|
1.IntroductionVision sensors have advantages of flexibility and noncontact measurement. Planar mirrors are utilized in a wide variety of applications, such as fringe-reflection measurement,1 planar catadioptric vision sensors,2 stereovision by a single camera, and calibration of multiple cameras with nonoverlapping fields of view (FOVs).3,4 A common major challenge of these applications is the pose estimation of the planar mirror. It is important and necessary to develop accurate and flexible pose estimation methods for planar mirrors. There has been some research on the pose estimation of planar mirrors. Lebraly et al.5 used a planar mirror to create an overlap between cameras by attaching markers on the mirror’s surface. However, pasting markers on the mirror was inconvenient and troublesome. Sturm et al.,6 Rodrigues et al.,7 Hesch et al.,8 and Kosuke Takahashi et al.9 proposed different pose estimation methods based on the fact that successive reflections in two-mirror planes were a fixed-axis rotation around the line of intersection of two mirror planes. Reference 10 proposed a practical visual inertial calibration method, which is based on visual observations taken in front of a planar mirror. Reference 11 presented a design and construction of a new system for stereovision with a single CCD camera. Reference 12 proposed a method for the calibration, which is used the orthogonal phase-shifted sinusoidal patterns during the calibration process. Xiao et al.1 calculated the transformation between the camera and the LCD screen by a markerless flat mirror. Li et al.13 measured the pose of a flat mirror by capturing the reflected structured light pattern. The translation vector was calculated by the mirror motion without any tilt. However, pure translations can hardly be achieved without external equipment. Ren et al.14 calibrated a stereodeflectometry system, where two cameras captured the fringe patterns produced by an LCD screen reflected by a flat mirror. The above methods require that the mirror is placed at several positions. However, these methods fail when the camera and the mirror are in fixed positions. In Ref. 15, an algorithm is based on an orthogonality constraint, which should be satisfied by all families of reflections of a single reference object. It can return a unique solution of three P3P problems from three mirrored images. Reference 16 presents a method that utilizes reasonable assumptions about geometric features such as a plane or an intersecting line of two planes. It obtains an initial solution via singular value decomposition (SVD), which is then refined via nonlinear optimization, which does not require any assumption about structures but estimates the parameters of geometric features. In Ref. 17, an accurate closed-form solution with explicit geometric meanings is proposed. It minimizes an error metric based on all reflections of rotation, and then derives the optimal solution that gives the real camera rotation directly by SVD of the sum of mirrored camera rotations. Then, it can minimize object-space collinearity error. This paper is also inspired by Shah18 and Li et al.17 Shah found the best rotation matrix with determinant 1. In contrast, we seek the best orthogonal matrix with determinant , and the best reflection matrix can be constructed by a matrix with determinant . Li et al.17 calculated the camera rotation by minimizing the error of all rotation reflections. In this paper, we initialize the normal vector by minimizing the error of reflections of the unit direction vectors. In this paper, we focus on the pose estimation of the planar mirror based on a one-dimensional (1-D) target. 1-D target has advantages of simple structure and flexibility. First, instead of moving the mirror, we move the target in front of the camera to provide multiple (at least two) views of both the original and virtual targets. The camera and the mirror are fixed during the process. Next, the normal vector of the planar mirror is estimated based on the direction vectors of the original and virtual targets. The distance between the camera and the mirror is calculated based on reflection equations. Last, the closed-form solution thus obtained is refined using nonlinear optimization by minimizing the reprojection error of reflection projection. Compared with conventional methods, the proposed method is suitable for the pose estimation of planar mirrors, especially when camera and the mirror are in fixed positions. After obtaining the position information in the plane mirror, the camera can be observed the carrier virtual image through plane mirror, then the rotation matrix, and translation vector of the image machine coordinate system and the carrier coordinate system can be calculated. The calibration of the installation parameters of the image machine is completed. The rest of this paper is organized as follows: preliminary work is introduced in Sec. 2. The principle of the proposed method is described in Sec. 3. Synthetic data and real-data experiments are performed in Sec. 4. The conclusions are given in Sec. 5. 2.Preliminaries2.1.Camera ModelPerspective projection is used as the camera model. The parameters and coordinates of the perspective projection model are shown in Fig. 1. The relation between a spatial point in and the corresponding image point can be described as follows: where is a scale factor, is the intrinsic matrix in camera. and are the equivalent focal lengths in horizontal and vertical directions, respectively. () denotes the principal point.As shown in Fig. 2(a), both the original and the virtual targets are in the camera’s FOV. is the coordinate frame of the camera, denoted as . 