You are on page 1of 15

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL.

70, 2021 5016715

A Convenient and High-Accuracy Multicamera


Calibration Method Based on Imperfect
Spherical Objects
Liming Tao , Renbo Xia , Jibin Zhao , Tao Zhang , Yueling Chen , and Shengpeng Fu

Abstract— Multicamera systems have important applications systems are the most widely used systems in the industry due
in industrial online measurement, attracting wide interest due to their simple structure and well-developed theories. However,
to their encouraging performance. However, how to develop a their simple structure limits their applications in complex and
convenient and high-accuracy multicamera calibration model is
a big challenge. Traditional calibration methods based on 2-D large-scale parts inspection [5]–[7]. It is a good choice to
objects require the preparation of high-precision and large-scale use multicamera systems to solve this problem. For example,
calibration objects and have poor adaptability to the spatial in the literature [8], multiple static and locally overlapping
distribution of cameras. To address these issues, a convenient cameras are used to form a multicamera system to detect
and accurate multicamera calibration method is proposed based various geometric structures on automobile parts. The mea-
on an imperfect spherical object. Specifically, a special calibration
object, i.e., an imperfect sphere with many coded targets is surement accuracy of multicamera systems depends greatly on
designed for calibrating the multicamera system, where the the calibration accuracy [9]–[11]. Therefore, how to develop
optical axes of the cameras are different in orientations and a high-efficiency and high-accuracy multicamera calibration
converge into the measurement field. Then, a Euclidean recon- method is of great significance for industrial measurement,
struction method is employed to calculate the camera poses and especially in the scenarios for complex and large-scale parts
the spatial coordinates of feature points in the local coordinate
system after the adaptive grouping of cameras is completed. inspection.
Moreover, a graph theory-based optimal path transformation Multicamera calibration methods can be divided into two
algorithm is developed to obtain the camera poses and spatial categories: the methods with physical calibration objects and
point coordinates in the global coordinate system (GCS), and the methods with virtual calibration objects. The calibration
a spherical projection optimization algorithm is designed to methods with physical calibration objects need to prepare a
spheroidize the spatial coordinates of feature points. In the end,
the camera poses and the spatial coordinates of feature points high-precision calibration object on which many feature points
in the GCS are optimized by a bundle adjustment algorithm. are evenly distributed. According to the image coordinates and
We build a multicamera system and conduct extensive experi- the known spatial coordinates of the feature points, the exter-
ments to demonstrate the superiority of the proposed method. nal parameters of the multicamera system can be obtained
Index Terms— 3-D calibration object, bundle adjustment, by using space resection methods [12]–[15]. These methods
machine vision, multicamera calibration. require the size of the physical calibration object to be large
enough to fully fill the field-of-view (FOV) of each camera in
I. I NTRODUCTION the multicamera system. Thus, it is difficult to manufacture and
maintain this kind of calibration object, which is not suitable
M ACHINE vision system has played an increasingly
important role in industrial inspection in recent years
because it allows noncontact, real-time, and high-accuracy
for on-site applications. To solve this problem, the calibration
methods with virtual calibration objects are developed. The
virtual calibration objects are composed of many virtual spatial
measurements [1]–[4]. Monocular vision and binocular vision
feature points which are mainly generated in two ways: 1) by
Manuscript received July 13, 2021; revised August 17, 2021; accepted regularly moving a dedicated 2-D calibration board in the
September 6, 2021. Date of publication September 16, 2021; date of current measurement field and 2) by moving a 1-D reference bar in
version September 29, 2021. This work was supported by the National Nature
Science Foundation of China under Grant 52075532, Grant 91948203, and the measurement field. These methods are more suitable for
Grant 51805527. The Associate Editor coordinating the review process was on-site calibration of the camera’s external parameters, which
Anirban Mukherjee. (Corresponding authors: Renbo Xia; Jibin Zhao.) can also expand the measurement range of the systems at
Liming Tao and Tao Zhang are with the State Key Laboratory of Robotics,
Shenyang Institute of Automation, Institutes for Robotics and Intelligent the same time [16]–[20]. However, these methods suffer a
Manufacturing, Chinese Academy of Sciences, Shenyang 110016, China, and serious common problem, i.e., the placement of the calibration
also with the University of Chinese Academy of Sciences, Beijing 100049, objects or reference bars requires the operator to have certain
China (e-mail: taoliming@sia.cn; zhangtaosia@gmail.com).
Renbo Xia, Jibin Zhao, Yueling Chen, and Shengpeng Fu are with background knowledge, and the operation process is rela-
the State Key Laboratory of Robotics, Shenyang Institute of Automation, tively cumbersome. In addition, their calibration accuracy for
Institutes for Robotics and Intelligent Manufacturing, Chinese Academy of multicamera systems still needs to be continuously improved
Sciences, Shenyang 110016, China (e-mail: xiarb@sia.cn; jbzhao@sia.cn;
chenyueling@sia.cn; fushengpeng@sia.cn). in order to meet the sophisticated industrial measurement
Digital Object Identifier 10.1109/TIM.2021.3113132 applications.
1557-9662 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
5016715 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

To improve the accuracy and convenience of calibration, practical and produces good results when calibrating a single
a multicamera calibration method is proposed by moving camera, it is unreasonable to simply repeat the calibration pro-
an imperfect spherical object (a perfectly spherical object is cedure for each camera in the multicamera system. The pose
also feasible) in the measurement field. To be more specific, between the calibration object and the camera can be obtained
a calibration object is designed by sticking dot-distribution as a by-product of the calibration of the camera’s internal
coded targets on a sphere that does not need to have high parameters. Based on this information, the rigid transformation
sphericity, which could further avoid the difficulty and cost of between any pair of cameras can be determined by the plane
manufacturing the high-precision calibration object. In addi- of the arbitrarily selected calibration model. In theory, these
tion, it is also beneficial for cameras with different poses to transformations should be constant, regardless of the plane
taking photographs from any perspective by using a spherical on which they are calculated. However, this constraint is not
calibration object, which also greatly simplifies the calibra- necessarily satisfied in the presence of noise because each
tion processes. Then, an optimal path transformation (OPT) camera is independently calibrated. This inconsistency will
algorithm for estimating camera poses is developed, and a reduce the estimation accuracy of the relative poses between
spherical projection optimization (SPO) algorithm for well the cameras, which may cause a serious measurement error in
initializing the spatial coordinates of feature points in the the multicamera system. Using the known metric structure and
global coordinate system (GCS) is designed in order to achieve consistency constraints, Sturm and Triggs [25] and Sturm [26]
high-accuracy multicamera calibration. Finally, a multicamera propose a general method for the relative pose estimation
system is designed and extensive experimental results show between cameras and model planes based on factorization.
that the proposed method can meet the requirement of multi- But this method requires all cameras to be calibrated in
camera system calibration. advance. Ueshiba and Tomita [27] propose an algorithm for
The main contributions of the proposed method are sum- simultaneous estimation of relative rigid displacement between
marized as follows. cameras and the internal parameters based on the Zhang’s and
1) A calibration object that is easy to obtain and low Sturm’s algorithms. This algorithm inherits the convenience
cost are designed. Meanwhile, an OPT algorithm is and flexibility of the original method and yields a better
developed for estimating the camera poses in the GCS, result than the preceding algorithm. But it also encounters the
and an SPO algorithm for well initializing the spatial problem of simultaneous visibility. Another visibility issue is
coordinates of feature points in the GCS is designed to that the features are imaged obliquely if the angle θ between
achieve high-accuracy multicamera calibration. the normal of the 2-D calibration object and the optical axis
2) A calibration method for a multicamera system with of the camera is large, which leads to severe feature extraction
randomly distributed cameras is realized, which is very errors and calibration errors. Zhao et al. [28] design a special
beneficial for ensuring the visibility and detection accu- 2-D calibration object by replacing the corner features with
racy of feature points. No matter how the axis of the spherical features to improve the accuracy of feature extraction
camera points to the measurement field, there is always during oblique imaging, but the promotion of this calibration
a part of the feature points on the sphere that can be object is unfavorable due to the difficulty of preparation.
captured from an almost orthographic angle. In addition, when θ is greater than 90◦ , the camera cannot
3) A multicamera system is built and calibrated with the capture the features due to the occlusion of the calibration
proposed calibration method, and extensive experiments object itself. Feng et al. [29], [30] propose a new calibration
conducted on this system demonstrate the high-precision object by replacing conventional materials with transparent
parameter calibration results and convenient calibration glass. However, in addition to the trouble of oblique imaging,
process of our proposed method. this method also introduces errors caused by refraction.
The remainder of this article is organized as follows. In order to overcome the limitations of imaging 2-D calibra-
In Section II the related work is introduced, and in Section III tion objects, 3-D calibration objects with different shapes have
the proposed method is discussed in detail. In Section IV, been used to calibrate multiple cameras. Huang et al. [15]
we present the experiments to verify the superior performance propose a multicamera method based on a cube calibration
of the proposed method. Finally, we draw a conclusion in object, each side of which is like a 2-D checkerboard calibra-
Section V. tion object. The features on each side of the cube are measured
in advance and unified to the GCS. Therefore, the pose
transformation matrix between the cameras can be calculated
II. R ELATED W ORK
by taking one image with each camera. But we only give
A. Calibration Methods With Physical Calibration Objects the calibration results of two cameras in the experiment and
The physical calibration objects include the 2-D calibration claim that the error transfer should be considered if the method
objects and the 3-D calibration objects. The 2-D calibration is extended to calibrate multiple cameras. Zhang et al. [14]
objects, such as checkerboard, have been widely used in the design a 3-D calibration object with five planes, and each
field of computer vision due to their easy manufacturing plane contains many circular markers. This method can realize
and convenient operation [21]–[24]. The camera calibration simultaneous calibration of intrinsic and extrinsic parameters
method based on a 2-D calibration object is pioneered by of multiple cameras but requires all cameras to see part of the
Zhang [12] and is originally mainly used to obtain the internal calibration object at the same time. Therefore, it is not suitable
parameters of the camera. Although this method is very for calibrating systems where some cameras do not have a

