in a 2D/3D matching
framework
I have been interested in a way of robustly tracking points, line segments, rectangles and 3D corners in streams of video images, as an input for a 3D recovery technique. Basically, endpoints are tracked using point correlation techniques. After that, an edge detector (a thresholded/smoothed Canny-Deriche detector) is run on the image,
and the various geometric primitives (aka segments, rectangles and corners)
are matched on the edges. For the rectangle and the corner, the 3D pose of the
object can be computed from vanishing points or angular constraints. Therefore,
we reconstruct them prior to matching, to match directly a 3D model onto the
images, allowing only relevant deformations.
The following results have been obtained starting from primitives drawn in a first image, then tracked automatically along the complete sequence.
|
The Sport sequence (90 frames) |
The StainX sequence (40 frames) |
|
|
The Begijnhof sequence (50 frames) |
The Praxitele sequence (30 frames) |
Acknowledgements
This work has been done during my PhD in the MIRAGES
Team at Inria Rocquencourt.
I would like to thank the VISICS
lab for the Begijnhof sequence and Alias|Wavefront
and SGI for the stainX sequence.
Further details
Tracking geometric primitives in video streams, P. L. Bazin and
J. M. Vezien, Proceedings of the 4th IMVIP conference, 2000. (postscript).
