On Image Matching and Feature Tracking for Embedded Systems: A State-of-the-Art - Advances in Heuristic Signal Processing and Applications - page 350

human computer interaction, stereo vision, 3D reconstruction, multimodal fusion.
processing, remote sensing, etc.
16.1 Introduction
A digital image is a representation of a scene by a regular grid of numbers. The
places in the grid are pixels and the numbers are the pixel values. Each number codes
the intensity of the signal received from a small part of the scene. An image sequence
records the changes in the scene over time, for example, changes in illumination or
changes due to relative motion between the sensing device and the objects in the
scene.
Image processing includes both local and global operations on images. The most
important local operation is the construction of features which summarize the infor-
mation in small regions of the image. Important global operations include segmen-
tation and the matching of features between consecutive images in a sequence of
images. The applications of feature matching include stereo vision, structure from
motion, and the detection and tracking of moving objects.
There are many different types of feature, however, hardware limits and recent
results in psycho-cognitive vision suggest that point features are good candidates
for high quality image processing which can be carried out on dedicated (parallel)
hardware. Each point feature is a single location in the image such that the values
of the surrounding pixels have some easily identified property. Point features can be
efficiently and reliably matched or tracked.
The applications of image and feature matching and tracking are very numerous.
Even so, more than 50 years of research have not resulted in unique algorithms for
matching and tracking. Indeed, vision algorithms usually depend heavily on specific
heuristics which are tailored to the application. However, despite the many papers
and communications on tracking and matching published in recent years [ 103 ], few
of them take into account the constraints on implementation found in embedded or
wearable systems. This chapter overviews some methods for image matching and
feature tracking which are implemented in academic or research prototype systems
or which could be implemented in hardware suitable for an embedded or wearable
system. The algorithms are not presented in detail. The basic concepts of each algo-
rithm are described and the effects of their system integration on the quality of the
results are discussed.
The rest of the chapter is organized as follows: Sect. 16.2 addresses the match-
ing concept, extraction of the interest/saliency points as basic primitive for targeted
hardware implementation, the most popular algorithms for image matching, and
hardware supports for image/vision matching operations based graphs. Section 16.3
discusses the feature tracking problem, correlation, and Bayesian approaches. Sec-
tion 16.4 provides some final comments on the considered approaches of matching
and tracking, and some potential research points for new hardware developments.