INS-Camera Calibration

without Ground Control Points

Daniel Bender∗ , Marek Schikora∗ , Jürgen Sturm† and Daniel Cremers†

∗ Department Sensor Data and Information Fusion, Fraunhofer FKIE, Wachtberg, Germany
Email: daniel.bender@fkie.fraunhofer.de, marek.schikora@fkie.fraunhofer.de
† Computer Vision Group, Technical University of Munich, Garching, Germany

Email: sturmju@in.tum.de, cremers@in.tum.de

Abstract—In this paper we present an approach for perform-

ing system calibration of a sensor suite consisting of a fixed
mounted camera and an inertial navigation system without the
usage of a photogrammetric calibration site. The aim of the
presented work is to obtain accurate direct georeferencing of cam-
era images collected with manned or unmanned aerial systems
for search and rescue tasks. These time-critical missions require
a straightforward calibration which can be performed without
additional equipment. This induces an in-flight calibration of
the intrinsic camera parameters in addition to the mounting
offsets between the camera and the inertial navigation system.
The optimization of these values can be done by introducing
them as parameters into a bundle adjustment process. We show
how to solve this by exploiting a graph optimization framework.
The evaluation of the proposed approach with data from flight Fig. 1. Our goal is to calibrate the static coordinate system offsets between
experiments leads to an improvement of roughly factor six a camera and an INS, which enables the utilization of the INS measurements
compared to a terrestrial calibration. as exterior camera orientations. For our flight experiments we integrated the
sensors into a payload pod, which was mounted beneath one wing of a manned
ultra light airplane.
I. I NTRODUCTION
Aerial photogrammetry is widely used for orthophoto and
digital terrain model creation. Nowadays these measuring
flights are usually performed by an aircraft equipped with
a digital large-format camera and a GPS corrected inertial
navigation system (INS). This equipment enables the direct
determination of the camera positions and orientations for
the captured images. Therefore it provides, in contrast to the
classical bundle adjustment, a very flexible flight planning
and no necessity of ground control points (GCP) during the
operation. Nevertheless, the essential system calibration is
generally realized by a flight over a photogrammetric test site
with a high number of signaled GCPs.
Time-critical surveillance and rescue tasks have the same Fig. 2. Visualization of a flight course performed for the evaluation of our
demand for a flexible flight planning and the direct determi- approach. The altitude was about 300 meters.
nation of object positions from image observations [1], [2].
In contrast to aerial photogrammetry, there is a need for a
straightforward system calibration, which can be performed edges. We determine the optimal configuration of the state
without the time-consuming and cost-intensive usage of large- variables by using a graph optimization framework [5]. In
scale calibration areas. Within the so called INS-camera cal- contrast to previous work relying on BA [3], our approach is
ibration the static coordinate system transformation between more intuitive and can easily be adapted when other calibration
the rigidly mounted sensors is determined. More precisely the parameters become relevant.
translational offsets (lever-arm) and the angle misalignments
(boresight) have to be estimated. The most accurate calibration Our previous work on simulated data indicated high accu-
procedures integrate these parameters as unknowns in a bundle racies for the graph based system calibration [6]. We extend
adjustment (BA) [3], [4]. the procedure to perform the calibration with real-world data
from sensors mounted on an aerial vehicle. In our experiments,
In this work, we formulate the system calibration as a we used a manned ultra light airplane with a wingspan of
graph optimization problem by representing the parameters 9.6 meters (Fig. 1) flying in heights of 300 to 800 meters
to optimize as vertices and the observations between them as (Fig. 2). Our evaluation shows that a calibration can be
performed without the usage of GCPs. The paper starts with an
overview on related research areas. It follows the description
of the realized system calibration. Finally experimental results
achieved with this approach are shown and discussed.

