Trinocular stereo system demo

We propose a trinocular stereo system which general features are:

Uses straight line segments as features.
The segments are presented using a probabilistic method.
Fusion is done with Extended and Iterated Kalman Filter.
The matching/reconstruction is computed using a predict-match-update loop.

The defining feature of the systems is that the matching considers both the image segments collinearity and the matching of the segments midpoints. However the weight for the midpoint matching is lower than
the weight for the collinearity.

The main qualities of the system are:

Low spurious matches.
Solution contains both and estimate of the 3D segments location and their covariance.

You can download the full stereo environment which includes, segment extaction, reconstruction and visualization. All the interfaces are written in java and the computations in ANSI C. This program was written by J.L Laviña as his final studies proyect.

A papers describing the computations:

Montiel J.M.M, Montano L. "Probabilistic structure from camera location using straight segments"
Image Vision Computing. Elsevier, Vol. 17, issue 3-4, March 1999, pp. 263-279. (Abstract)
Martinez Montiel, J.M Zhang, Z and Montano L. "Segment-Based Structure from an Imprecisely Located Camera" 1996 IEEE Int. Symposium on Computer Vision, ISCV'96. pp182-187 Coral Gables, Florida., USA, 21-23 Nov1996. (abstract), (paper, 300Kb, PostScript compressed)

You can see the 3D reconstruction and the correspondences with a java applet wirtten by David Berna Sanjuán. The available examples are:

Stairs image

In order to show the trade-off assignment for the weight of the pairing between the image segments midpoint, 3 different reconstructions are shown for the same image triplet :

Person Image

To show the performance on a more difficult triplet see the following reconstruction

reconstruction

back to JMMM home page
please, send your comments to JMMMontiel
Last Updated Jan 12 2000 by JMMM