南方基地知识库

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.