基于神经网络的虚拟靶标大视场双目相机标定技术

针对大视场双目相机标定中的精度低和非线性畸变问题,提出了一种结合BP神经网络的大尺寸虚拟靶标标定技术。鉴于单角点棋盘格具有易制作、高精度的特性,构建大尺寸虚拟靶标;利用神经网络的非线性映射能力直接建立角点的像素坐标和世界坐标映射关系;用建立的映射网络对测试样本进行三维重建,并与传统的线性标定方法进行对比实验。结果表明,该方法操作简单,且重建距离相对误差为0.92%,优于传统的线性标定方法,可用于大视场双目相机标定。

Large-scale binocular camera calibration combining neural network with virtual target

Aiming at the problem of low precision and nonlinear distortion of binocular camera calibration in large large-scale, a new calibration technique is proposed which combines the BP neural network and the large dimension virtual target.In view of the easy production and high precision characteristic of single corner board, a large size virtual target is built.By using the neural network that has a nonlinear mapping ability, the mapping relationship between the pixel coordinates and the 3D coordinates of corner is directly established.The test samples are reconstructed with the established mapping network, and the results are compared with the traditional linear calibration method.The results show that the method is simple in operation and the relative error is 0.92%, which is better than the traditional linear calibration method, and can be used for the binocular camera calibration in large-scale.