-
为了验证文中提出算法,设计了数值仿真分析与实际标定试验。
-
根据对两种多摄像机标定算法的描述,多摄像机标定误差主要来源于图像坐标提取误差、摄像机标定内部参数误差。该节将通过仿真实验研究定位噪声与原理误差对多摄机标定精度的影响。
实验中,摄像机的分辨率为2048×2048 pixel,4个摄像的内部参数
$ \left( {\alpha,\;\beta,\; \gamma,\;{u_0},{v_0}} \right) $ 均设置为:(1100, 1100,0, 1024, 1024),一阶畸变系数${k_{i1}} = 0.01$ ,二阶畸变系数${k_{i2}} = 0.001$ 。如图2所示,将4个摄像机设置在边长为6000的正方形$ {O_1}{O_2}{O_3}{O_4} $ 的4个顶点处,摄像机光轴与正方形的对角线重合;一维靶标杆AC之间的距离$AC = {\text{5}}00\;{\rm{mm}}$ 。仿真实验中利用残差与重建误差来验证文中所提算法的优越性,其中残差公式定义为:
式中:
${d_x} = {x_i} - {\hat x_i}$ ;${d_{x'}} = {x'_i} - {\hat x'_i}$ ;${d_{x''}} = {x''_i} - {\hat x''_i}$ ;${d_{x'''}} = {x'''_i} - {\hat x'''_i}$, ${x_i}$ ,$ {x'_i} $ ,${x''_i}$ ,${x'''_i}$ 表示特征点在4个摄像机中的同名成像点;${\hat x_i}$ ,$ {\hat x'_i} $ ,${\hat x''_i}$ ,${\hat x'''_i}$ 表示根据求解得到的摄像机矩阵将空间三维特征点反向投影到4个摄像机中的成像点。重建误差公式定义为:
式中:
$\left( {{x_s},{y_s},{z_s}} \right)$ 表示仿真生产的空间三维点坐标;$\left( {{x_r},{y_r},{z_r}} \right)$ 表示三维重建生产的空间点坐标。 -
每次标定实验时,令线段AC在正方形中心附近做20次任意的刚体运动,为了保证算法的统计特性,在每一固定成像误差下重复进行100次标定实验,且保证标定点都在4个摄像机的成像平面内。给图像坐标添噪声水平在0.0~1.1 pixel的成像误差。统计每一误差等级下标定杆长度解算误差模的均值。其算法精度对比实验结果如图3所示。
根据图3可知:
(1) 五种摄像机的标定误差等级都随图像坐标定位结果的信噪比的减小而增大,但平面法无法同时被4个摄像机同时观测到,因此在某些视角下无法实现四摄像机测量系统的标定;
(2) 相比与其他摄像机标定算法,当图像坐标定位精度为1 pixel时,文中提出的多摄像机标定算法的反向投影误差保持在3 mm以内;
(3) 加权基本矩阵法、三焦点张量法与文中提出的算法之间的标定大致相当,但由于文中所提出的四摄像机标定算法直接分解四焦点张量即可实现,因此其效率更高。
-
6426个标定点被生成且均匀分布在一个半径为2 000 mm的球体内,重建不确定性大小为1.0 pixel,均值为0。每一个点的重建误差
$er{r_{rec}}$ 被统计,其归一化的误差统计结果如图4所示。其统计参数如表1所示。Re-projection coordinate/mm Proposed Weighted Trifocal Fundamental Mean 2.31 2.27 2.99 3.59 Std 0.03 0.05 0.35 0.62 Max 2.40 2.50 4.33 8.11 RMSE 1.52 1.50 1.73 1.89 Table 1. Error statistics of quad-camera calibration results
根据图4可知,文中提出的算法误差最为集中,主要集中在2.2~2.4 mm之间,且呈现正态分布的形式,这说明文中的方法对于每一误差项的影响是均匀的。其他算法都有一个拖尾效应,在其误差分布上可以明显地看出耦合的有指数分布,这说明不同的位置,其误差因素的影响是不一样的,因此文中提出的方法更加得合理。
表1给出了四种一维标定算法误差的均值、标准差、最大值、以及均方根,对比结果可以得出,文中提出的算法的误差均值与均方根相比于加权基本矩阵法稍大以外,其他统计指标都更小,这说明了文中提出的算法即可达到加权基本矩阵法的精度,又具有更集中的误差分布,及更快的标定效率。
-
为了验证算法的可行性与精度,搭建了如图5(a)所示多摄像机测量系统。该系统由4台摄像机、4个LED红外光源,4个磁盘存储阵列、一个T型一维标定杆,一个相机同步控制器和1台计算机组成;其中摄像机的型号为IO Industries Flare 4 M140,CCD分辨率为2 048×2 048 pixel,像元大小为0.0055 mm;镜头为焦距为25 mm的CHIOPT HC1605 A;红外光源波长为850 nm;一维标定杆上标定球直径为d=12.7 ,标定杆长度500 mm,其表面的红外反光材料为3 M7610;系统标定时将4个摄像机如图5(b)所示的布局方式固定在衍架上。
-
如图6所示,在测量空间中挥舞标定杆80次,使得标定点均匀地分布在测量空间内,并采集其图像。标定点图像坐标采用参考文献[14]提出的方法进行定位,其定位误差为1 pixel。
-
根据上述摄像机的内部参数以及采集到的1D标定杆图像坐标,然后利用文中提出的四摄像机标定方法进行标定,其结果(每一次挥杆两靶标点之间距离误差统计)如图7、图8所示。其中摄像机的内部参数由二维标定算法[5]分别进行精确标定得到。系统中每一台摄像机的内部参数如表2所示。摄像机的外部参数如表3所示。
Figure 7. Statistical error chart of spatial point locating of proposed multi-camera calibration method
Figure 8. Statistical error chart of spatial point locating of proposed multi-camera calibration method
(fu, fv) (u0, v0) (k1, k2) Camera1 [4542.585 4542.015] [1020.998 1014.689] [−0.081 0.813] Camera2 [4598.596 4596.746] [1050.463 1051.540] [−0.067 0.451] Camera3 [4595.456 4596.327] [1052.499 1030.833] [−0.066 0.447] Camera4 [4587.792 4587.845] [1038.169 1025.135] [−0.079 0.646] Table 2. Intrinsic parameters of the four cameras
