945
INTEGRATED RANGE CAMERA CALIBRATION
USING IMAGE SEQUENCES FROM HAND-HELD OPERATION
Wilfried Karel
Christian Doppler Laboratory for Spatial Data from Laser Scanning and Remote Sensing at the Institute of
Photogrammetry and Remote Sensing, Vienna University of Technology Gusshausstrasse 27-29 / E122, 1040 Vienna,
Austria - wk@ipf.tuwien.ac.at
KEY WORDS: Imaging Systems, 3D Sensors, Calibration, Close Range Photogrammetry, Video Camera, Machine Vision
ABSTRACT:
This article concentrates on the integrated self-calibration of both the interior orientation and the distance measurement system of a
time-of-flight range camera that employs amplitude-modulated, continuous-wave, near-infrared light (photonic mixer device - PMD).
In contrast to other approaches that conduct specialized experiments for the investigation of individual, potential distortion factors, in
the presented approach all calculations are based on the same data set being captured under near-real-world conditions, guided by
hand and without auxiliary devices serving as high-order reference. Flat, circular targets stuck on a planar whiteboard and with
known positions are automatically tracked throughout the amplitude layer of long image sequences. These image observations are
introduced into a bundle block adjustment, which on the one hand results in the determination of the camera’s interior orientation
and its temporal variation. On the other hand, the reconstructed exterior orientations and the known planarity of the imaged board
allow for the derivation of reference values of the actual distance observations. These deviations are checked on relations with the
reference distance itself, the observed signal amplitude, the integration time, the angle of incidence, and with both the position in the
field of view and in object space. Eased by the automatic reconstruction of the camera’s trajectory and attitude, several thousand
frames are processed, leading to comprehensive statistics.
1. INTRODUCTION
Range Imaging (RIM) denotes the capture of distances at the
pixels of a focal plane array using simultaneous time-of-flight
measurements. Thereby, the round-trip time of the emitted
signal and hence the object distance may be determined in
various ways. Currently, alternatives employing nanosecond
Laser pulses are highly investigated. So-called Flash LADARs
or Laser Radars utilize avalanche photo diodes (APD) for
photon detection, eventually supported by photo cathodes for
signal amplification (Stettner et al., 2004), and may facilitate
maximum observable distances of up to a kilometre. Another
technique, called Time-Correlated Single Photon Counting,
employs the most sensitive single photon avalanche diodes
(SPAD) as detectors (Aull et al., 2002, Niclass et al., 2007,
Pancheri et al., 2007, Wallace et al., 2001), minimizing the
requirements on the illumination power. In Multiple Double
Short Time Integration (MDSI), different fractions of the echo
energy are captured in consecutive images by varying the
shutter speed and using conventional chips that integrate the
irradiance (Mengel et al., 2001, Elkhalili et al., 2004). The
requirements for nanosecond Laser pulses or high-speed
shutters and the integration of highly precise, miniaturized
timing circuitry however boost the complexity and costs of all
these systems.
Photonic mixer devices (PMD, lock-in pixels) employ
incoherent, near-infrared, amplitude-modulated, continuous
wave (AM-CW) light and determine the signal phase shift and
hence the object distance by mixing the emitted with the
returned signal at every pixel (Spirig et al., 1997, Lange et al.,
1999). As on the one hand, the illumination unit of PMD close-
range cameras may be realized with low-cost LED arrays, and
on the other hand, the operation point of the system is limited to
the single frequency of modulation, implying that the demands
on the electronic components are lower (Buttgen et al., 2005),
the purchase costs are comparatively low, lying in the range of
professional SLR cameras.
The present article concentrates on this latter technique,
implemented in the Swissranger™ SR-3000 by MESA Imaging
AG. This instrument samples the correlation function of the
emitted and returned signal one after another at every quadrant
of the modulation period. It features a sensor resolution of
144x176 pixels, a fix-focus lens, range and amplitude data
encoded with 16 bit, an overall ranging precision of a few
centimetres, and a maximum range of 7.5m when using the
default modulation frequency. As other PMD cameras, its
ranging system suffers from large systematic distortions
reaching decimetres, why comprehensive calibration methods
are needed in order to harness the potentials of RIM for
geometry- and quality-oriented realms.
1.1 Related Work
For the correct reconstruction of the object space imaged by a
range camera, knowledge of its interior orientation is an
essential prerequisite in combination with undistorted range
observations. Westfeld (2007) determines the intrinsic
projection parameters by application of conventional
photogrammetric techniques to amplitude images, and reports
an unstable position of the principal point. Having calibrated
the camera optics, Lindner and Kolb (2006) gather range
images at known distances from a planar target and experience
the deviations of the range measurements as being partly of
periodic nature. Steitz and Pannekamp (2005) report influences
of the angle of incidence and the surface type on the range
observations. Kahlmann et al. (2006) perform elaborate
laboratory experiments and also reveal the partly cyclic non-
linearities of the distance observations. Additionally,