. Ferreira. The see-through system: From implementation

compared to state-of-the-art superpixel methods. IEEE Trans. on Pattern Anal. and Mach.

R. Szelisk. Panoramic video textures. ACM Trans. on Graphics, 24(3):821–827, 2005.

accurate correspondences selection in structure from motion: A game-theoretic perspective.

Int. J. Computer Vision, 97(1):36–53, 2012.