Index Parameters Test environment 4 m×4 m×2 m Sample frequency 100 Hz Exposure time 1 ms Camera1 projection matrix $ { {{\boldsymbol{P}}}_1} = \left[ {\begin{array}{*{20}{c} } 1&0&0&0 \\ 0&1&0&0 \\ 0&0&1&0 \end{array} } \right] $ Camera2 projection matrix $ { {\boldsymbol{P} }_2} = \left[ {\begin{array}{*{20}{c} } { {\text{0} }{\text{.002} } }&{1.000}&{0.001}&{ - 32.651} \\ { - 1.000}&{0.002}&{ - 0.002}&{80.536} \\ { - 0.002}&{ - 0.001}&{1.000}&{4\;832.358} \end{array} } \right] $ Camera3 projection matrix $ { {\boldsymbol{P} }_3} = \left[ {\begin{array}{*{20}{c} } { -0.999 }&{0.017}&{0.000}&{ - 11.579} \\ {0.017}&{0.999}&{ - 0.001}&{3.798} \\ {0.000}&{ - 0.001}&{ - 1.000}&{7\;988.482} \end{array} } \right] $ Camera4 projection matrix $ { {\boldsymbol{P} }_4} = \left[ {\begin{array}{*{20}{c} } { {\text{0} }{\text{.002} } }&{1.000}&{0.001}&{ - 14.485} \\ { - 1.000}&{0.002}&{ - 0.002}&{ - 26.452} \\ { - 0.002}&{ - 0.001}&{1.000}&{4\;615.600} \end{array} } \right] $ Table 3. Experimental test parameters
(1) 文中提出的四摄像机标定算法对空间标准杆长度测量误差小于4 mm(3σ),满足室内运动目标的三维定位需求;
(2) 实际测量值大于仿真分析所得值的原因是因为现场光照条件复杂,标定靶标点的图像灰度分布不满足高斯分布的假设,且对摄像机的内部参数标定也存在误差,在利用这些数据进行标定时必然会将误差传递到定位结果上。
Calibration method of quad-camera measurement system based on tensor decompose
doi: 10.3788/IRLA20220103
- Received Date: 2022-02-14
- Rev Recd Date: 2022-03-24
- Publish Date: 2022-09-28
-
Key words:
- measurement /
- machine vision /
- multi-camera calibration /
- 1D calibration /
- quadrifocal tensor
Abstract: Camera calibration is essential for vision measurement. In order to realize the accurate calibration of the multi-camera measurement system, the existing multi-camera calibration methods based on 1D (one-dimensional) targets are analyzed, and a multi-camera calibration method based on tensor decomposition is proposed. The method includes three aspects: (1) A mathematical model of multi-camera calibration is established with pin-hole model and rigid body transformation theory; (2) A problem of multi-camera calibration approach based on the fundamental matrix is analyzed, which is the calibration results coupled with each other; (3) In order to address this problem, the quad-focal tensor of a quad-camera measurement system is introduced to the calibration process, which includes that the quad-tensor is solved by using the image point of 1D calibration objects, and a reduced method for the quad-focal tensor is used to refine the camera matrices. Finally, the effectiveness and accuracy of the method are verified by experiments. The results indicate that (1) The calibration of the quad-camera 3D measurement system can be realized by only 3 groups of calibration target images, and the calibration operation efficiency is higher than that of the basic matrix method, and (2) the accuracy of the multi-camera measurement system reaches 4 mm (3σ) in the range of 4000 mm×4000 mm×2000 mm, which is more accurate than traditional approaches with only 3 groups calibration images. The proposed method meets the requirements of accurate measurement of indoor moving target pose.