II. R ELATED W ORK

A related process of the INS-camera calibration is the
so called hand-eye calibration. Given a camera mounted on
a robot arm, the rigid-body transformation between the co-
ordinate systems of these devices is estimated. As a result,
measurements from the acquired images can be transformed
into the robot arm coordinate system to interact with objects
recognized and located in the images. The calibration out of
corresponding robot arm and camera poses was realized with
a direct solution for the rotational part followed by solving
the equations for the translation [7]. Alternatively a nonlinearNorth2
optimization of the whole transformation at the same time was Fig. 3. Both, the navigation systems of the INS and our world reference
proposed in [8]. frame are based on East, North, Up (ENU) coordinate systems. They are local
Cartesian coordinate systems with the origin tangential to the earth ellipsoid.
These approaches had a big influence on the calibration Further we use the global earth-centered, earth fixed (ECEF) coordinate
between an inertial measurement unit (IMU) and a camera. system, to convert between ENU-coordinates and GPS measurements taken in
Measurements of the IMU in form of rotational velocities latitude (ϕ) and longitude (λ) as polar coordinates.
and linear accelerations can be integrated to determine the
positions, velocities and orientations of the device. The small
estimation errors which are summed up over time can be flights at photogrammetric calibration sites [4]. In contrast to
corrected by camera measurements, if the transformation be- this our goal is to achieve direct georeferencing with small
tween the two devices is known. In [9], the estimation was manned or unmanned aerial systems, without relying on time-
realized with a modified hand-eye calibration algorithm. Other consuming and cost-intensive GCPs. We phrase the calibration
approaches utilize a Kalman filter for the estimation of the pose problem as a graph that we solve using a standard graph
transformation [10], [11]. This sensor combination is known optimizer. This approach allows the intuitive integration of the
as a vision-aided INS. mounting offsets as parameters to optimize and can easily be
adapted when more parameters become relevant.
By contrast, GPS-aided INS exploit the GPS measurements
to correct the IMU estimations. In conjunction with GPS
correction signals from ground control stations, accuracies in III. INS-C AMERA C ALIBRATION
the range of a few centimeters for the positions and a few In this section, we describe the performed calibration of a
hundredths of a degree for the orientations are achievable. Thus sensor suite comprising a camera and an INS. First, we define
the INS provides a reliable stand-alone source describing its the utilized coordinate systems and our parametrization of the
own movement. This leads to the estimation of the rigid-body rigid-body motions as well as the camera mapping. Then, we
transformation between the INS and the camera with methods state the problem considered in this work and show how to
similar to the hand-eye calibration. In a first step the camera solve it with a graph based approach.
movement is calculated with a structure from motion (SFM)
approach and refined in a BA procedure. The observations A. Preliminaries
of ground control points are used to scale the 3D model to
real world coordinates. In a second step, the transformation The INS measures the orientation as Euler angles with
between the two devices is estimated by relating these abso- regard to local navigation systems. The latter are local East,
lute camera poses to time synchronized measurements from North, Up (ENU) coordinate systems, each with the origin
the INS. This widely used approach is known as two-step in the current device position described through a GPS mea-
procedure [12]. The advantage is that each bundle adjustment surement in form of latitude, longitude and altitude (Fig. 3).
package can be used without modifications. On the other hand, This implies tiny differences between the orientations of these
the integration of the mounting parameters as variables to coordinate systems and thus also in the measured angles for
optimize in the BA is possible. This approach is known as successive timestamps. Our INS-camera calibration will be
single-step calibration and induces a simpler estimation of the performed in an area of less than one square kilometer, which
mounting offsets due to more flexible flight courses [3]. The allows to neglect these differences. Besides, we perform the
estimation of the INS-camera calibration has been examined INS-camera calibration in an ENU coordinate system W with
widely within the OEEPE test [4]. They conclude that this the origin in the middle of the observed area. This prevents the
approach is a serious alternative for many applications, even need for earth curvature corrections which have to be taken into
though it does not achieve the mapping accuracies of classical consideration for mapping frames like the Universal Transverse
bundle adjustment. Mercator (UTM) coordinate system.
The studies for the INS-camera calibration performed in The rigid body motions gCiW describe the camera poses
the last decades were targeted at aircrafts equipped with high- at the exposure times ti , i = 1, 2, ..., n with regard to the
precision INS and metric large-format cameras performing world frame W. Likewise, gIiW specify the corresponding
configurations of the INS as rigid body motions. In general,
a rigid body motion g ∈ SE(3) describes how the points camera (C)
of a rigid object change over time. Instead of considering INS (I)
the continuous path of the movement, we bring into focus
gCI ?
the mapping between the initial and the final configuration of
the rigid body motion. This movement can be described by a
rotation matrix R ∈ SO(3) and a translation vector t ∈ R3 .
Consequently the rigid body displacement G of a 3D point
p ∈ R3 can be performed as gIiW
gCiW
G : SE(3) × R3 → R3 , G(g, p) = Rp + t . (1)
The representation of the rotational part in form of the overde-
termined rotation matrix R is not suitable for the optimization
performed in this work. Thus we use a minimal representation
in form of YXZ Euler angles (ψ, θ, φ) ∈ R3 , called yaw, pitch world frame (W)
and roll. These describe the rotational part of the movement as
a mapping from R3 to SO(3) by consecutive rotations around
Fig. 4. The estimation of the static rigid body motion gCI between the devices,
the principal axes as enables by composition with the INS measurement gIiW the description of the
"
cos ψ − sin ψ 0 1 0
#"
0
#"
cos φ 0 sin φ
# camera pose gCiW .
R = sin ψ cos ψ 0 0 cos θ − sin θ 0 1 0 .
0 0 1 0 sin θ cos θ − sin φ 0 cos φ
(2)
The YXZ rotation order leads to singularities at θ = ±π/2.
This corresponds to a pitch angle of ±90◦ , which will never
be achieved in practice by the considered aerial platforms and
therefore does not constitute a problem.
Further we define the set of intrinsic camera calibration
parameters
k = {fx , fy , ox , oy , k1 , k2 , k3 , p1 , p2 } , (3)
whereby (fx , fy ) describe the focal length and (ox , oy ) the
principal point of the camera. Further we express the radial
distortion with the parameters (k1 , k2 , k3 ) and the tangential
distortion with (p1 , p2 ). The projection π performs the map-
ping from a transformed 3D point G(g, p) = (x, y, z)> to Fig. 5. The classical checkerboard calibration for a camera equipped with
wide angle lens and focus at infinity requires bulky calibration patterns.
pixel coordinates by
 >
rfx x rfy y
π(k, G(g, p)) = − ox + tx , − oy + ty , static and especially comprises that the rigid body motions
z z
(4) describing the movements from the INS to the camera at
with the radial distortion factor r and the tangential distortion various exposure times are constant
offsets tx and ty being defined as follows [13]: !
 2 2 3 ∀k, l ∈ {1, 2, ..., n} : gCk Ik = gCl Il , (8)
x + y2
 2
x + y2
 2
x + y2

