基于图像深度感知中的立体匹配和深度增强算法研究

发布时间：2018-01-10 09:02

本文关键词：基于图像深度感知中的立体匹配和深度增强算法研究　出处：《清华大学》2015年博士论文　论文类型：学位论文

【摘要】：基于图像深度感知,即从图像中恢复深度信息,是计算机视觉领域的核心问题之一。它在传统的三维建模,机器人导航以及新兴的移动消费领域都有着广泛的应用。基于图像深度感知的重要组成部分是立体匹配和深度增强,立体匹配研究如何得到图像间的像素级匹配关系进而获取深度信息,深度增强主要研究如何进一步改善深度信息的质量。在新的应用场景下,立体匹配和深度增强的研究工作面临着诸多困难和挑战。针对立体匹配所面临的光照变化、弱纹理等挑战,本文利用特征空间和尺度空间提出了新的立体匹配算法和算法框架来改善匹配精度;针对深度增强的随机性、未知性等困难,本文提出了新的深度增强算法,突出了置信估计的重要作用,具体而言本文的主要贡献如下:?提出了一种二值立体匹配算法。本文将二值特征描述子引入到立体匹配问题中,用彩色一致性近似视差一致性,构造了自适应的像素对采样算法来解决边界膨胀问题。自适应采样通过二值掩码实现,可以通过位运算快速计算匹配代价。在视差修正阶段,本文提出了加权投票的方法,进一步改善匹配精度。通过在公开数据集上的评测,证实了二值立体匹配算法具有较好的精度、实时性、光照鲁棒性和性能可控性。?提出了一种多尺度立体匹配算法框架。本文将立体匹配中的代价聚合建模为加权最小二乘问题,利用该模型分析了目前先进的代价聚合算法。本文将代价聚合的优化模型扩展到尺度空间得到多尺度代价聚合优化模型,并在优化目标中添加约束尺度间视差响应一致性的正则项,将多尺度信息引入到传统的代价聚合方法中。多尺度代价聚合优化模型能够集成已有的代价聚合算法,得到了立体匹配算法框架。通过在多个公开数据集上的对比测试,本文提出的算法框架能够有效地改善已有算法的匹配精度。?提出了一种基于置信估计的深度增强算法。本文将单目深度置信估计问题建模为双标记的马尔可夫随机场,能够快速估计深度置信度。在此基础上,本文提出了一种基于置信度的搜索算法来修正深度。相较于传统的联合滤波的方法,基于置信度的搜索算法能够处理局部区域中的大面积深度错误。在深度超分辨和深度去噪等应用场景下的对比实验证明了本文算法能够得到更为精确的深度信息。
[Abstract]:Depth sensing based on image, that is, recovering depth information from image, is one of the core problems in the field of computer vision. Robot navigation as well as the emerging field of mobile consumption have a wide range of applications. Stereo matching and depth enhancement are important components of depth sensing based on image. Stereo matching studies how to obtain the pixel level matching relationship between images and obtain depth information. Depth enhancement mainly studies how to further improve the quality of depth information. The research of stereo matching and depth enhancement is faced with many difficulties and challenges. In this paper, a new stereo matching algorithm and algorithm framework is proposed to improve the matching accuracy by using the feature space and the scale space. In view of the randomness and uncertainty of depth enhancement, a new depth enhancement algorithm is proposed in this paper, which highlights the important role of confidence estimation. Specifically, the main contributions of this paper are as follows:? A binary stereo matching algorithm is proposed. In this paper, the binary feature descriptor is introduced into the stereo matching problem, and the parallax consistency is approximated by color consistency. An adaptive pixel pair sampling algorithm is constructed to solve the problem of boundary expansion. Adaptive sampling is realized by binary mask, which can quickly calculate the matching cost by bit operation. In this paper, a weighted voting method is proposed to further improve the matching accuracy. Through the evaluation on the open data set, it is proved that the binary stereo matching algorithm has good accuracy and real-time performance. Illumination robustness and performance controllability.? A multi-scale stereo matching algorithm framework is proposed. In this paper, the cost aggregation in stereo matching is modeled as a weighted least squares problem. Using this model, the advanced cost aggregation algorithm is analyzed. In this paper, the optimization model of cost aggregation is extended to the scale space to obtain the multi-scale cost aggregation optimization model. The canonical terms of the parallax response consistency between the constraint scales are added to the optimization target. Multi-scale information is introduced into traditional cost aggregation methods. Multi-scale cost aggregation optimization model can integrate existing cost aggregation algorithms. A stereo matching algorithm framework is obtained. By comparing and testing on multiple open datasets, the proposed algorithm framework can effectively improve the matching accuracy of existing algorithms. This paper presents a depth enhancement algorithm based on confidence estimation. In this paper, the monocular depth confidence estimation problem is modeled as a double marker Markov random field, which can estimate the depth confidence quickly. In this paper, a confidence based search algorithm is proposed to correct the depth, compared with the traditional joint filtering method. The confidence based search algorithm can deal with large area depth errors in local area. The contrast experiments in depth super-resolution and depth denoising show that the proposed algorithm can obtain more accurate depth information. .
【学位授予单位】：清华大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41

【相似文献】