摘要
计算机视觉中,摄像机标定作为摄像测量技术的前提,是必不可少的一个环节。针对目前基于神经网络的相机标定方法训练精度不够高的问题,提出了一种基于双神经网络的相机标定方法。该方法从成像模型出发,推导出相机坐标Z_(c)是世界坐标Z_(w)和像素坐标u,v的函数,在考虑了Z_(c)的变化的基础上,将成像模型简化成两个函数关系式,使用两个神经网络进行标定,分化了单个神经网络的任务量的同时又充分遵循了成像模型。实验结果表明,较其余基于神经网络的相机标定方法,该方法提高了相机标定的精度,在400 mm×300 mm标定范围内平均标定误差为0.178 6 mm,验证了所提方法的可行性和有效性。
In computer vision,camera calibration as the premise of camera measurement technology,is an essential part.Aiming at the problem that the training accuracy of camera calibration method based on neural network is not high enough,a camera calibration method based on double neural network was proposed.Starting from the imaging model,it was deduced that the camera coordinate Z_(c) was a function of the world coordinate and the pixel coordinate.On the basis of considering Z_(c),the imaging model was simplified into two function relations,and two neural networks were used for calibration,which not only differentiated the task amount of single neural network,but also fully followed the imaging model.The experimental results show that compared with other calibration methods based on neural network,this method improves the accuracy of camera calibration.And the average calibration error is 0.178 6 feasibility and effectiveness of proposed method.
作者
陈文艺
许洁
杨辉
Chen Wenyi;Xu Jie;Yang Hui(Industry School of Modern Post,Xi’an University of Posts and Telecommunications,Xi’an 710061,China;Collaborative Innovation Center for Modern Post,Xi’an University of Posts and Telecommunications,Xi’an 710121,China)
出处
《红外与激光工程》
EI
CSCD
北大核心
2021年第11期284-292,共9页
Infrared and Laser Engineering
关键词
计算机视觉
神经网络
相机标定
成像模型
computer vision
neural network
camera calibration
imaging model