r = 1 + k1 + k2 + k3 , where n is the number of images. Therefore we can simplify
z2 z2 z2
(5) the notation by omitting the time indices for the rigid body
xy
 2
3x + y 2
 motion gCI = (ψB , θB , φB , xL , yL , zL ) describing the mounting
tx = 2p1 + p2 , (6) offsets as boresight and lever-arm. The composition with the
z z2 measured INS movement leads to the camera motion
 2
x + 3y 2

xy gCiW = gCI gIiW . (9)
ty = p 1 2
+ 2p2 . (7)
z z
Thus the knowledge of the mounting offsets is required to
B. Problem Formulation describe the camera poses with the INS measurements.
The objective of this research is the direct determination of The use of a wide angle lens in conjunction with focus at
the camera positions and orientations for the captured images infinity makes the use of standard camera calibration proce-
by exploiting measurements from a high-precision INS. In dures laborious (Fig. 5). Furthermore, the estimations for the
order to describe the rigid body motion gCiW of the camera intrinsic camera parameters k tend to differ slightly from a
using the measured rigid body motion gIiW of the INS (Fig. laboratory calibration, due to the climate and environmental
4), the devices have to be synchronized in time and rigidly conditions [14]. Thus it is an advantage to optimize these
mounted. This induces that the offsets between them are parameters by using images from a measurement flight.
 Therefore the joint calibration of the intrinsic camera
parameters k in conjunction with the mounting offsets gCI out gCI gIiW gCiW xij pj
of synchronized data from measurement flights is the problem Ii
considered in this work.