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
TAO et al.: CONVENIENT AND HIGH-ACCURACY MULTICAMERA CALIBRATION METHOD 5016715

sufficiently large public FOV. In order to calibrate the circular


camera array, Abedi et al. [13] design a pyramid-shaped
3-D calibration object with a triangular pattern. The pyramid
peak is visible to all cameras and is considered an important
constraint for calibration. However, the accurate extraction of
the towering peak in the image is not an easy task. In addition,
this method is not suitable for calibrating multicamera systems
in the spherical arrangement. In a word, the traditional 3-D
calibration object is usually used in a specific multicamera
system, and all the feature positions need to be accurately
measured in advance. Therefore, its versatility is poor and the
cost is high.

B. Calibration Methods With Virtual Calibration Objects


The method based on the virtual calibration object is ini-
tially applied to off-line photogrammetry [31], [32]. Due to
the convenience of the method and the structural similarity
between the multicamera system (also known as the online
photogrammetric system) and the offline photogrammetric Fig. 1. Schematic of calibrating a multicamera system with cameras randomly
system, the method is also applied to the multicamera systems. distributed in space by our proposed method.
For offline photogrammetry, the images of the measured
object are asynchronously collected by a camera at different object and develop a new theoretical model for multicamera
positions, which requires that we should calibrate the intrinsic calibration. In the future, we would introduce multiple sensing
parameters of the camera and the extrinsic parameters at each modalities, such as tactile sensing, to further improve the
position. However, for online photogrammetry, the images of performance of the model [34]–[36].
the measured object are simultaneously collected by multiple
cameras at different positions, which requires that we should III. P ROPOSED M ETHOD
calibrate the intrinsic and extrinsic parameters of each camera.
In other words, the online photogrammetry system needs to A. Problem Statement
calibrate more parameters in order to achieve the purpose of Without loss of generality, the schematic of multicamera
measurement. In addition, the position, posture, and a number calibration is shown in Fig. 1. Let Ow X w Yw Z w denotes the
of cameras cannot be freely arranged, which cannot guarantee GCS with the coordinate origin Ow located on the ground
that the virtual calibration object has a good distribution and plane or at the optical center of the reference camera. The pro-
acquisition angle in the images. Therefore, the calibration posed calibration object with Ni features, whose coordinates
of a multicamera system is more challenging than offline are represented as a set {X i(t) |i = 1, 2, . . . , Ni }, is moved Nt
photogrammetry. times, where the positions are represented as set {P S (t) |t =
In the latest research based on virtual calibration objects, 1, 2, . . . , Nt }, in the measurement volume. The images of the
the calibration model of offline photogrammetry has been calibration object are captured by N j cameras denoted as
widely adopted. Svoboda et al. [19] use the virtual calibration {C j | j = 1, 2, . . . , N j }. The subscript i and j denote the index
object generated by the laser pointer to efficiently calibrate of the feature and the camera, respectively. The superscript t
the multicamera system. However, this method requires an denotes the index of the position of the calibration object. Each
environment with high contrast with the laser point and is camera has a coordinate system O j X j Y j Z j , called camera
easily affected by the reflective plane in the camera room. coordinate system, with the coordinate origin O j located at the
Yu et al. [17] realize the calibration of the multicamera system optical center of this camera. x i,(t)j denotes the projection pixel
based on the virtual calibration object formed by moving an point of the feature X i(t) on the image plane of the j th camera.
optical reference bar with three marks in the measurement When the positions of the feature X i(t) are fixed, the factors
volume. However, this method is inefficient, since it requires a that affect the position of the projection point x i,(t)j , including
lot of manual participation. Sun et al. [20] use a similar method extrinsic parameters and intrinsic parameters. For convenience,
to achieve the calibration of a multicamera system based on let X and X denote the vector and homogeneous vector of
the virtual calibration object formed by moving a rigid optical feature point X, respectively.
reference bar with two marks in the measurement volume. The mapping from a spatial point X i(t) to a 2-D pixel point
Truong et al. [18] and Aslan et al. [33] achieve multicamera (t)
x i, j can be expressed as
calibration by extracting the features of the human body  
moving in the measurement volume. However, these methods xi,(t)j = G r j , t j , ℘1 j , ℘2 j , Xi(t) (1)
have large calibration errors and cannot be used for calibration
in industrial measurement. The proposed method in this article where r j and t j are the rotation and translation transfor-
can be classified as a method based on virtual feature points; mation matrices of the j th camera, respectively. ℘1 j =
but unlike previous methods, we design an imperfect spherical { f u j , fv j , u 0 j , v 0 j } is called the intrinsic parameters which

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
5016715 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

Fig. 2. Flowchart of the proposed method. First, a decoding and feature matching algorithm is employed to obtain the homologous feature points after the
adaptive grouping of cameras is completed. Then, the Euclidean reconstruction method is conducted to calculate the camera poses and spatial coordinates of
feature points in the local coordinate system. Moreover, the SPO algorithm and the OPT algorithm are adopted for well initializing the camera poses and the
spatial coordinates of feature points in the GCS. Finally, the bundle adjustment algorithm is employed to obtain high-accuracy camera parameters.

describes the mapping from the spatial point X i(t) to the Assumption 3: The manufacturing position error of the
point x i,(t)j on the image plane of the j th camera. ℘2 j = points on coded targets is less than ±0.1 mm.
{k1 j , k2 j , k3 j , s1 j , s2 j } is the distortion coefficient of the j th
camera. The goal of this article is to obtain the optimal internal B. Methodology
and external parameters of all cameras by minimizing the The entire flow of our proposed method is shown in Fig. 2.
following objective function: Different from the existing calibration algorithms, we develop

