2020 年 10 月
图 学 学 报
October 2020
第 41 卷 第 5期
JOURNAL OF GRAPHICS
Vo l . 4 1 N o . 5
收稿日期:2020-04-11;定稿日期:2020-05-28
Received:11 April,2020;Finalized:28 May,2020
基金项目:国家重点研发计划项目(2016YFB0501805)
Foundation items:National Key Research and Development Programs of China (2016YFB0501805)
第一作者:陈 炎(1994),男,河北衡水人,硕士研究生。主要研究方向为机器视觉。E-mail:13122358665@163.com
First author:CHEN Yan (1994–), male, master student. His main research interests cover machine vision. E-mail:13122358665@163.com
通信作者:杨丽丽(1974–),女,内蒙古赤峰人,副教授,博士。主要研究方向计算机网络与智能信息处理。E-mail:llyang@cau.edu.cn
Corresponding author:YANG Li-li (1974–), female, associate professor, Ph.D. Her main research interests cover computer networks and intelligent
information processing. E-mail:llyang@cau.edu.cn
双目视觉的匹配算法综述
陈 炎, 杨丽丽, 王振鹏
(中国农业大学信息与电气工程学院,北京 100083)
摘要:双目视觉是获取对现实世界立体感知的重要方法,在自动驾驶等领域得到了普遍
的应用。立体匹配是实现双目感知的前提,该算法对左右摄像机拍摄的照片进行像素级的匹配,
生成稠密视差图,从而获取了三维坐标信息。概述了立体匹配算法近 20 年来的发展过程,围绕
基于人工特征和深度学习两个方向进行了综述,对算法实现过程中的代价计算、代价聚合、视
差计算和视差求精进行分析讨论,评估了算法的准确性和时间复杂度。最后总结了立体匹配算
法面对的挑战和对未来发展的展望。
关键词:立体匹配;双目视觉;立体感知;深度学习;计算机视觉
中图分类号:TP 391.4 DOI:10.11996/JG.j.2095-302X.2020050702
文献标识码:A 文 章 编 号:2095-302X(2020)05-0702-07
Literature survey on stereo vision matching algorithms
CHEN Yan, YANG Li-li, WANG Zhen-peng
(College of Information and Electrical Engineering, China Agricultural University, Beijing 100083, China)
Abstract: Stereo vision is an important means to acquire stereoscopic perception of the real world,
which is widely used in autonomous driving and other fields. Stereo matching algorithm serves as the
basis for the binocular system to realize perception. This algorithm matches the images captured by
the left and right cameras and calculates the depth map, thus providing coordinates for the 3D
modeling. There are four stages of stereo matching algorithm, including cost computation, cost
aggregation, disparity selection, and disparity refinement. This literature survey summarized the
development of stereo matching algorithm from such two research fields as artificial feature and deep
learning in the past 20 years, evaluated the accuracy and time complexity, and presented the
challenges and prospects of the development of stereo matching algorithms.
Keywords: stereo matching; stereo vision; stereo perception; deep learning; computer vision
立体视觉是机器认识世界的重要手段。利用
2 幅图像形成的视差可以计算得到像素点的深
度,从而获得三维的立体感知
[1-2]
。其工作流程主
要包含 4 步:双目标定、图像校正、立体匹配和三
维重建,而立体匹配是关键的步骤
[3]
。
立体匹配算法可以实现 2 幅图像像素级的匹