On Image Matching and Feature Tracking for Embedded Systems: A State-of-the-Art - Advances in Heuristic Signal Processing and Applications - page 367

Fig. 16.15 Visible spectrum radial symmetry guided particle filter. Left image summarizes the
principle of the pdf update. Right image shows the tracking results for x and y pupil center position;
in red : raw true data, in blue : the estimated data
can be initialized by either using the first measurement or by training the filter using
sample sequences.
Several methods based on particle filtering for feature tracking and for combining
object models have been proposed. An application of particle filtering to eye track-
ing has been recently proposed. Hansen and Pece [ 36 ] combine a particle filter with
a generalized Laplacian for coding grey-level differences, and a Gaussian distribu-
tion for deformation modeling; Wu et al. [ 98 ] use a particle filter and a 3D model for
the eye; Martinez [ 59 ] combines a particle filter with the radial symmetry in visible
spectrum images of the eye in order to update the associated pdf (Fig. 16.15 ).
A priori knowledge of the targeted object can be provided during an initializa-
tion stage using interest point detection. However, all these methods are not robust
against occlusions or changes in the appearance of the object and have a high com-
putational complexity. There is no suitable hardware for real time processing.
16.4 Final Comments and Potential Future Developments
This chapter has described algorithms for image and image sequence matching.
Many of these algorithms are suitable for embedded hardware implementation. The
new hardware oriented computational structures can be used not only for image
matching, but for image processing and image analysis in general.
Image matching, despite its long history and significant progress, is still an ac-
tive research area. 2D matching is better understood than 3D matching, however,
2D matching still lacks pertinent approaches to deal with occlusion and to include
additional information about the matching.
There are very few efficient methods for 3D-3D matching and for 2D-3D match-
ing. For future wearable and autonomous systems, it seems necessary to build the
matching model using additional sensors in order to obtain reliable feature tracking