郑红, 刘振强. 基于精确模型的云台摄像机自标定[J]. 机器人, 2013, 35(3): 326-331,338. DOI: 10.3724/SP.J.1218.2013.00326
引用本文: 郑红, 刘振强. 基于精确模型的云台摄像机自标定[J]. 机器人, 2013, 35(3): 326-331,338. DOI: 10.3724/SP.J.1218.2013.00326
ZHENG Hong, LIU Zhenqiang. Self-Calibration of Pan-Tilt Camera Based on Accurate Model[J]. ROBOT, 2013, 35(3): 326-331,338. DOI: 10.3724/SP.J.1218.2013.00326
Citation: ZHENG Hong, LIU Zhenqiang. Self-Calibration of Pan-Tilt Camera Based on Accurate Model[J]. ROBOT, 2013, 35(3): 326-331,338. DOI: 10.3724/SP.J.1218.2013.00326

基于精确模型的云台摄像机自标定

Self-Calibration of Pan-Tilt Camera Based on Accurate Model

  • 摘要: 为了实现非理想结构云台摄像机的精确自标定,构建了一种云台摄像机的精确模型,提出了基于摄像机旋转自约束特征的内参数及云台结构参数的自标定方法.首先, 基于云台摄像机的结构特性建立了区别于理想模型的精确云台摄像机模型,以描述云台水平旋转轴、竖直旋转轴以及摄像机之间存在的相对方向及位置偏差.然后, 利用云台摄像机的旋转自约束特性,计算旋转过程中的不变量,进而结合配极关系建立绝对二次曲线像的约束方程,通过Cholesky分解获得摄像机的内参数.最后, 在求得的内参数矩阵的基础上,通过旋转轴和截面的投影特性求解云台结构参数.实验结果表明,在0.5像素噪声水平下本文方法焦距标定误差低于0.73%, 主点标定误差低于0.52%,同时实际图像重投影误差均值为2.38,皆优于基于理想模型的自标定方法.整个标定过程只利用云台摄像机自主旋转的几何约束, 不依赖外界场景结构信息或标定物.

     

    Abstract: To realize accurate self-calibration of pan-tilt camera with nonideal structure, an accurate pan-tilt camera model is built, and a self-calibration method for intrinsic camera parameters and pan-tilt structural parameters is proposed based on self-constraint characteristics in rotation. Firstly, according to the structural feature of the system, an accurate pan-tilt camera model, different from the ideal model, is built to describe the relative orientation and position offset among pan axes, tilt axes and the camera. Secondly, fixed entities in rotation are calculated by taking advantage of the self-constraint characteristics in rotation, and then combining with the polarizing constraint, the constraint equations of the image for an absolute conic are established and computed to obtain the intrinsic camera parameters by Cholesky's factorization. Finally, based on the intrinsic matrix of the camera, the pan-tilt structural parameters are solved by using the projection characteristics of the rotation axes and cross sections. Experimental results indicate that, for 0.5 pixel noise level, the calibration errors in focal length are less than 0.73% and in principle points are less than 0.52%, and mean error in reprojection of real images is 2.38, all of which are better than those by self-calibration method based on ideal model. The whole calibration process only utilizes the geometric constraint in active rotation of pan-tilt camera, without the information of outside scene or calibration object.

     

/

返回文章
返回