Sebastian Bronte's WebPage - Master Thesis Results

Master Thesis Results

Stereo face tracking system

As it was said before, these works are maid for Pedro Jiménez's PhD Thesis.

Briefly, the main system is based on a 3D tracking algorithm taking its input from a stereo system, which detects features inside a region of interest, and then, triangulate and track features to estimate the pose of the object. In this particular case, the objects are faces.

My main contribution to this system was to increase the number of points in the head model. The model used in this case was aquired from scan-laser and it was decimated to make the system works with the model. Some samples of the process followed to decimate the model are described in the section below.

Tu sum up, the increase of model points makes the system more robust against fast rotations. On the other hand, due to the high number of points, the estimation of the pose is a little bit worse because RANSAC has a fixed number of iterations, which is function of the number of points. Also the increase of model points makes the processing time too high, again, function of the points considered.

go to top

3D scan-laser models

As it is shown in the title, a procedure to obtain and decimate models from a 3D scan-laser is exposed.

Considering that it is imposible to obtain a complete 3D reconstruction from a single view, to take (at least) 3 views of the same object from different viewpoints is necesary for a complete reconstruction. In the particular case of head pose tracking, 3 views are needed: 1 frontal and 2 lateral views to match common points and add new ones from the ears.

Once we have a dense model from the different views, we proceed to decimate the model using feature points from the image associated to the model, and then the 3D coordinates for each view.

Having all the decimated points from each view, we proceed to estimate the rotation and translation matrix from lateral to frontal views, and then mix al the points into only one to integrate it to the tracking algorithm.

Some result samples are shown below, modeling a robot and a face from dense reconstructions to easy to use models. For each sample, we show some samples and also some acquired models in vrml format. To see them go to http://cic.nist.gov/vrml/vbdetect.html and download a plug-in to see the content of the vrml file.