Nt   
N j Ni
 
 2

the OPT algorithm and SPO algorithm, which reduce the
arg min λxi,(t)j − G r j , t j , ℘1 j , ℘2 j , Xi(t)  (2) possibility of global nonlinear optimization getting stuck in

t=1 j =1 i=1 a local minimum and improve the calibration accuracy.
 1) Calibration Object Design: A calibration object is
where  = {r j , t j , ℘1 j , ℘2 j , Xi(t) }, and xi,(t)j is the actual designed by pasting the dot-distribution coded targets onto
projection of the spatial point X i(t) . λ is a binary variable. the sphere, as given in Fig. 3. More specifically, the diameter
When X i(t) is visible to a camera, λ takes 1, otherwise, of the sphere is about 500 mm, and there are approximately
takes 0. Nt represents the total number of positions, where 200 coded targets on its surface, each of which has a size of
the spherical calibration object is placed. N j represents the 28 mm × 28 mm. Each dot-distribution coded target contains
total number of cameras, making up the measurement system. eight subtargets, which can further provide eight features,
Ni represents the total number of all feature points on the namely, {A, B, C, D, E, M1, M2, M3}. A, B, C, D, and
spherical calibration object. E are template points, and M1, M2, and M3 are coding
It is worth noticing that there are some common necessary points. The diameter of each feature point is about 3 mm.
assumptions needed to be claimed in this article. The reasons why we design the spherical calibration object
Assumption 1: All camera intrinsic parameters remain stable are as follows: 1) spherical objects are easy to obtain, and
in the calibration process. the globe for teaching is sufficient to meet the requirements;
Assumption 2: Each camera has a common FOV with at 2) it is helpful to improve the detection accuracy of feature
least another two cameras. points because some feature points on the sphere are always

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
TAO et al.: CONVENIENT AND HIGH-ACCURACY MULTICAMERA CALIBRATION METHOD 5016715

where F j,g and e j,g are the fundamental matrix and epipole
between the two cameras, respectively. Without loss of gen-
erality, assuming that the projective depth of X i in the j th
camera is 1, the projection depth of the point under the
other two cameras can be obtained according to (4). Then,
the projection equations of n features in the common FOV of
three cameras can be formulated as
⎡ ⎤ ⎡ ⎤
μ1,1 x1,1 · · · μn,1 xn,1 Qp1  
Fig. 3. Designed calibration object and its related components. (a) Four dif- ⎣ μ1,2 x12 · · · μn,2 xn,2 ⎦ = ⎣ Qp2 ⎦ Xp1 · · · Xpn . (5)
ferent dot-distribution coded targets. (b) Sphere without special requirements μ1,3 x1,3 · · · μn,3 xn,3 Qp3
for sphericity. (c) Designed calibration object.
Using the SVD-based method [16], [23] to factorize the
matrix on the left side of (5), whose rank does not exceed 4
orthophotographed; and 3) it can adapt to various arrangements at most, the 3-D coordinates [ Xp1 · · · Xpn ] and camera poses
of cameras. [ Qp1 Qp2 Qp3 ]T in the projective space can be recovered.
2) Estimation of the Camera Poses and the Spatial Coor- In order to further obtain the corresponding 3-D coordinates
dinates of the Feature Points in Local Coordinate System: and camera poses in the Euclidean space, the transformation
Notice that we do not focus on feature matching, since the matrix H from projective space to Euclidean space is intro-
use of coded targets makes it very easy and reliable. More duced, which should satisfy the following constraints:
details of feature matching are given in the literature [37]. ⎧  
Once all cameras are installed on the rigid framework, their ⎪
⎪ Q H = α K r t

⎪ 1 1 1 1 1
internal and external parameters are considered unchanged ⎨  
during the calibration process. Therefore, the feature points Q2 H = α2 K2 r2 t2 (6)

⎪  
in two photographs taken by the same camera at different ⎪

⎩Q3 H = α3 K3 r3 t3
positions of the calibration object can be merged into the same
image plane. In other words, this approach is equivalent to where αi (i = 1, 2, 3) is a nonzero real number. Solving (5)
achieving the same effect as the use of a large-scale calibration and (6) together, H can be solved. QH and H−1 X correspond
object if the distance between the coded targets does not need to the pose of the camera and the coordinate of feature point in
to be known during the calibration process. Thus, we no longer the Euclidean space, respectively. Therefore, the transforma-
specifically declare the position of the calibrated object where tion relationship from projective reconstruction to Euclidean
the spatial feature points are located. We use the symbol X i reconstruction can be expressed as follows:
to replace X i(t) and x i, j to replace x i,(t)j in the following text. ⎧   T
In a multicamera system, not all cameras have a common ⎪
⎨ Q1 Q2 Q3 = Qp1 Qp2 Qp3 H,
FOV. Therefore, the cameras must be automatically grouped     (7)

⎩ X1 · · · Xn = H−1 Xp1 · · · Xpn .
according to the number of common feature points in the
FOV. Each group contains k cameras, and each camera can
Let Qg and Q j denote the pose matrices of the gth and j th
participate in multiple groups. In this article, k takes the
cameras in the local coordinate system, respectively. Then,
value 3. The cameras in each group are calibrated in order to
the relative pose transformation between the gth and j th
obtain the poses of the cameras in the local coordinate system,
cameras can be expressed as
which will be converted to the GCS. More specifically, we first
group any three cameras to get all possible combinations. Q j,g = Qg · Q−1
j . (8)
Then, we calculate the number of common features in each
group of cameras, and those groups whose number of common Accordingly, the coordinates of the i th point in j th camera
features is lower than the threshold THpt are discarded. Here, coordinate system can be obtained as follows:
we take one group as an example to introduce the calculation Xi, j = Q−1
j · Xi . (9)
method of camera pose and spatial point coordinates.
According to the mathematical model of pinhole camera 3) Initialization of the Camera Poses and the Spatial Coor-
perspective imaging, the mapping relationship between a spa- dinates of the Feature Points in the GCS: We choose the
tial point X i and an image point x i, j can be expressed as camera, which is shared by most groups as a reference camera,
  and define the coordinate system of the reference camera as
μi, j xi, j = K j r j t j Xi (3) the GCS. After the camera poses and the spatial coordinates
where K j ( j = 1, 2, 3) is the intrinsic matrix of j th camera, of the feature points in the local coordinate system of each
and μi, j is the unknown scale factor which represents the group are obtained, the next step is to transform the local
projective depth from the spatial point X i to the j th camera. coordinate systems into the GCS to perform the multicamera
The relationship between the scale factors of point X i in the global calibration. However, there are many possible trans-
j th camera and gth camera can be defined as formation paths. And different paths will lead to different
 T global calibration accuracy. Thus, it is necessary to develop an
e j,g × xi, j F j,g · xi,g OPT algorithm to improve the estimation accuracy of camera
μi, j = μi,g (4)
e j,g × xi, j 2 poses and the spatial coordinates of the feature points in

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
5016715 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