C. Algorithm Overview k

We generate initial camera poses by combining the mea-

sured INS poses with the initial mounting offsets determined Fig. 7. The objective function of the stated problem can be illustrated by
from construction drawings. The images and initial camera a hyper-graph. The measurements (boxes) are presented as links between
poses are the input for a SFM approach, which calculates a the nodes concerning each multiple sets of variables (circles). For improved
sparse 3D point cloud out of consistent image observations overview multiple state variables and measurements of the same type are
visualized in a stacked view unrelated to their number of occurrence.
[15]. A refinement of these parameters is performed by the
optimization of a non-linear function, which can be phrased
as a graph. Our entire workflow is depicted in Fig. 6. D. Graph Optimization
We optimize the system calibration by introducing con-
straints between the intrinsic camera parameters, camera poses
camera INS and 3D points. The resulting non-linear error function can be
represented as a graph by introducing the variables to optimize
as nodes and the observations between them as edges. We
determine the optimal configuration of the state variables with
a graph optimizer [5]. In the following we will present our
objective function and how to express it in the graph-based
INS formulation.
image
image poses
image The calibration parameters represented as the intrinsic
camera parameters k and the mounting offsets gCI are added as
nodes to the graph. Furthermore, we add each 3D point pj and
each camera pose gCiW as a node to the graph. The connection
between these nodes is given by inserting observations as edges
initial initial into the graph. A pixel measurement connects three different
SFM camera system nodes, namely: a camera, a 3D point and the intrinsic camera
poses calibration parameters. This constraint can be realized with a hyperedge,
which is able to connect an arbitrary number of nodes. The
edge of an INS measurement connects the corresponding rigid
body motion of the camera with the mounting offsets. A
initial visualization of the graph is presented in Fig. 7.
3D points
and image Further, we have to define error functions, which measure
observations
how well measurements are described by the state variables
they are connecting. The first constraint measures the error
occurring from the reprojection of a 3D point into the image.
The error function for this constraint can be expressed as

exij (k, gCiW , pj , xij ) = π(k, G(gCiW , pj )) − xij . (10)

graph
optimization
The resulting error vector has dimension two and is 0 if
the pixel observation xij is perfectly described by the state
variables. The second error function states how well the INS
measurements can be described by the composition of the
camera poses gCiW and the mounting offsets gCI as follows:
refined refined
refined
camera system egi (gCI , gCiW , gIiW ) = (gCI−1 gCiW )gI−1
iW
. (11)
3D points
poses calibration
Using a minimal representation for the rigid body motions in
form of three Euler angles and a three-dimensional translation,
Fig. 6. Overview of the algorithm. Data from the INS and the camera are we receive a 6-dimensional error vector, which is 0 if the
processed to determine the system calibration. The concatenation of the latter parameters perfectly satisfy the measurement.
with the measured INS poses enables the very accurate determination of the
camera poses.
Without limiting the generality, we refer to the whole
state vector [k, gCiW , pj , gCI ] as y and formulate our objective
 TABLE I. C ALIBRATION RESULTS FROM THE CHECKERBOARD
function as follows: METHOD AND THE PRESENTED APPROACH
Xn X m n
X
min x > x x
eij (y) Ωij eij (y) + egi (y)> Ωgi egi (y) , (12) ψB θB φB fx fy ox oy
y
i j i [◦ ] [◦ ] [◦ ] [pel] [pel] [pel] [pel]

where the information matrices Ωxij and Ωgi respectively checker

flight 1
0.0
0.846
0.0
0.215
0.0
-0.072
3334.68
3342.89
3343.5
3334.88
1744.32
1730.6
1238.06
1227.9
represent the inverse covariance of the pixel observations and flight 2 0.816 0.205 -0.068 3343.4 3335.44 1724.04 1231.19
INS measurements. A numerical solution of Equation (12) flight 3 0.795 0.205 -0.074 3343.73 3335.95 1725.39 1230.74
flight 4 0.805 0.193 -0.078 3346.28 3338.36 1730.62 1234.04
is computed with the Levenberg-Marquardt (LM) algorithm
and therefore a good initial guess y̆ of the state vector is
needed. Iteratively, the first order Taylor expansion around the
current guess y̆ is used to approximate Equation (12) and
optimize the local increments ∆y by solving the resulting
sparse linear system. The center for the next iteration is
obtained by adding the optimized increments to the current
guess. This is done by using the motion composition for the
state variables represented by rigid body motions and a simple
addition for the 3D points and intrinsic camera parameters. For
a detailed description of the LM algorithm we refer the reader
to [5], [16].

