基于卷积神经网络的立体匹配技术研究

发布时间：2018-03-18 15:28

本文选题：立体匹配　切入点：匹配代价　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：双目立体视觉是计算机视觉的关键技术之一,它在传统的三维场景重建、智能机器人导航、目标跟踪以及新兴的无人驾驶汽车、虚拟现实、移动终端等领域都具有广泛的应用。立体匹配作为双目立体视觉中最为重要的一个环节,一直是目前研究的焦点和热点。近年来,深度学习技术快速发展,在图像、语音、文本等领域获得了巨大的成功。卷积神经网络更是结合了深度学习技术、人工神经网络和图像局部关联性等特性,能有效提取图像特征,特别适用于计算机视觉任务,在图像分类、物体检测、目标跟踪等任务中取得了显著的成绩。目前,如何将卷积神经网络应用到立体匹配任务中,获得高准确率、快速的算法,仍是一个有待解决的问题。本文整理和总结了国内外关于深度学习和立体匹配技术的研究成果,以经典的卷积神经网络模型为基础,将其应用到立体匹配任务当中,主要工作如下:1)提出了一种基于卷积神经网络的匹配代价计算算法。在分析传统匹配代价算法原理和现有数据集限制的基础上,本文利用卷积神经网络计算图像块的匹配程度,避免了需要人工选择特征。实验结果表明,本文提出的基于卷积神经网络的匹配代价计算算法大幅减少了初始匹配误差,为后续步骤建立良好基础,且具有较好的光照鲁棒性。同时研究了卷积神经网络超参数对算法效果和速度的影响。2)提出了一种基于卷积神经网络的立体匹配算法框架。本文参考传统立体匹配四个步骤,使用基于卷积神经网络的匹配代价计算算法计算匹配代价,将半全局匹配算法和基于十字的代价聚合算法结合用于代价聚合,视差计算完成后使用多种后处理算法对初始视差图进行修正。通过在公开数据集的测试表明,本文提出的算法框架在准确度和性能上都已经达到了领域内先进水平。
[Abstract]:Binocular stereo vision is one of the key technologies of computer vision. It is used in traditional 3D scene reconstruction, intelligent robot navigation, target tracking, and emerging driverless cars, virtual reality, etc. Stereo matching, as the most important link in binocular stereo vision, has been the focus and hotspot of current research. In recent years, depth learning technology has developed rapidly in image, speech, speech, etc. The convolution neural network combines the characteristics of deep learning technology, artificial neural network and image local correlation, so it can extract image features effectively, especially for computer vision tasks. Significant achievements have been made in image classification, object detection, target tracking and so on. At present, how to apply convolution neural network to stereo matching task to obtain high accuracy and fast algorithm, It is still a problem to be solved. This paper summarizes the research results of depth learning and stereo matching at home and abroad, and applies them to stereo matching task based on classical convolution neural network model. The main work is as follows: (1) A matching cost calculation algorithm based on convolution neural network is proposed. On the basis of analyzing the principle of traditional matching cost algorithm and the limitation of existing data set, this paper uses convolutional neural network to calculate the matching degree of image block. The experimental results show that the proposed algorithm based on convolution neural network greatly reduces the initial matching error and establishes a good foundation for the following steps. At the same time, the effect of hyperparameter of convolution neural network on algorithm effect and speed. 2) A stereo matching algorithm framework based on convolution neural network is proposed. This paper refers to the four steps of traditional stereo matching. The matching cost calculation algorithm based on convolution neural network is used to calculate the matching cost. The semi-global matching algorithm and the cross-based cost aggregation algorithm are combined for the cost aggregation. After the parallax calculation is completed, a variety of post-processing algorithms are used to correct the initial parallax map. The experimental results in the open data set show that the proposed algorithm framework has reached the advanced level of accuracy and performance in the field.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】