the GCS. In addition, an SPO algorithm is developed to better


initialize the spatial coordinates of the feature points in the
GCS based on a spheroidization operation. More details of the
OPT algorithm and the SPO algorithm are given as follows.
Optimal Path Transformation Algorithm: Graph theory is
used to clearly describe the transformation path optimization
problem. In this article, the camera is defined as the vertex of
the graph, and the pose transformation matrix between the
cameras is defined as the edge. The weight of each edge
depends on the reprojection error and the number of com-
mon features between cameras. In general, the weight of the
edge represents the distance between two vertices. However,
the purpose of this article is to achieve the highest accuracy Fig. 4. Two-dimensional adjacency matrix. The node Si, j stores all useful
camera pose in GCS. Therefore, the weight of the edge should information between Ci and C j , including transformation matrix, spatial
points, and edge weight.
reflect the pose accuracy between the two cameras. Inspired by
the A star search algorithm [38], a calculation algorithm for
the OPT is developed based on the graph. A heuristic function After obtaining the weights of the edges of the graph
is designed to improve the efficiency of finding the optimal through (13), the next step is to design an algorithm to find the
path by predicting the shortest path from the current vertex to optimal transformation path. To describe the algorithm clearly,
the target vertex. we let S j,g = {Q j,g ,  j,g , a j,g } represents the information of
The first step of the algorithm is to determine the weight each edge in the graph, where  j,g = {Xi, j |i = 1, 2, . . . , N j,g }
{a j,g | j = 1, 2, . . . , N j , g = 1, 2, . . . , N j } of each edge of the is the set of feature points in the common FOV of the j th
graph. The reprojection error is used to evaluate the recon- camera and the gth camera that is calculated by (7)–(9). It is
struction accuracy of the vision system, which also numeri- easy to see that Q j,g = Q−1 g, j and a j,g = ag, j according
cally reflects the correctness of the camera pose. The mean of to the properties of the transformation matrix and (11)–(13).
the reprojection error of the j th camera is defined as In addition, we define a 2-D adjacency matrix M = {S j,g | j =
1, 2, . . . , N j , g = 1, 2, . . . , N j }, as shown in Fig. 4, to facili-
1   
n
E Rj =  x  − G r j , t j , ℘ 1 j , ℘ 2 j , Xi  (10) tate the OTP algorithm to find the optimal transformation path.
i, j
n i=1 We first apply the method described in Section III-B.2 to
where n denotes the number of feature points that success- complete the calculation of the camera pose and spatial point
fully reconstructed with a group of cameras. Then, the first coordinates in the local coordinate system. Then, the relevant
component of weight in the edge can be defined as elements in the adjacency matrix are filled. If an element in
⎧ ⎧ the adjacency matrix is recalculated by another group cameras,

⎪ ⎪
⎨ E R j ∈ (0, T Her ]& the one with a smaller edge weight is selected as this element.

⎨ E R j + E Rg , if E R ∈ (0, T H ]& Here, we regard the adjacency matrix as a map and the
a (1) ⎪ j er
j,g = 2 ⎩ (11)

⎪ C , C ∈ Same Group. elements in the adjacency matrix as map nodes. We define


j g
the camera to be transformed as the starting node on the
+∞, otherwise
map and the reference camera as the target node. The OPT
where E R j and E Rg denote the mean of the reprojection algorithm not only considers the edge weight a j,g but also
error of the j th and gth camera, respectively. THer is the considers the number of nodes that pass through during the
reprojection error threshold. pose transformation. Thus, the estimated shortest path length
As a rule of thumb, the larger the number of common f ( p) from the starting node to the target node through pth
feature points in the FOV of two cameras, the more accurate node is defined as
the camera pose estimation can be obtained. Therefore, we use
the reciprocal of the number of common feature points to f ( p) = bm g( p) + h( p) (14)
establish the second component of the weight of the edge, where p denotes the next node on the map to be searched. m
which can be expressed as denotes the number of nodes passed from the starting node to
1 the pth node during the pose transformation. b is a constant,
a (2)
j,g = (12) and b ≥1. b =1 means that the influence of the number
N j,g
of transformations is not considered. g( p) denotes the actual
where N j,g denotes the number of common feature points of length of the path from starting node to pth node. h( p) is
the j th and gth cameras. a heuristic function that denotes the estimated shortest path
Using (11) and (12) jointly, the weight of edges in the graph from the pth node to the target node. Assuming that Nmin is
can be defined as the minimum number of nodes passed from the pth node to
a j,g = ξ1 a (1) (2)
j,g + ξ2 a j,g (13) the target node, then h( p) can be defined as
h( p) = Nmin × amin (15)
where ξ1 and ξ2 represent the coefficients of a (1) (2)
j,g and a j,g ,
respectively. where amin = min{a j,g | j = 1, 2, . . . , N j , g = 1, 2, . . . , N j }.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
TAO et al.: CONVENIENT AND HIGH-ACCURACY MULTICAMERA CALIBRATION METHOD 5016715

Fig. 5. Example of finding the optimal transformation path in a four-camera system. (a) All feasible transformation paths. The blue arrow indicates the
optional path, and the red arrow indicates the optimal path. (b) Process of finding the optimal transformation path. The part marked in red represents the
minimum value of the current search node. The equations in the box filled with light cyan represent the calculation process of the next nodes to be searched.

In order to find the optimal transformation path, we first Algorithm 1 OPT Algorithm
calculate the f ( p) of each node according to (14) and regard Input:
the node with the smallest f ( p) as the current node. Then, Ki : intrinsic matrices;
calculate the f ( p) of the next node (ignoring the nodes that the xi, j : image point coordinates;
current path has passed), and regard the node with the smallest T H pt , T H er , ξ1 , ξ2 , b: thresholds and coefficients.
f ( p) as the current node again. By analogy, the search ends Output:
until the target node becomes the current node. An example Camera poses and spatial point positions in the GCS.
of finding the optimal transformation path in a four-camera 1: Initialize the adjacency matrix M with S j,g =
system is shown in Fig. 5, where b = 1.1. {i denti t y matri x, empt y set, i n f i ni t y};
Suppose the transformation path of one camera 2: Group any three cameras to get all possible combinations
obtained by the above-mentioned method is Path = CombSet = {Cm Setυ |υ = 1, 2, · · · , Ncs };
{Ss,s , Ss,1 , S1,2 , . . . , Sn,r , Sr,r }, where Ss,s and Sr,r are the 3: for υ ← 1 to Ncs do
starting node and target node, respectively. Then, the pose of 4: Match the features and count the number of features
the camera in the GCS can be expressed as successfully matched N j,g ;
Qs,r = Qs,1 Q1,2 , . . . , Qn,r . (16) 5: Calculate the relative poses Q j,g and spatial points P j,g
of Cm Setυ by Eq. (3) to (9);
Similarly, the corresponding spatial points in the GCS can 6: Calculate the edge weight a j,g based on xi, j , K j , N j,g
be expressed as by Eq. (10) to (13);
−1 −1 7: if a j,g < the corresponding element in M then
Xi,r = Q−1
n,r , . . . , Q1,2 Qs,1 Xi,s . (17)
8: Replace corresponding element in M with Q j,g , P j,g ,
We summarized the proposed OPT algorithm in a j,g ;
Algorithm 1. 9: end if
The OPT algorithm is a new proposed algorithm considering 10: end for
the factors which affect the camera calibration accuracy. It is 11: Find the minimum value of a j,g in M;
suitable for multicamera calibration where local calibration 12: Calculate Nmin by the A star search algorithm [38];
is performed first and then a global calibration. Dijkstra’s 13: for j ← 1 to N j do
algorithm is the most similar to the proposed algorithm, 14: Calculate the optimal path Path j based on M by
which is another approach that could be used. But Dijkstra’s Eq. (14) to (15);
algorithm has the problem of lower search efficiency. Our 15: Calculate camera poses and point positions in the GCS
algorithm only calculates the length of a part of paths, which by Eq. (16) to (17);
has the time complexity of O(n log(n)), and n is the number 16: end for
of nodes. While Dijkstra’s algorithm needs to calculate the
length of all paths, which has the time complexity of O(n 2 ),
n is the number of nodes [39]. by this algorithm are composed of feature points on multiple
Spherical Projection Optimization Algorithm: We have spherical calibration objects which actually correspond to the
mentioned that the initial value has a great influence on the same spherical calibration object placed in different positions.
BA optimization. And we have obtained a good initial value Therefore, in order to implement the algorithm, the spatial
of the camera pose through the OPT algorithm. Next, we will points need to be divided into multiple subsets according to the
develop an SPO algorithm to obtain a better initial value of sphere to which they belong. Next, sphere fitting is performed
the spatial point coordinates. Because the spatial points used on the spatial points in each subset to obtain the parameters of

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
5016715 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

