Multiple-View Video for Traffic
Surveillance (OIS)
Project-Page
3D Scene Reconstruction
If Multiple-View images are
available for a specific scene, a navigation through
the scene can be provided. The necessary interpolation of intermediate views however
strongly depends on the appropriate images. Within the OIS-project original
images strongly differ in lighting conditions and shadowing, since convergence
angles between different input views often exceed 90°. Therefore algorithms
based on block or luminance similarities (e.g. disparity estimation) cannot be
used. In our Multiple-View data, the images were calibrated, using a number of
reference points in the real 3D-scene. Projected image positions for each view
were than provided for each view. Thus, it was possible to match correspondent
image content by calculating the appropriate projection matrix between two
input images. The following example shows intermediate views between two input
images of a traffic crossroad.
(1.5MB, please be patient, may take a
while)
Navigation between different viewpoints
The images represent
background information that was obtained from traffic sequences by
Moving-Object-Detection and segmentation (see Object Segmentation, Tracking
and 3D Reconstruction). Furthermore, no 3D-Reconstruction was used in this
example, since intermediate views were directly created from the warped
original images.
People:
Michael
Droese
Karsten Mueller
Aljoscha Smolic
Last modified: May 27, 2002