ressed to
ion. The
direction,
ied. This
'eference
iour with
ensional
guideline
etrisches
| seine
ndeswehr
Distortion
- 6, June
. (2002):
dak DCS
002.
raphische
F.
ammetrie;
ns for an
isurement
letry and
rung von
tionen der
e element
e Sensing,
eras using
ering and
)1-1205.
Cipal point
odak DCS
): Camera
Elements;
systems —
'üsseldorf.
REAL-TIME PHOTOGRAMMETRIC ALGORITHMS FOR ROBOT CALIBRATION
J. Hefele
ifp, Institute for Photogrammetry, Geschwister-Scholl-Str. 24 70174 Stuttgart, Germany
juergen.hefele@ifp.uni-stuttgart.de
Commission V, WG V/1
KEY WORDS: Industry, Robotics, Calibration, Measurement, Close Range
ABSTRACT
The aim of this project is to investigate a new photogrammetric approach to determine the pose of the robot endeffector in real-time
for updating the robot model. Specifically, the two fundamental photogrammetric algorithms are investigated for that purpose:
intersection and resection. In both cases, cameras are mounted on the moving robot observing targets fixed on the floor. In the first
approach the camera pose (exterior orientation) with respect to the target co-ordinate system can be measured directly by using the
collinearity equation. In the second approach, the stereo-camera measures the position of the observed targets with respect to the
camera co-ordinate system. If the target co-ordinates are available the camera position with respect to the target co-ordinate system
can be determined. With a standard algorithm for hand-eye calibration the misalignment in-between camera and end-effector is
computed. Different set-ups with respect to ease of implementation, accuracy, and workspace size are compared. These setups were
simulated and verified in several investigations. In our test environment we an industrial robot KUKA KR15/2 and digital CCD
cameras with near infrared illumination were used.
1. INTRODUCTION
By ISO 8373 (International Standard Organization) industrial
robots are defined as freely programmable appliances with a
series of rigid components connected by joints. One end of the
component chain is fixed while the other end (end-effector) can
be moved by computer control. If there are for example six or
more revolute joints, the industrial robot can reach every point
of its working cell with every orientation. State-of-the-art
industrial robots are able to move objects with weights up to
500 kg and with a repeatability of 0.3 mm or better. In most
cases the joints are powered by electric motors whereas very
heavy robots are powered hydraulically. Because of the
maximum power at a relatively high speed of the electro
motors, the speed must be geared down. Forward kinematic
describes the relation between the motion of each joint and the
motion of the end-effector and thereby the position and
orientation (pose) of the end-effector in arbitrary coordinate
systems can be computed. Therefore robot model parameters
such as length of segments, distance between two adjacent
segments and rotation angle for revolute joint between two
segments have to be known. The rotation angle can be
measured exactly by a position encoder between motor and gear
unit. Other parameters are defined by the design plan of the
robot. But due to manufacturing and other environmental
influence they do not operate accurately.
The overall errors can be subdivided in geometric errors such as
etolerance of the segment length,
eangle error,
and non-geometric such as
e gear elasticity
esegment elasticity
etemperature influence.
By a robot-calibration the influence of several errors can be
eliminated (Whitney 1986, Heisel 1998, Wiest 2001), but there
are still remaining time dependent errors such as temperature
influence, tear and wear. These errors can only be reduced by
the design of the robot in order to get it in accordance with the
mathematical model. Still, the absolute accuracy of a robot is
much lower composed to the repeatability and can be as large as
several millimeters. The disadvantage of the approach of
absolute calibration is obvious since replacement of robot
components requires a complete recalibration. In addition, the
main disadvantage is that off-line programming is not possible,
as the required accuracy cannot be reached. To remove the
disadvantages an external measurement system for the direct
measurement of the robots pose is required. There are several
photogrammetric and non-photogrammetric measurement
systems on the market (Wiest 2001). Also, industry acceptance
of camera-based systems has increased in the last few years.
The main disadvantage of these systems is that they cannot be
used during production. Furthermore, the costs of these systems
are very high, sometimes higher than the cost of a robot. On the
other hand photogrammetry is certainly able to determine the
robots pose very accurately by using industrial standard
cameras at low cost (Maas 1997). In the ideal case the
frequency of the direct measurement of the robot pose and the
frequency of the robot control loop are the same. For modern
industrial robots the frequency of the control loop is between
200 and 1000 Hz. With inexpensive standard industrial camera
this frequency is not possible, because of the high data rate.
However, in our special case, only those parts of an image
containing targets are necessary for measurement. In the near
future CMOS-Cameras with direct access to the pixel will be
available. With these cameras a measurement rate of 200 Hz
and more is possible.
Nevertheless the problem of shadowing effects remains. For
example, a robot moving into a car body to fix a new element
does not have a direct view to targets fixed outside the car body.
In this case, the robot pose cannot be determined. Therefore, it
is easier to use direct pose measurement for updating the robot
model than to correct the control loop directly.
-33—