Algorithm 2 SPO Algorithm where T j,r and Xi,r denote the camera translation vectors and
Input: spatial points before scale transformation, respectively. T j and
Xi,r : The spatial points obtained by algorithm 1. Xi denote the camera translation vectors and spatial points
Output: after scale transformation, respectively. It is worth noticing that
Optimized spatial points. if the position accuracy of the features of the dot-distribution
1: Divide Xi,r into Nt groups {SpSett |t = 1, 2, · · · , Nt }, coded targets is not high enough, the global scale can still
(t) (t) be obtained by using the reference bar reconstruction method,
where SpSett = {Xi,r |i = 1, 2, . . . , Ni,r };
2: for t ← 1 to Nt do as described in Fig. 2.
3: Fit the sphere with SpSett to get the sphere parameters;
(t) IV. E XPERIMENTAL R ESULTS
4: for i ← 1 to Ni,r do
5: Calculate the projection point of the spatial point on A. Experiment Settings
the sphere; To verify the performance of the proposed method, we build
6: Save the projection point. a multicamera measurement system with ten cameras mounted
7: end for on a rigid frame, as shown in Fig. 6(a). The measurement
8: end for system is a rectangular parallelepiped structure with a size
of 1500 mm × 1000 mm × 1200 mm, which consists of ten
AVT G-319B cameras distributed in two rows on the top plane.
the sphere. Then, all the points are projected to the spherical To improve the lighting conditions of the system, ten ring light
surface along the radius direction. The obtained projection sources are coaxially distributed with the camera. In addition,
points are considered as candidates for better initial values of as shown in Fig. 6(b), a clearer photograph of the actual globe
spatial points. We summarized the proposed SPO algorithm in is provided, which enlarges a part of the globe to show the
Algorithm 2. stretch of the dot-distribution coded targets more clearly.
4) Global Optimization and Scale Transformation: Four experiments, including the calibration object layout
Although we have obtained the camera poses and point experiments, the OPT algorithm experiments, the SPO algo-
positions by the proposed OPT algorithm and SPO algo- rithm experiments, and the comparative experiments, are con-
rithm, the obtained values do not meet the requirement of ducted to prove the effectiveness of our proposed method.
high-precision industrial measurement. In order to obtain more The reprojection error E rpj is employed to evaluate the model
accurate camera poses and point positions, we employ the performance, which is defined as follows:
bundle adjustment algorithm to optimize the reprojection error,
1  
N j Ni,r
2
which can further refine these parameters. More specifically, E rpj = x − x i, j (20)
given the spatial point coordinates in the GCS and the initial Nall j =1 i=1 i, j
poses of the cameras, we refine the parameters by optimizing
where x i, j and x i, j denote the reprojected and captured image
the reprojection error defined via (2).
point, respectively. Nall is the cardinal of these sets.
According to the calibration theory, the camera poses and
As shown in Fig. 6(c), two reference bars (i.e., Aref and
the spatial point positions obtained by global optimization are
Bref ) are used for reconstruction accuracy experiments in this
different from the actual value by a scale. Normally, this scale
article. The reference bar Aref has four markers, which are
can be obtained by calculating the ratio of the reconstructed
denoted as A1, A2, A3, and A4. The reference bar Bref
length of the reference bar and its theoretical length.
also has four markers, which are denoted as B1, B2, B3,
According to Assumption 3, the distance between the
and B4. Five subreference bars, i.e., A1A2, A3A4, B1B2,
template points of the dot-distribution coded targets on the
B1B3, and B1B4 are used in this article. The reference
calibration object is a fixed value, which also obeys the normal
length of each subbar, which is denoted as L, is obtained
distribution. Therefore, we can use the two most distant points,
using a high-accuracy three-coordinate imager and is given
i.e., point A and point E, as shown in Fig. 3(a), to calculate
in Fig. 6(c). We employ the proposed multicamera system to
the scale factor, which can be expressed as
reconstruct the reference bars and denote the reconstruction
NAE 
  length as L rcs . Then, the reconstruction error E rcs can be
1  (A) (E) 
scale = Xi,r − Xi,r  (18) defined as
d · NAE i=1
E rcs = L − L rcs . (21)
where Xi,r (A) and Xi,r (E) denote the reconstruction coordinates
of point A and point E on the i th dot-distribution coded
target, respectively. d denotes the actual distance between B. Calibration Object Layout Experiments
point A and point E. After obtaining the scale, we can calculate The layout (i.e., the number of placement times and place-
the spatial points and the camera poses in the GCS by the ment position) of calibration objects has a great influence
following equation: on the calibration result because it determines the number of
 feature points in the captured images.
T j = scale · T j,r A heuristic method is employed to find the best layout. More
(19)
Xi = scale · Xi,r specifically, as shown in Fig. 7, seven different layouts are

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
TAO et al.: CONVENIENT AND HIGH-ACCURACY MULTICAMERA CALIBRATION METHOD 5016715

Fig. 8. Results of calibration object layout experiments. (a) Reprojection


errors of the feature points under different layouts. (b) Reconstruction errors
of subreference bars under different layouts.

TABLE I
T HRESHOLD AND C OEFFICIENT VALUES U SED IN THE A LGORITHMS

error increases slightly and then gradually becomes stable


with the increasing of the number of placement times. The
maximum reprojection error is less than 0.2. These experi-
mental results show that the proposed model is convergent.
In addition, as shown in Fig. 8(b), when the number of
placement times is small, the reference bar reconstruction error
is relatively large. With the increasing of the number of place-
ment times, the reconstruction error gradually decreases and
tends to be stable at the error of 0.1 mm. These experimental
phenomena show that increasing the number of placement
times appropriately can effectively improve the calibration
performance. However, when the number of placement times
exceeds 4, its effect on the reconstruction error is limited.
Increasing the number of placement times would also increase
the complexity of the calibration process. Thus, we set the
number of placement times to be 4 in this article in order to
balance calibration accuracy and complexity.

C. OPT Algorithm Experiments


Fig. 6. Equipment used in the experiment. (a) Diagram of the multicamera To prove the effectiveness of the proposed OPT algorithm,
system with ten cameras. (b) Much clearer photograph of the calibration we conduct two calibration experiments. The first experi-
object. (c) Reference bar and its geometric information.
ment employs the proposed OPT algorithm in the calibration
process. The second experiment does not employ the OPT
algorithm, which is called the none-OPT algorithm, as shown
in Algorithm 3. There are several thresholds and coefficients
that should be set appropriately for the algorithms, which are
given in Table I.
As shown in Fig. 9, the fifth camera is set as the refer-
ence camera on which the GCS is established. The number
on each line denotes the weight of each edge, which is
Fig. 7. Schematic of seven different layouts of the calibration objects. calculated by (13). After getting the weight, the optimal
transformation path can be obtained via (14). For instance,
employed, where calibration objects are placed as evenly as as shown in Fig. 9(a), according to (14), the optimal path from
possible. “camera2” to the reference camera “camera5” is {C2 , C1 , C5 }.
Two evaluation metrics, i.e., the reprojection error and the Similarly, the shortest path for “camera10” to “camera5” is
reference bar reconstruction error are used to evaluate the cal- {C10 , C8 , C5 }. As a contrast, as shown in Fig. 9(b), we directly
ibration performance. As shown in Fig. 8(a), the reprojection employ the path from the adjacency matrix. More specifically,

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
5016715 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

Algorithm 3 None-OPT Algorithm


Input:
Ki : intrinsic matrices;
xi, j : image point coordinates;
T H pt : threshold.
Output:
Camera poses and spatial point positions in the GCS.
1: Group any three cameras that contains reference camera,
CombSet = {Cm Setυ |υ = 1, 2, · · · , Ncs };
2: for υ ← 1 to Ncs do
Fig. 10. Experimental results of the OPT algorithm experiments. (a) Repro-
3: if There are unoriented cameras in the Cm Setυ then
jection error of the features on the images captured by the ten cameras.
4: Match the features and count the number of features (b) Reconstruction error of five subreference bars.
successfully matched, N j,g ;
5: Calculate the poses Q j,r and 3D points P j,r of the
Cm Setυ by Eq. (3) to (9);
6: end if
7: end for
8: while There are unoriented cameras do
9: Find oriented cameras, {Cso |so = 1, 2, · · · , Nso };
10: for so ← 1 to Nso do
11: Set Cso as reference camera, and repeat Step 1 to 7;
12: Calculate camera poses and point positions in the GCS
by Eq. (16) to (17);
13: end for
14: end while