IV. E VALUATION
In this section, we present achieved results from flight
experiments to validate the proposed approach. The equipped Fig. 8. Visualization of the camera poses and 3D points introduced as vertices
INS is based on fibre optic gyroscopes, which have a stability in our graph optimization.
up to some hundredths of a degree per hour. In combination
with real time kinematic enhanced GPS measurements very
accurate pose information are generated. The optical system estimation of the vertical lever-arm and all other mounting
consist of a downward looking camera with 8 megapixels and internal camera parameters can be decoupled from each
and a wide angle lens with a field of view of 54 ◦ in the other. Due to these observations, we fixed the lever-arm to the
horizontal and 42 ◦ in the vertical direction. The sensors were terrestrial measurements in our optimization.
integrated into a payload pod, which was mounted beneath the
wing of a manned ultra light airplane (Fig. 1). The performed The resulting calibration parameters show a hight stability
flight course has to introduce measurements that constrain all for different flights and differs from the initial checkerboard
dimensions of the calibration parameters. For our platform calibration (Table I). This holds especially for the distortion
small movements in all axes occur even for straight and level parameters, that we omitted here for the sake of readability.
flights, which aim at a constant heading and altitude by accom- The differences between the optimized boresight angles ψB ,
plishing immediate corrections to unintentional movements. θB and φB are very close to the stated accuracy of the INS.
We performed a total of four flights within two days. At an Small variations for the intrinsic camera parameters occur
altitude of 300 meters and above, we captured two images per most likely due to different climate conditions during flight
second at a speed of approximately 125 km/h. The flights were execution. To evaluate the accuracy of the achieved results
performed as crossing straight lines (Fig. 2). To achieve a high we performed a least-square forward intersection for the pixel
image overlap, we use only images within a circle of a radius observations of five GCP. The image coordinates of these
of 600 meters around the central point. This results in a total
number of nearly 700 images for the first two flights and 300
images for the other two.
These images were used to calculate an initial 3D point
cloud of the observed area with a SFM approach under con-
sideration of the camera poses [17]. The latter were generated
by a concatenation of the INS measurements with the initial
mounting offsets determined through terrestrial measurements.
The output of the SFM (Fig. 8) was used as input for
the graph optimization. Thereby the pixel observations were
introduced as measurements with an accuracy of 1 pixel. Since
no quality log files of the INS were available we considered
the manufacturer information of an accuracy of 2 cm in the
position, 0.04 ◦ for the yaw and 0.01 ◦ for the pitch and roll
angles. Our previous work [6] reveals that the optimization
of the translational part of the mounting offsets leads to a Fig. 9. We used five ground control points to verify the accuracy of our
lower accuracy compared to usual terrestrial measurements. approach. These were placed at dominant image corners to allow easy manual
As further stated in [18], at least one GCP is needed for the measurement of their image coordinates (orange circles).
 TABLE II. M EAN E UCLIDIAN DISTANCE BETWEEN FIVE GROUND
CONTROL POINTS AND THE FORWARD INTERSECTION FOR THE INITIAL
R EFERENCES
AND THE OPTIMIZED CALIBRATION ( LEFT ) AS WELL AS THE CALIBRATION [1] M. Schikora, D. Bender, W. Koch, and D. Cremers, “Multi-target multi-
FROM FLIGHT 1 FOR ALL FLIGHTS ( RIGHT ) sensor localization and tracking using passive antennas and optical
sensors on UAVs,” Proc. SPIE Security + Defence,, vol. 7833, pp. 1–9,
init. opt. gain opt. opt. flight 1 gain 2010.
[m] [m] [factor] [m] [m] [factor]
[2] M. Schikora, D. Bender, and W. Koch, “Airborne emitter tracking by
flight 1 3.17 0.53 5.98 flight 1 0.53 0.53 1.0 fusing heterogeneous bearing data,” in Proc. of the 17th International
flight 2 2.94 0.47 6.26 flight 2 0.47 0.55 0.85 Conference on Information Fusion (FUSION), 2014.
flight 3 2.02 0.37 5.46 flight 3 0.37 0.66 0.56
flight 4 4.45 0.44 10.11 flight 4 0.44 0.58 0.76 [3] L. Pinto and G. Forlani, “A single step calibration procedure for
IMU/GPS in aerial photogrammetry,” International Archives of the
Photogrammetry, Remote Sensing and Spatial Information Sciences,,
vol. 34 Part B3, pp. 210–219, 2002.
points were measured manually (Fig. 9) and used to perform [4] C. Heipke, K. Jacobsen, and H. Wegmann, “Analysis of the results
