PE&RS November 2015 - page 847

Automatic Co-Registration of Pan-Tilt-Zoom (PTZ)
Video Images with 3D Wireframe Models
Ravi Ancil Persad, Costas Armenakis, and Gunho Sohn
Abstract
We propose an algorithm for the automatic co-registration of
Pan-Tilt-Zoom (PTZ) camera video images with 3D wireframe
models. The proposed method automatically retrieves chang-
ing camera focal length and angular parameters, due to the
motion of PTZ cameras by matching linear features between PTZ
video images and 3D CAD wireframe models. The developed
feature-matching schema is based on a novel evidence-based
hypothesis-verification optimization framework referred to
as Line-based Randomized RANdom SAmple Consensus (LR-
RANSAC). LR-RANSAC introduces a fast and stable pre-verifica-
tion test into the optimization process to avoid unnecessary
verification of erroneous hypotheses. An evidence-based veri-
fication follows to optimally select the PTZ camera parameters,
where an original line-based approach for full-verification,
-exploiting local geometrical cues on the image scene-, eval-
uates the pre-verified hypotheses. Tests on an indoor dataset
produced a 0.06 mm error in focal length estimation and
rotational errors in the order of 0.18° to 0.24°. Experiments on
the outdoor dataset resulted in a 0.07 mm error for focal length
and rotational errors ranging from 0.19° to 0.30°.
Introduction
Indoor and outdoor video images from surveillance monitor-
ing systems are used in computer vision tasks such as event
detection and tracking moving objects (e.g., pedestrians and
vehicles) (Adam et al., 2008; Leibe et al., 2008). The projection
of 2D object positions, detected from monocular single images,
into the 3D space is possible by determining the orientation
parameters of the camera with respect to the orthogonal axes
of 3D wireframe model. However, automatic estimation of
these parameters is critical, particularly for pan-tilt-zoom (PTZ)
surveillance cameras where the angular camera parameters
and the focal length change spontaneously with the camera
motion, while the position of the camera remains fixed. Our
research work is motivated by the problem of image-based
automatic localization of dynamic objects for augmented real-
ity applications using surveillance cameras (Figure 1). Using
virtual 3D environments populated with building models of
3D cityscapes and road networks, avatars of pedestrians and
vehicles detected from 2D video can be rendered in the 3D en-
vironments. In related works (Baklouti et al., 2009; Kim et al.,
2009), the transformation parameters to move from 2D to 3D
object positions are manually established. We introduce a new
approach for automatically estimating the parameters of PTZ
surveillance cameras relative to the virtual 3D scene models.
This would enable the transformation of image positions of
moving objects into the virtual 3D environment during times
of camera motion. Using the wireframes of 3D scene models,
we match their 3D lines with corresponding 2 D image lines.
Geomatics Engineering, GeoICT Lab, Department of Earth and
Space Science and Engineering, Lassonde School of Engineer-
ing, York University, 4700 Keele Street, Toronto, Ontario, M3J
1P3, Canada ( armenc@yorku.ca).
Photogrammetric Engineering & Remote Sensing
Vol. 81, No. 11, November 2015, pp. 847–859.
0099-1112/15/847–859
© 2015 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.81.11.847
Figure 1. Concept of video and 3D city models for augmented reality surveillance.