On Image Matching and Feature Tracking for Embedded Systems: A State-of-the-Art - Advances in Heuristic Signal Processing and Applications - page 351

16.2 Matching: Concepts, Algorithms, and Architectures
Matching and tracking are essential functions for many vision-based processes.
However, they still present difficult problems, especially for implementation in em-
bedded or wearable systems.
16.2.1 Matching: Basic Concepts
In general, matching deals with the identification of certain attributes or character-
istics associated with a given relationship. As far as image matching is considered,
this definition can be reformulated as follows: matching two images of the same (2D
or 3D) scene involves the identification of the geometric (2D or 3D) transformation
that superposes one image onto another. The parameters of the transformation may
be estimated from matches between selected features, sometimes referred to as con-
trol points.
The mathematical model of the geometric transformation depends on the image
acquisition sensors, the required accuracy, and any bounds on algorithmic complex-
ity. The image acquisition system may produce noisy data subject to geometric dis-
tortions. If the error model is known, then it is useful to pre-process images with
an error reverse model. In the case of noise, a Gaussian model is often appropriate
even if the true but unknown distribution is known not to be Gaussian. The most
common geometric image distortions, i.e., deviations from rectilinear projection,
are radial and tangential distortions. The former include quadratic, barrel, and pin-
cushion distortions. These distortions can be corrected with the Brown distortion
model [ 14 ]. Camera calibration can include distortion error correction [ 2 , 97 , 101 ].
16.2.2 Characteristics for Matching
Matching can be performed in the space domain or in the frequency domain. In
applications which require fast matching using parallel hardware, matching in the
space domain is preferred.
Image matching methods are classified as area based or feature based, as follows:
1. Area-based methods: matching is carried out using a direct comparison of pixel
values. A dense set of image matches is often sought.
2. Feature based methods vary according to the selected features. Possible features
include:
- Locally salient features, for example, local extrema, in which an image region
differs in some systematic way from its neighboring regions [ 61 , 82 , 93 ].
- Predefined features in which a feature is defined by a specific property of a set
of pixels, for example, a standard deviation in a particular range or a particular
texture [ 42 , 61 ].
- Application specific features [ 33 , 64 ], for example, faces and gestures.