Next: References Up: Visual Tracking of Solid Previous: 4 Tracking the 3D

5 Experimental Results and Concluding Remarks

   
Figure 2: Rendered image model (Maple V) and three frames out of the sequence with overlaid silhouette and B-spline control points.

The prototype system has been tested for smooth curved objects. It performs well on synthetic image sequences. Because of the lack of physical CAD-models for objects, it was not tested on real image sequences. However, we demonstrated the robustness of the 2D contour tracker in [ 10 ] on real image sequences of polyhedra.

In figure 2 sample frames of a synthetic image sequence are shown. The object was tracked through a sequence of 50 frames, where the object was rotated in and by 0.5 degree per frame and in by one degree per frame. Translation was applied to keep the rotating object in the field of view. Tracking results are plotted in figure 3 . The tracking was performed on a DEC5000/230 station (60 MHz) achieving an average frame rate of 0.85 fps. However, until now, little effort has been made to improve the performance of the current prototype system.

   
Figure 3: Tracking results for the angles measured in degree over a sequence of 40 frames: Ground truth is plotted with linespoints and estimates are represented by dots.

We are currently extending the system for concave objects exhibiting T-junctions in the occluding contour and on piecewise smooth objects with creases at the surface. Furthermore, we investigate the influence of 2D tracking errors on the 3D pose estimation.

In [ 10 ] it was shown that with the presented framework tracking of polyhedra can be executed in real time on modest hardware. The currently used, slightly outdated, hardware suggests that tracking of smooth objects can also be implemented in real time.

We see the main contribution of the presented method in the split of the tracking process into two autonomous working processes, which interchange information and influence and control each other, respectively. This opens a natural way to distribute the computational load. The particular tracking methods for the contour as well as for the pose tracking are exchangeable and can be chosen depending on the local and global geometry of the object's surface.