2.2.Pose Estimation of One-Dimensional TargetAs shown in Fig. 2, and denote the spatial coordinates of the ’th original and the ’th virtual feature points at the ’th position in , respectively. and are the corresponding image coordinates, , . is the number of target positions and is the number of feature points on the target. The original target and the virtual target at ’th position are uniquely characterized by , and , respectively. and are the unit direction vectors of the original and virtual targets at th position, respectively. For the th position, and can be obtained by the method proposed by Liu et al.19,20 It requires that at least three collinear feature points on the target. Suppose , , and , then where can be computed from the distances between the points , , .Suppose that in the image coordinate system, , , and . Combining Eqs. (1) and (2), we can obtain By solving Eq. (3), the coordinates of , , and can be retrieved and the initial value of the normal vector can be obtained. Due to the image noise, if there are more than three feature points, results of different point triplets, e.g., () and (), are not equivalent. Let be the image coordinates of the reprojection feature point where is a scale factor, is the known interval of two adjacent feature points on the target, and denotes the intrinsic matrix in camera.To utilize the redundant information of feature points, and are refined by minimizing the following reprojection error using Levenberg–Marquardt algorithm:21 where . . and are the pitch angle and yaw angle of the ’th unit direction vector with respect to , respectively. is the distance function between two points. The initial value of () can be obtained by Liu’s method.19The pose of the ’th virtual target can also be obtained similarly. 2.3.Planar Mirror ReflectionAs shown in Fig. 2(b), camera is in front of a planar mirror . The mirror is uniquely determined by its unit normal vector and the distance with respect to .2,17 A spatial point lies on the mirror plane if and only if it satisfies The Householder matrix 22,23 is widely used in the mirror reflection model, and its equation is as follows: The Householder transformation is a linear transformation that transforms a vector into a mirror image reflected by a hyperplane. The Householder transformation can zero some elements of a vector and keep the norm of the vector constant. The schematic diagram of a mirror surface emission model is shown in Fig. 3. The equation of the plane is as follows: If , then If , then The reflection in can be represented by the following matrix: Through the conversion of can obtain the position of the virtual mirror corresponding to the space point in the camera coordinate system. The space point and the corresponding virtual point are related as follows: 3.Principle of the Mirror Pose Estimation3.1.Analysis of the Unique SolutionThe calculation of the normal vector is equivalent to solving the Householder matrix . According to Eqs. (11) and (12), and are related by the Householder matrix as follows: From Eq. (13), one position is not enough to obtain a unique solution. For two positions, we have According to the principle of mirror image, assume that target real image is positioned for the right-handed coordinate system, and in the target virtual image, the left and right, upside down, and up and down directions remain the same with respect to the left-hand target image coordinate system.4 Therefore, the relationship between real image and virtual image is as follows: Combining Eqs. (14) and (15), we have If is not parallel to , matrix is invertible. Thus, can be uniquely determined as follows: If is parallel to , the target motion is a pure translation, det , thus a unique solution cannot be found. 3.2.Computing the Normal Vector nIf the target is placed more than two times, the best Householder matrix can be computed from the solution to the orthogonal Procrustes problem.21 According to Eq. (13), the sum of the normal vector error is where denotes the Frobenius norm, . If measurements are noiseless, .The best approximation of matrix is equivalent to the matrix that solves the minimization problem of Eq. (18). According to the properties of the Householder matrix, Eq. (18) is expanded as where .It can be seen that minimization of is equivalent to maximize . Therefore, the matrix that solves the minimization problem of Eq. (18) is equivalent to that solves There are some researches18,22 on finding the best rotation matrix with determinant . However, the Householder matrix is not a rotation but actually a reflection with determinant . To construct the best reflection, a lemma is introduced. Lemma 1For a given matrix and a Householder matrix where is the full SVD of , the singular values are in decreasing order, andAccording to Eqs. (24) and (25), we can come to the conclusion that is an orthogonal matrix with determinant 1 and hence a rotation matrix. Suppose that , , we have Therefore, the trace is maximized if . As is an orthogonal matrix, this means that would have to be the identity matrix. Then According to two sets of the original and virtual direction vectors, the best reflection matrix can be obtained by Theorem 1. Theorem 1.The solution to the maximization problem of Eq. (20) is as follows: where is the full SVD of with decreasing singular values and is subject to Eq. (22).Proof.From Lemma 1, the maximization problem of Eq. (20) is achieved if a Householder matrix can be constructed such that Let , thenWhen the Householder matrix is determined, the unit normal vector is obtained. 