Fig. 11. Schematic of the effect of the SPO algorithm. (a) Spatial points
before performing the SPO algorithm and a curved surface fit with them.
(b) Spatial points after performing the SPO algorithm and a curved surface
Fig. 9. Transformation path of cameras. The black solid dots represent the fit with them. The different colors indicate different z-axis coordinate values.
camera, and the solid lines and the dotted lines represent the used and unused All spatial points are obtained from real experiments.
transformation paths, respectively. (a) Path obtained by the proposed OPT
algorithm. (b) Path obtained by the none OPT algorithm. 0.11 mm, which is also much smaller than that obtained by
the none-OPT algorithm.
taking the path of “camera2” to “camera5” for example,
the path length is 0.1221 which is obtained by the proposed D. SPO Algorithm Experiments
OPT algorithm, and the path length obtained by the none-OPT As shown in Fig. 11(a), we visualize some of the spatial
algorithm is 0.188. This experiment shows the proposed OPT points before performing the SPO algorithm and fit a curved
algorithm can actually help obtain the shortest transformation surface with them to better illustrate the distribution of spatial
path. points. Obviously, the spatial points are not on the same
In order to further prove the effectiveness of the proposed sphere. Therefore, we fit the spherical surface and project
OPT algorithm, we also use the reprojection error and subref- the spatial points to the spherical surface along the radius
erence bar reconstruction error to evaluate the performance. direction. The result is shown in Fig. 11(b). Theoretically,
As shown in Fig. 10(a), the average value of the reprojec- it is reasonable to regard the obtained projection points as
tion error obtained by the proposed OPT algorithm is about candidates for better initial values of spatial points.
0.09 pixels, which is much smaller than that obtained by the Furthermore, we conduct two experiments to prove the
none-OPT algorithm (0.15 pixels). As shown in Fig. 10(b), effectiveness of the proposed SPO algorithm. The first experi-
the average value of the reconstruction error of the subrefer- ment employs the SPO algorithm. The second experiment does
ence bar obtained by the proposed OPT algorithm is about not employ the SPO algorithm, i.e., the spatial points obtained

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
TAO et al.: CONVENIENT AND HIGH-ACCURACY MULTICAMERA CALIBRATION METHOD 5016715

TABLE II
R EPROJECTION E RROR OF THE SPO A LGORITHM AND THE N ONE -SPO A LGORITHM

by Algorithm 1 or Algorithm 3 are directly used as the initial


value of BA optimization, called the none-SPO algorithm.
The inaccurate intrinsic parameters and feature extraction
have the most serious impact on the spatial point. Therefore,
to prove the superiority of the SPO method more objectively,
we add different degrees of Gaussian noise (10%–25%) to
the spatial point to simulate the spatial point error caused by
the inaccurate intrinsic parameters and the inaccurate feature Fig. 12. Locations of the used cameras and the layouts of the calibration
extraction. The process of using the noise is as follows. object.

1) Generate Gaussian random numbers, Rx,i , R y,i , and


Rz,i , between −1 and 1.
2) Add a%(a = 10, 15, 20, 25) Gaussian noise to the
spatial point coordinates


⎨ X i,R = X i,r + a% × R × Rx,i
Yi,R = Yi,r + a% × R × R y,i (22)


Z i,R = Z i,r + a% × R × Rz,i
where R is a constant, and R = 250 mm, which
is approximately equal to the radius of the calibration Fig. 13. Results of experiments with varying number of cameras. (a) Average
object; X i,R , Yi,R , and Z i,R are spatial point coordinate values of the reprojection error. (b) Standard deviation error bars of the
reconstruction of the reference bar.
components with Gaussian noise.
By the way, the main noise present in our industrial setting The reprojection error and the reference bar reconstruc-
is vibration, but this is different from the noise mentioned in tion error are used to evaluate the calibration performance.
this article. As shown in Fig. 13(a), the largest reprojection error is
Table II lists the reprojection errors of ten cameras under 0.875 pixels, the smallest reprojection error is 0.0593 pixels.
different noise data. The following conclusions can be drawn: The standard deviation error bars of the reconstruction of
1) SPO always performs best under all perturbations; 2) when reference bar, A3A4, are shown in Fig. 13(b), which is
the perturbation is small (10% and 15%), SPO slightly per- obtained by 20 measurements. It can be seen that the mean
forms better than none-SPO; and 3) with the increasing noise, values of reconstruction errors are less than 0.05 mm, and
SPO performs much better than none-SPO. For instance, the standard deviations of reconstruction errors are less than
when the noise is 20% or 25%, the none-SPO even becomes 0.1 mm. Therefore, we can conclude that our proposed scheme
nonconvergent. These encouraging results show that the pro- performs well for calibrating multicamera systems with differ-
posed SPO algorithm can actually improve the calibration ent numbers of cameras.
accuracy.
F. Comparative Experiments
E. Experiments With Varying Number of Cameras We compare the proposed method with the classical cali-
Experiments were performed with different numbers of bration method and Deng et al.’s [16] method to verify the
cameras on the system described in Section IV-A. The loca- performance of the proposed method. For the experiment of
tions of the used cameras and the layouts of the calibration the classical calibration method, the 2-D calibration object
object are shown in Fig. 12. (Fig. 14) that used in the experiment is carefully manufactured

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
5016715 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

TABLE III
M EASUREMENT U NCERTAINTY OF S UBREFERENCE BARS

Fig. 14. Clearer image of the 2-D calibration object. Please check our
previous work [24] for more details about this object. Fig. 15. Some calibration images of the experiments collected by the cameras.
(a) Some images of the proposed method. (b) Some images of the classical
calibration method. (c) Some images of Deng’s method.
using high-precision manufacturing equipment, and the dis-
tance between adjacent feature points is 33 ± 0.01 mm [24].
It is worth noticing that we calibrate four adjacent cameras how the reference bars are placed is shown in Fig. 18. The
each time. For instance, we first calibrate {C1, C2, C3, C4} results also perform statistical analysis for the evaluation of
and {C3, C4, C5, C6}, and then use the public cameras the measurement uncertainty of the data with the standard
(C3 and C4) to transform all cameras to the GCS. For international methods [40]. The measurement uncertainties
the experiment of Deng’s method, the reference bar, A3A4, are shown in Table III. It can be seen that the average
in Fig. 6(c) is used for the calibration object. In this experi- reconstruction errors of the proposed method are slightly
ment, 3000 frames of images of the reference bar are collected. smaller than the result obtained by the other two methods.
Some calibration images of the experiments collected by the The standard deviations of our proposed method are less
cameras are shown in Fig. 15. than 0.125 mm, which are much smaller than the result
As shown in Fig. 16, the average absolute error of the pro- obtained by the other two methods, 0.2 mm and 0.175 mm,
posed method in the x-axis direction and the y-axis direction respectively. The measurement uncertainties of our method
are 0.045262 and 0.048272 pixels, respectively, which are are also significantly smaller than the other two methods.
smaller than the result obtained by the other two methods. The Therefore, we can conclude that our method is more robust
standard deviations of our proposed method in the x-axis and and could achieve better calibration performance.
y-axis directions are 0.06323 and 0.06287, respectively, which The time cost is a crucial factor in practical applications,
are much smaller than the result obtained by the other two and thus, we further analyze the time cost of the three meth-
methods. Therefore, we can conclude that our method is more ods. We implement our proposed algorithm using the C++
robust and could achieve better calibration performance, which language and perform the experiment on a personal computer
also can be displayed more intuitively from error distribution with the capability of 2.60-GHz CPU and 16.0-GB memory.
in Fig. 16. As shown in Fig. 19(a), we obtain the result by repeating
The standard deviation error bars of the subreference the experiment ten times, and the average time consumption
bars reconstruction with three calibration methods are shown is 168.71 s for our method, 237.89 s for classical method,
in Fig. 17, which is obtained by 20 measurements and gives 319.20 s for Deng et al.’s [16] method. Therefore, we can
the stability of the measured distances. The schematic of conclude that we are outstanding in time consumption.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
TAO et al.: CONVENIENT AND HIGH-ACCURACY MULTICAMERA CALIBRATION METHOD 5016715

Fig. 16. Reprojection error of all spatial points successfully obtained. (a) Result of our proposed method. (b) Result of the 2-D classical calibration method.
(c) Result of Deng et al.’s [16] method.