a forward intersection with the initial and optimized camera of the OEEPE test ’Integrated sensor orientation’,” Integrated sensor
poses. This leads to 3D coordinates, which were compared to orientation - Test report and workshop proceedings, OEEPE Official
Publications No., vol. 43, pp. 31–49, 2002.
values measured with a mobile GPS-receiver. The latter stated
[5] R. Kuemmerle and G. Grisetti, “g2o: A General Framework for Graph
a horizontal accuracy of about 30 cm and a vertical accuracy of Optimization,” in Proc. of the IEEE International Conference on
about 50 cm. Our results from the forward intersection are in Robotics and Automation (ICRA), 2011, pp. 3607–3613.
the same range, which shows the performance of our approach [6] D. Bender, M. Schikora, J. Sturm, and D. Cremers, “Graph-based bundle
(Table II). We assume that the larger initial error of flight adjustment for ins-camera calibration,” International Archives of the
4 occurs due to the range from 300 to 800 meters for the Photogrammetry, Remote Sensing and Spatial Information Sciences,,
altitude, which was smaller for the other flights. Nevertheless, vol. XL-1/W2, pp. 39–44, 2013.
the results using the optimized calibration parameters for flight [7] R. Tsai and R. Lenz, “A new technique for fully autonomous and effi-
cient 3D robotics hand/eye calibration,” in Proc. of the 4th International
4 are in the same range as for the other flights. The generation Symposium on Robotics Research, 1989, pp. 287–297.
of the camera poses out of calibration results from flight 1 for
[8] R. Horaud and F. Dornaika, “Hand-Eye Calibration,” International
all flights leads to a slightly decreasing performance (Table II). Journal of Robotics Research,, vol. 14, no. 3, pp. 195–210, 1995.
Given an altitude of 300 meters and above, the accuracies are [9] J. Lobo and J. Dias, “Relative Pose Calibration Between Visual and
high and clearly outperform our terrestrial calibration. Inertial Sensors,” International Journal of Robotics Research,, vol. 26,
no. 6, pp. 561–575, 2007.
V. C ONCLUSION AND F UTURE W ORK [10] F. Mirzaei and S. Roumeliotis, “A Kalman filter-based algorithm
for IMU-camera calibration: Observability analysis and performance
In this paper we presented a graph-based approach for evaluation,” IEEE Transactions on Robotics,, vol. 24, no. 5, pp. 1143–
the system calibration of a sensor suite consisting of a fixed 1156, 2008.
mounted camera and an INS. We showed how to phrase the [11] S. Weiss and M. Achtelik, “Versatile distributed pose estimation and
sensor self-calibration for an autonomous mav,” in Proc. of the IEEE
optimization problem as a graph and estimated the mounting International Conference on Robotics and Automation (ICRA), 2012,
offsets between the devices and the intrinsic camera parameters pp. 31–38.
with a graph optimization framework. Our evaluation points [12] M. Cramer, D. Stallmann, and N. Haala, “Direct georeferencing using
out that a straightforward system calibration without the usage GPS/inertial exterior orientations for photogrammetric applications,”
of GCP leads to results which show high potential for cost- International Archives of the Photogrammetry, Remote Sensing and
saving in flight calibrations. It was shown that the calibration Spatial Information Sciences,, vol. 33 Part B3, pp. 198–205, 2000.
results can be used for consecutive flights, but the highest [13] D. C. Brown, “Decentering distortion of lenses,” Photogrammetric
Engineering, vol. 32, no. 3, pp. 444–462, 1966.
precision will be obtained by performing the INS-camera
calibration during the mission. Compared to our terrestrial [14] K. Jacobsen, “Aspects of handling image orientation by direct sensor
orientation,” in Proc. of the ASPRS Annual Convention, 2001.
calibration, we achieved in our experiments an improvement of
[15] R. Hartley and A. Zisserman, Multiple View Geometry in Computer
roughly factor six, which can be even higher for other setups Vision. Cambridge University Press, 2004.
due to larger angle misalignments between the devices. [16] M. Lourakis and A. Argyros, “SBA: A software package for generic
Future work will investigate the proposed procedure in sparse bundle adjustment,” ACM Transactions on Mathematical Soft-
ware, vol. 36, no. 1, pp. 1–30, 2009.
more detail. Furthermore, we will perform experiments with
[17] C. Wu, “Towards linear-time incremental structure from motion,” in
cost-efficient MEMS INS and evaluate if the approach is International Conference on 3D Vision, 2013, pp. 127–134.
also usable to perform the INS-camera calibration if no RTK [18] A. Kersting, A. Habib, and K. Bang, “Mounting Parameters Calibra-
corrections are performed. tion of GPS/INS-Assisted Photogrammetric Systems,” in International
Workshop on Multi-Platform/Multi-Sensor Remote Sensing and Map-
ping (M2RSM), 2011, pp. 1–6.