3.3.Computing the Distance lFrom Eq. (12), we have Therefore, the distance can be computed as where and can be calculated according to the Eq. (3). can be calculated according to Eq. (17).3.4.Nonlinear OptimizationThe results obtained in Sec. 3.2 and Sec. 3.3, which were called as the closed-form solution. The closed-form calculation is fast, but it is sensitive to the noise. In this subsection, the closed-form solution is refined with nonlinear optimization by minimizing the reprojection error of reflection. According to the camera model, we have where s is a scale factor, denotes the intrinsic matrix in camera, is a identity matrix, and 0 is a zero vector. is the reprojected image coordinates.The mirror normal vector and the distance are refined by minimizing the following function using the Levenberg–Marquardt algorithm21 where , . and are the pitch angle and yaw angle of the normal vector with respect to , respectively. is the distance between image points and is the image coordinate corresponding to the imaginary point . After optimization, the optimal value of the normal vector and the distance is obtained.3.5.Calculating the Relative Position of the Image Machine and the CarrierFigure 4 shows the relationship diagram of the vector coordinate system , the vector image frame , and the machine coordinate system , where the and stand for the rotation vector, the and represent the translation vector, and the mean mirror reflection matrix. The image of the mounting carrier in the mirror is observed by the plane mirror, so as to obtain the transformation relationship between the plane frame and the vector virtual frame . According to the above results, the transformation relationship between the image coordinate system and the carrier image coordinates can be obtained directly. Thus, the calibration of the installation parameters of the imager can be completed. where is the rotation matrix, is the translation vector, subscripts and , respectively, represent the conversion relationship between the camera coordinate system and the virtual carrier coordinate , and the transformation relationship between the camera coordinate system and the carrier image coordinates .3.6.Method SummaryThe calibration method of the installation parameters of the image machine based on the plane mirror is proposed in this section, which keeps the relative position of the image machine and the plane mirror unchanged. The main processes of the proposed method are as follows:
4.Experimental ResultsThrough the previous derivation, the mirror position (the mirror vector and the image machine-mirror distance ) are the key parameters in the calibration process of the installation parameters of the image machine. Therefore, this section takes the pose error of plane mirror as the accuracy evaluation index, analyzes the related factors that affect the accuracy through simulation data and physical experiments, and compares it with other calibration methods. 4.1.Simulation DataTo evaluate the accuracy, the estimation errors are defined as follows:
In this experiment, one simulated camera and a planar mirror are used to construct the system. The intrinsic parameters of the camera are , , . The image resolution is . The mirror normal vector , i.e., , . The distance , the interval . The positions of target are randomly generated. Only the original and virtual targets in the camera’s FOV are used. Gaussian noise with zero mean and different standard deviation σ is added to the image coordinates of feature points. The effects of several factors on the accuracy of the pose estimation are evaluated quantitatively by synthetic data as follows. 4.1.1.Influence of the characteristic point noise on the accuracyIn this experiment, , . Gaussian noise is added to the image coordinates of the feature points. The noise level varies from 0 to 1.0 pixel. For each level, 100 independent trials are performed to compute the relative errors of , and RMS errors of . From Fig. 5, the error increases linearly with the noise level. It indicates that the closed-form solution can be effectively adjusted by the proposed method. If the noise level of a system is , relative errors of the normal vector and the distance are and 0.5%, respectively, which are applicable for common applications. 4.1.2.Influence of the number of target placement on the accuracyIn this experiment, Gaussian noise with is added to the image coordinates of the feature points. varies from 2 to 12, is set to 9. For each level, 100 independent trials are performed and are used to compute relative errors. From Fig. 6, the error decreases with the increasing number of target positions. When the target is placed only twice, the pose estimation error is rather high. The estimation error decreases rapidly as the number of positions increases. When there are enough positions, the contribution of additional positions decreases. 4.1.3.Influence of the number of target feature points on the accuracyIn this experiment, Gaussian noise with is added to the image coordinates of the feature points. is set to 10, varies from 3 to 16. For each level, 100 independent trials are performed and are used to evaluate relative errors. From Fig. 7, increasing number of feature points benefits the estimation accuracy. When there are only three feature points on the target, the estimation error reaches maximum. The accuracy improves sharply when increases from 3 to 8. When , the accuracy stables at the certain level. 4.2.Physical ExperimentsAs shown in Fig. 8(a), a Canon 60D digital camera and a planar mirror are in fixed positions. The image resolution of the camera is . The intrinsic parameters of the vision sensor are calibrated using the Bouguet’s calibration toolbox,15 as shown in Table 1. We use the intersection of two rows of checkerboards equivalently to the 1-D target. The distance between neighboring feature points in the checkerboard is 20 mm, i.e., . The two-row checkerboard is used to define 1-D target for feature extraction. Table 1Intrinsic parameters of the camera.