Fig. 17. Standard deviation error bars of the subreference bars reconstruction with three calibration methods. The height of each error bar represents the
standard deviation, and the blue dot represents the average of 20 measurements. (a) Our proposed method. (b) 2-D classical method. (c) Deng’s method.

Fig. 19. Time consumption of different methods. (a) Total time. (b) Time
of each stage.

consumes significantly less time than the other two meth-


ods. In the stage of local calibration and global calibration,
our method consumes more time than the classic calibration
method, but in terms of total time, we are still the winner.

Fig. 18. Schematic of how the reference bars are placed in the measurement
volume. G. Another System Experiments
To better evaluate the proposed method, we build another
To be more specific, the whole process of the methods can multicamera system with different cameras, different lenses,
mainly be separated into three steps, including image acqui- and different camera layouts. The measurement system is a
sition, local calibration, and global calibration. Notice that rectangular parallelepiped structure with a size of 1290 mm ×
the time of image acquisition includes the time of manually 1150 mm × 990 mm, which consists of 12 MV-CA050-20GM
moving the calibration object and the time of automatically cameras equipped with V1226-MPZ lenses distributed on two
acquiring the image by the cameras. The time consumption parallel planes, as shown in Fig. 20. There are 12 ring light
of the involved three steps is shown in Fig. 19(b). As can be sources coaxially distributed with the camera to improve the
found, in the image acquisition stage, our proposed method lighting conditions of the system.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
5016715 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 70, 2021

The SPO algorithm is proposed to well initialize the spatial


coordinates of feature points in the GCS, which can help
obtain a more robust initialization value and further improve
calibration accuracy. Finally, we build a multicamera system
with ten cameras, and extensive experiments conducted on
this system prove that the proposed method can improve
calibration accuracy effectively and meet the requirement of
industrial measurement.

R EFERENCES
[1] X. Ban, H. Wang, T. Chen, Y. Wang, and Y. Xiao, “Monocular visual
odometry based on depth and optical flow using deep learning,” IEEE
Trans. Instrum. Meas., vol. 70, pp. 1–19, 2021.
[2] L. Song, S. Sun, Y. Yang, X. Zhu, Q. Guo, and H. Yang, “A multi-
view stereo measurement system based on a laser scanner for fine
workpieces,” Sensors, vol. 19, no. 2, p. 381, Jan. 2019.
[3] M. Jiang, R. Sogabe, K. Shimasaki, S. Hu, T. Senoo, and I. Ishii,
“500-Fps omnidirectional visual tracking using three-axis active vision
Fig. 20. Another system with different cameras, different lenses, and different system,” IEEE Trans. Instrum. Meas., vol. 70, pp. 1–11, 2021.
camera layouts. [4] A. Ruchay, V. Kober, K. Dorofeev, V. Kolpakov, and S. Miroshnikov,
“Accurate body measurement of live cattle using three depth cameras
and non-rigid 3-D shape recovery,” Comput. Electron. Agricult., vol. 179,
Dec. 2020, Art. no. 105821.
[5] X. Ziang, “Principle analysis of computer vision and its application
research,” in Proc. 7th Int. Conf. Adv. Mater. Comput. Sci. Ottawa, ON,
Canada: Clausius Scientific Pr, 2018, pp. 478–482.
[6] C.-Z. Dong and F. N. Catbas, “A review of computer vision—based
structural health monitoring at local and global levels,” Struct. Health
Monitor., vol. 20, no. 2, pp. 692–743, Mar. 2021.
[7] C. Chen, G. Wang, Y. Liu, and Y. He, “Online automatic measurement
of deflection for automobile based on digital image,” in Proc. IEEE Int.
Conf. Veh. Electron. Saf., Jul. 2013, pp. 42–47.
[8] T. Birdal, E. Bala, T. Eren, and S. Ilic, “Online inspection of 3D parts
via a locally overlapping camera network,” in Proc. IEEE Winter Conf.
Fig. 21. Calibration result of 12-camera system with our proposed method. Appl. Comput. Vis. (WACV), Mar. 2016, pp. 1–10.
(a) Reprojection error of each camera. (b) Standard deviation error bars of the [9] E. Khoramshahi et al., “Accurate calibration scheme for a multi-camera
measurement of the subreference bar. The height of the error bars represents mobile mapping system,” Remote Sens., vol. 11, no. 23, p. 2778,
the standard deviation, the blue dots represent the average of 20 measurements, Nov. 2019.
and the red symbols represent a part of the measured value. [10] G. Straube, C. Zhang, A. Yaroshchuk, S. Lübbecke, and G. Notni, “Mod-
elling and calibration of multi-camera-systems for 3D industrial supervi-
As shown in Fig. 21(a), the maximum reprojection error sion applications,” Proc. SPIE, vol. 11144, Sep. 2019, Art. no. 111440D.
is 0.12 pixels, which is similar to the experimental results [11] Z. Pusztai, I. Eichhardt, and L. Hajder, “Accurate calibration of
of the ten-camera system with the same method. With multi-LiDAR-multi-camera systems,” Sensors, vol. 18, no. 7, p. 2139,
Jul. 2018.
20 measurements of each subreference bar, the standard devia- [12] Z. Zhang, “A flexible new technique for camera calibration,” IEEE Trans.
tions of the measurements are shown in Fig. 21(b), which give Pattern Anal. Mach. Intell., vol. 22, no. 11, pp. 1330–1334, Nov. 2000.
the stability of the measured distances. The standard deviations [13] F. Abedi, Y. Yang, and Q. Liu, “Group geometric calibration and recti-
fication for circular multi-camera imaging system,” Opt. Exp., vol. 26,
of the reconstruction error are better than 0.125 mm, which no. 23, pp. 30596–30613, 2018.
is similar to the experimental result of the ten-camera system [14] J. Zhang, J. Zhu, H. Deng, Z. Chai, M. Ma, and X. Zhong, “Multi-
with the same method. Therefore, we can conclude that the camera calibration method based on a multi-plane stereo target,” Appl.
Opt., vol. 58, no. 34, pp. 9353–9359, 2019.
proposed method still has perfect performance for systems [15] L. Huang, F. Da, and S. Gai, “Research on multi-camera calibration and
with different layouts, different cameras, and different lenses. point cloud correction method based on three-dimensional calibration
object,” Opt. Lasers Eng., vol. 115, pp. 32–41, Apr. 2019.
[16] H. Deng, K. Yang, Q. Quan, and K.-Y. Cai, “Accurate and flexible
V. C ONCLUSION calibration method for a class of visual sensor networks,” IEEE Sensors
In this article, we propose a high-accuracy and convenient J., vol. 20, no. 6, pp. 3257–3269, Mar. 2020.
[17] Z. J. Yu et al., “Online calibration for a stereo vision measurement
method for the calibration of a multicamera system, which system,” Key Eng. Mater., vols. 295–296, pp. 723–728, Oct. 2005.
does not need to prepare a high-accuracy calibration object. [18] A. M. Truong, W. Philips, J. Guan, N. Deligiannis, and L. Abrahamyan,
To be specific, we first design a simple 3-D calibration “Automatic extrinsic calibration of camera networks based on pedes-
trians,” in Proc. 13th Int. Conf. Distrib. Smart Cameras, Sep. 2019,
object, where dot-distribution coded targets are pasted on pp. 1–6.
a sphere. Then, we propose an efficient and high-accuracy [19] T. Svoboda, D. Martinec, and T. Pajdla, “A convenient multicamera
calibration method, which mainly consists of two powerful self-calibration for virtual environments,” Presence, Teleoper. Virtual
Environ., vol. 14, pp. 407–422, Aug. 2005.
modules, i.e., the optimal transformation path algorithm and [20] P. Sun, N.-G. Lu, M.-L. Dong, B.-X. Yan, and J. Wang, “Simultaneous
the SPO algorithm. The optimal transformation path algorithm all-parameters calibration and assessment of a stereo camera pair using
is designed to find the optimal path of the current camera a scale bar,” Sensors, vol. 18, no. 11, p. 3964, Nov. 2018.
[21] T. Yang, Q. Zhao, X. Wang, and D. Huang, “Accurate calibration
to the root camera, which reduces the length of the trans- approach for non-overlapping multi-camera system,” Opt. Laser Tech-
formation path and can further improve calibration precision. nol., vol. 110, pp. 78–86, Feb. 2019.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.
TAO et al.: CONVENIENT AND HIGH-ACCURACY MULTICAMERA CALIBRATION METHOD 5016715

