基于超分辨率角点检测算法的双目相机高精度标定方法  

High-Precision Binocular Camera Calibration Based on Super-Resolution Corner Point Detection Algorithm

作　　者：孙婷 陈娜 孟然 龙潜 谢启伟 Sun Ting;Chen Na;Meng Ran;Long Qian;Xie Qiwei(School of Mathematics and Statistics,Hubei University,Wuhan 430062,Hubei,China;Beijing Smarter Eye Technology Co.,Ltd.,Beijing 100023,China;Yunnan Observatories,Chinese Academy of Sciences,Kunming 650000,Yunnan,China;Research Base of Beijing Modern Manufacturing Development,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]湖北大学数学与统计学学院,湖北武汉430062 [2]北京中科慧眼科技有限公司,北京100023 [3]中国科学院云南天文台,云南昆明650000 [4]北京工业大学北京现代制造业发展研究基地,北京100124

出　　处：《激光与光电子学进展》2023年第8期355-363,共9页Laser & Optoelectronics Progress

基　　金：国家重点研发计划(2018AAA0103103);国家自然科学基金(61673381)。

摘　　要：双目相机标定是研究立体视觉的基础工作,标定的精度是视觉测量精度的关键。图像角点提取是相机标定的基础,但在现实应用场景中,外界影响使获取的图像不清晰,导致检测到的角点精度低,从而影响标定的精度。因此,提出一种基于超分辨率亚像素角点检测的端到端算法,从特征级解决低质量角点检测问题。首先,应用盲超分部分估计低分辨率图像模糊核,融合低分辨率图像特征重建出高分辨率图;然后,在此基础上得到角点亚像素位置;最后对双目相机进行高精度标定,并用测距实验对其进行检验。实验结果表明,所提基于超分辨的亚像素角点检测方法在真实场景下具有优越性。Binocular camera calibration is the foundation of stereo vision research,and its accuracy is the key to achieve precision in vision measurement.The basis of camera calibration is image corner extraction,but in real applications,the accuracy of detected corner points is often low due to unclear acquisition of images due to external influences,which affects calibration accuracy.To solve the problem of low-quality corner detection in terms of the feature level,an end-to-end algorithm based on super-resolution subpixel corner detection is proposed.First,the fuzzy kernel of a low-resolution image is estimated using the blind hyperspectral part,and its features are fused to reconstruct its high-resolution version.Subsequently,the subpixel position of the corner points is obtained.Finally,a binocular camera is calibrated with high accuracy and tested via ranging experiments.Experimental results show that the proposed subpixel corner detection method based on super-resolution has advantages in real scenarios.

关 键 词：成像系统 相机标定 角点检测 超分辨率 亚像素 自监督 深度学习 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]