As shown in Fig. 8(b), a checkerboard is placed between the camera and the mirror. Both the original and virtual checkerboards are in the camera’s FOV. The spatial coordinates of the corner points of the checkerboards can be obtained by the Bouguet’s toolbox. In this experiment, reflective error and the reprojection error are used to evaluate the accuracy where and are the spatial coordinates of the original and virtual corner points of the checkerboards, with respect to . and are the number of squares along two orthogonal directions, respectively.To verify the effect of several factors shown in the synthetic experiment, pose estimations of different position numbers and different feature points are implemented. 4.2.1.Influence of the number of target placement on the accuracyIn this experiment, the number of target positions is changed from 2 to 12, the number of feature points is set to 9. From Fig. 9, the result is similar to the result of the synthetic experiment in Sec. 4.1.2. An increasing number of target positions benefits the accuracy, especially when the number of target positions is relatively small. 4.2.2.Influence of the number of target feature points on the accuracyIn this experiment, the number of target positions is set to 10, the number of feature points is changed from 3 to 12. From Fig. 10, it also indicates a similar result of the synthetic experiment in Sec. 4.1.3. The estimation error peaks at , which corresponds to the minimum quantity of feature points. Then, the estimation error decreases sharply with the increasing number of feature points. When , the estimation error flattens out gradually. 4.2.3.DiscussionsEvidently, there exists a trade-off between the accuracy and the workload. According to the effect, analysis of the number of target positions and the number of feature points , and is a balanced choice. When the error of an image point is 0.5 pixel, the synthetic experiment shows that relative errors of the normal vector and the distance are 0.06% and 0.45%, respectively. It is acceptable for common applications such as the camera-to-base calibration.8 In addition, there is a gap between the results of synthetic data and real-data experiments. There may be some reasons for this. First, the feature point extraction method is a commonly used method with good generality and acceptable accuracy. Extraction algorithms with higher accuracy improve the accuracy, which will be further investigated. Second, the mirror used in this paper is a domestic appliance, not a precision optical device. Precision mirrors can be used to improve the accuracy. 4.3.ComparisonsThere are many methods for the pose estimation. Here, three typical methods are described as follows:
For each method, 10 images are captured to estimate the mirror pose. The reflective error Es is used to evaluate the accuracy. The results of different methods are shown in Table 2. Table 2The estimation errors of different methods.