[22] T. Werner, D. Harrer, and D. Henrich, “Efficient, precise, and convenient Renbo Xia received the bachelor’s and master’s
calibration of multi-camera systems by robot automation,” in Proc. Int. degrees from Harbin Institute of Technology,
Conf. Robot. Alpe-Adria Danube Region. Cham, Switzerland: Springer, Harbin, China, in 2000 and 2002, respectively, and
2018, pp. 669–677. the Ph.D. degree from the University of Chinese
[23] R. Xia, M. Hu, J. Zhao, S. Chen, and Y. Chen, “Global calibration of Academy of Sciences, Beijing, China, in 2006.
multi-cameras with non-overlapping fields of view based on photogram- He is currently a Research Fellow with
metry and reconfigurable target,” Meas. Sci. Technol., vol. 29, no. 6, the Shenyang Institute of Automation, Chinese
Jun. 2018, Art. no. 065005. Academy of Sciences, Beijing. His research interests
[24] R. Xia, J. Zhao, W. Liu, and J. Xu, “Fully automatic matching of circular include computer vision, image understanding, and
markers for camera calibration,” in Proc. 3rd Int. Conf. Intell. Syst. intelligent manufacturing.
Knowl. Eng., Nov. 2008, pp. 1065–1070.
[25] P. Sturm and B. Triggs, “A factorization based algorithm for multi-image
projective structure and motion,” in Proc. Eur. Conf. Comput. Vis. Berlin,
Germany: Springer, 1996, pp. 709–720.
[26] P. Sturm, “Algorithms for plane-based pose estimation,” in Proc. IEEE
Conf. Comput. Vis. Pattern Recognit. (CVPR), Jun. 2000, pp. 706–711.
[27] Ueshiba and Tomita, “Plane-based calibration algorithm for multi- Jibin Zhao received the bachelor’s degree in
camera systems via factorization of homography matrices,” in Proc. 9th mechanical engineering from Hefei University of
IEEE Int. Conf. Comput. Vis., Oct. 2003, p. 966. Technology, Hefei, China, in 1996, the master’s
[28] C. Zhao, J. Zhang, H. Deng, and L. Yu, “Multi-cameras calibration from degree from Shandong University, Jinan, China,
spherical targets,” Proc. SPIE, vol. 9903, Jan. 2016, Art. no. 99032T. in 2000, and the Ph.D. degree from the Graduate
[29] M. Feng, S. Huang, J. Wang, B. Yang, and T. Zheng, “Accurate School, Chinese Academy of Sciences, Beijing,
calibration of a multi-camera system based on flat refractive geometry,” China, in 2004.
Appl. Opt., vol. 56, no. 35, pp. 9724–9734, 2017. He is currently a Researcher with the Shenyang
[30] M. Feng, X. Jia, J. Wang, S. Feng, and T. Zheng, “Global calibration of Institute of Automation, Chinese Academy of Sci-
multi-cameras based on refractive projection and ray tracing,” Sensors, ences, Shenyang, China. His research interests
vol. 17, no. 11, p. 2494, Oct. 2017. include computer-aided design, rapid prototyping,
[31] C. S. Fraser, “Automatic camera calibration in close range photogram- and reverse engineering.
metry,” Photogramm. Eng. Remote Sens., vol. 79, no. 4, pp. 381–388,
Apr. 2013.
[32] T. Luhmann, C. Fraser, and H.-G. Maas, “Sensor modelling and camera
calibration for close-range photogrammetry,” ISPRS J. Photogramm.
Remote Sens., vol. 115, pp. 37–46, May 2016.
[33] C. T. Aslan, K. Bernardin, and R. Stiefelhagen, “Automatic calibration
Tao Zhang received the B.S. degree from Wuhan
of camera networks based on local motion features,” in Proc. Workshop
University of Technology, Wuhan, China, in 2016.
Multi-Camera Multi-Modal Sensor Fusion Algorithms Appl. (M2SFA2),
He is currently pursuing the Ph.D. degree with
2008, pp. 1–13.
the State Key Laboratory of Robotics, Shenyang
[34] T. Zhang, Y. Cong, G. Sun, Q. Wang, and Z. Ding, “Visual tactile fusion
Institute of Automation, Chinese Academy of Sci-
object clustering,” in Proc. AAAI Conf. Artif. Intell., vol. 34, no. 6, 2020,
ences, University of Chinese Academy of Sciences,
pp. 10426–10433.
Shenyang, China.
[35] T. Zhang, Y. Cong, J. Dong, and D. Hou, “Partial visual-tactile fused
His current research interests include tactile
learning for robotic object recognition,” IEEE Trans. Syst., Man, Cybern.
sensing, visual-tactile fusion, robotics, and multi-
Syst., early access, Jul. 21, 2021, doi: 10.1109/TSMC.2021.3096235.
view clustering.
[36] T. Zhang, Y. Cong, G. Sun, and J. Dong, “Visual-tactile fused graph
learning for object clustering,” IEEE Trans. Cybern., early access,
Jun. 16, 2021, doi: 10.1109/TCYB.2021.3080321.
[37] T. Luhmann, S. Robson, S. Kyle, and J. Boehm, Close-Range Pho-
togrammetry and 3D Imaging. Berlin, Germany: Walter de Gruyter,
2013.
[38] W. Zeng and R. L. Church, “Finding shortest paths on real road
networks: The case for A,” Int. J. Geograph. Inf. Sci., vol. 23, no. 4, Yueling Chen received the bachelor’s degree in
pp. 531–543, Apr. 2009. automation from Xidian University, Xi’an, China,
[39] A. Goyal, P. Mogha, R. Luthra, and N. Sangwan, “Path finding: A* or in 2011, and the master’s degree from the Pattern
Dijkstra’s,” Int. J. IT Eng., vol. 2, no. 1, pp. 1–15, 2014. Recognition and Intelligent System Laboratory,
[40] Uncertainty Measurement—Part 3: Guide to Expression Uncertainty Xidian University, in 2014.
Measurement (GUM: 1995), International Organization for Standard- She has been with the Shenyang Institute
ization, ISO, Geneva, Switzerland, 2008. of Automation, Chinese Academy of Sciences,
Shenyang, China, since 2014, where she is currently
an Assistant Professor. Her current research interests
include image processing and 3-D reconstruction.

Liming Tao received the M.E. degree from Shengpeng Fu received the B.S. degree in logistics
Shandong University of Technology, Zibo, China, engineering from Shandong University, Jinan, China,
in 2018. He is currently pursuing the Ph.D. in 2007, and the Ph.D. degree in mechanical
degree with the State Key Laboratory of Robotics, and electronic engineering from the University of
Shenyang Institute of Automation, Chinese Chinese Academy of Sciences, Beijing, China,
Academy of Sciences, University of Chinese in 2015.
Academy of Sciences, Shenyang, China. He is currently an Associate Researcher with
His current research interests include computer the Shenyang Institute of Automation, Chinese
vision and pattern recognition. Academy of Sciences, Shenyang, China. His
research interests include computer vision and pat-
tern recognition.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on November 09,2023 at 08:58:49 UTC from IEEE Xplore. Restrictions apply.

You might also like