From Table 2, the comparisons show that the proposed method achieves a good accuracy. Compared with the closed-form solution, the closed solution is simple and fast, but it is more sensitive to noise and larger in error. The same line determined by the selection of different points is different. Therefore, the normal vectors determined by these lines will also have certain differences, which will produce certain errors. Compared with the checkerboard of method 1, the errors of the two methods are relatively small. However, the number of feature points of 1-D target is reduced a lot, under the condition that the computation accuracy is not decreased. Therefore, according to the 1-D target of the known size, the calibration of the imaging machine can be achieved without reducing the precision, and the limitation of the complex calibration tool can be removed, so that the calibration is more flexible and convenient. The accuracy of method 3 is roughly the same as the closed-form solution. However, method 3 fails when the camera and the mirror cannot be moved. 5.ConclusionIn this paper, a pose estimation method for planar mirrors is presented. The method only needs a simple 1-D target and a digital camera. First, the camera captures the target in different positions. The pose of the target in each position is obtained. Then, the normal vector and the distance between the camera and the mirror are calculated by a closed-form method, followed by a nonlinear optimization. Our method is simple and flexible. A minimum of two nonparallel target positions and three collinear feature points on the target are needed. No prior information about the target motion and the camera’s pose is required. The proposed method is accurate and robust to noises. Both synthetic data and real experiments demonstrate the effectiveness of our method. In addition, both the camera and the mirror do not need to move during the process, so it is suitable for application where the camera and the mirror need to remain fixed. In some applications, the transformation between the body frame and the camera frame need to be accurately calibrated. However, cameras mounted on the body usually cannot observe the body directly. We plan to apply our method for the camera-to-body calibration in the future. AcknowledgmentsThe authors acknowledge the support by the Industrial Technology Development Program under Grant No. B1120131046. ReferencesY. Xiao, X. Su and W. Chen,
“Flexible geometrical calibration for fringe-reflection 3D measurement,”
Opt. Lett., 37
(4), 620
–622
(2012). https://doi.org/10.1364/OL.37.000620 OPLEDP 0146-9592 Google Scholar
G. L. Mariottini et al.,
“Planar mirrors for image-based robot localization and 3-D reconstruction,”
Mechatronics, 22 398
–409
(2012). https://doi.org/10.1016/j.mechatronics.2011.09.004 MECHER 0957-4158 Google Scholar
X. Feng and D. Pan,
“Research on the application of single camera stereo vision sensor in three-dimensional point measurement,”
J. Mod. Opt., 62
(15), 1204
–1210
(2015). https://doi.org/10.1080/09500340.2015.1024775 JMOPEW 0950-0340 Google Scholar
R. K. Kumar et al.,
“Simple calibration of non-overlapping cameras with a mirror,”
in Proc. of IEEE Conf. on Computer Vision and Pattern Recognition,
2602
–2608
(2008). Google Scholar
P. Lebraly et al.,
“Flexible extrinsic calibration of non-overlapping cameras using a planar mirror: Application to vision-based robotics,”
in Proc. of IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, (IROS, 2010),
5640
–5647
(2010). Google Scholar
P. Sturm and T. Bonfort,
“How to compute the pose of an object without a direct view?,”
Lect. Notes Comput. Sci., 3852 21
–31
(2006). https://doi.org/10.1007/11612704_3 Google Scholar
R. Rodrigues, J. P. Barreto and U. Nunes,
“Camera pose estimation using images of planar mirror reflections,”
Lect. Notes Comput. Sci., 6314 382
–395
(2010). https://doi.org/10.1007/978-3-642-15561-1_28 Google Scholar
J. A. Hesch, A. I. Mourikis, S. I. Roumeliotis,
“Mirror-based extrinsic camera calibration,”
Algorithmic Foundations of Robotics VIII, 57 Springer, Berlin, Heidelberg
(2009). https://doi.org/10.1007/978-3-642-00312-7_18 Google Scholar
K. Takahashi, S. Nobuhara and T. Matsuyama,
“Mirror-based camera pose estimation using an orthogonality constraint,”
IPSJ Trans. Comput. Vision Appl., 8 11
–19
(2016). https://doi.org/10.2197/ipsjtcva.8.11 Google Scholar
G. Panahandeh, M. Jansson and P. Händel,
“Calibration of an IMU-camera cluster using planar mirror reflection and its observability analysis,”
IEEE Trans. Instrum. Meas., 64
(1), 75
–88
(2015). https://doi.org/10.1109/TIM.2014.2329388 Google Scholar
T. P. Pachidis and J. N. Lygouras,
“Pseudostereo-vision system: a monocular stereo-vision system as a sensor for real-time robot applications,”
IEEE Trans. Instrum. Meas., 56
(6), 2547
–2560
(2007). https://doi.org/10.1109/TIM.2007.908231 Google Scholar
Y. Xu et al.,
“A calibration method for non-overlapping cameras based on mirrored absolute phase target,”
Int. J. Adv. Manuf. Technol.,
(2018). https://doi.org/10.1007/s00170-018-1704-8 Google Scholar
L. Li et al.,
“Flat mirror tilt and piston measurement based on structured light reflection,”
Opt. Express, 22
(22), 27707
–27716
(2014). https://doi.org/10.1364/OE.22.027707 OPEXFF 1094-4087 Google Scholar
H. Ren, F. Gao and X. Jiang,
“Iterative optimization calibration method for stereo deflectometry,”
Opt. Express, 23
(17), 22060
–22068
(2015). https://doi.org/10.1364/OE.23.022060 OPEXFF 1094-4087 Google Scholar
J. Y. Bouguet,
“Camera calibration toolbox for Matlab,”
(2013) http://www.vision.caltech.edu/bouguetj Google Scholar
T. Matsuyama, S. Nobuhara and K. Takahashi,
“A new mirror-based extrinsic camera calibration using an orthogonality constraint,”
in IEEE Conf. on Computer Vision and Pattern Recognition,
1051
–1058
(2012). Google Scholar
X. Li et al.,
“Accurate mirror-based camera pose estimation with explicit geometric meanings,”
Sci. China-Technol. Sci., 57
(12), 2504
–2513
(2014). https://doi.org/10.1007/s11431-014-5713-z Google Scholar
M. Shah,
“Comparing two sets of corresponding six degree of freedom data,”
Comput. Vis. Image Understanding, 115
(10), 1355
–1362
(2011). https://doi.org/10.1016/j.cviu.2011.05.007 CVIUF4 1077-3142 Google Scholar
Z. Liu et al.,
“Novel calibration method for non-overlapping multiple vision sensors based on 1D target,”
Opt. Lasers Eng., 49
(4), 570
–577
(2011). https://doi.org/10.1016/j.optlaseng.2010.11.002 Google Scholar
Z. Liu, F. J. Li and G. J. Zhang,
“An external parameter calibration method for multiple cameras based on laser rangefinder,”
Measurement, 47 954
–962
(2014). https://doi.org/10.1016/j.measurement.2013.10.029 0263-2241 Google Scholar
J. J. Moré,
“The Levenberg–Marquardt algorithm: implementation and theory,”
Numerical Analysis, 105
–116 Springer(1978). Google Scholar
F. Zhang,
“Matrix analysis and applications,”
Int. J. Inf. Syst. Sci., 309
(1), i
(2004). Google Scholar
K. Kanatani,
“Analysis of 3-D rotation fitting,”
IEEE Trans. Pattern Anal. Mach. Intell., 16
(5), 543
–549
(1994). https://doi.org/10.1109/34.291441 ITPIDJ 0162-8828 Google Scholar
Z. Y. Zhang,
“A flexible new technique for camera calibration,”
IEEE Trans. Pattern Anal. Mach. Intell., 22
(11), 1330
–1334
(2000). https://doi.org/10.1109/34.888718 ITPIDJ 0162-8828 Google Scholar
Y. Bok et al.,
“Extrinsic calibration of non-overlapping camera-laser system using structured environment,”
in IEEE/RSJ Int. Conf. on Intelligent Robots & Systems,
436
–443
(2014). Google Scholar
BiographyWenlei Liu received his BS degree at Shandong University of Science and Technology, Qingdao, in 2012 and his MS degree at Beihang University, Beijing, in 2016. Currently, he is pursuing his PhD at Beihang University, Beijing, China. He mainly studies the high-precision visual relative navigation, and also studies the data fusion of multi data sensors to achieve the purpose of cooperative detection and guidance. Sentang Wu received his PhD in dynamics, ballistics, and aircraft motion control systems from National Aviation University, Ukraine, in 1992. Currently, he is a professor of automation science and electrical engineering, and a PhD tutor at Beihang University, Beijing, China. He is also the navy missile expert in the National Defense Basic Research Institute and a member of the academic committee. His research interests include the theory and application of nonlinear stochastic systems, computer information processing and control, aircraft cooperative control, precision, and guidance. Xiaolong Wu received his BS degree in Sichuan University, Chengdu, in 2007 and his PhD in Beihang University, Beijing, in 2017. He mainly studies the information acquisition system of the missile’s autonomous formation, including the relative navigation method of the missile's autonomous formation, the calibration of the vision sensor’s installation parameters and the optimization of the field of view, and the cooperative detection and guidance of the missile's autonomous formation. |