[ Go to content ] [ Help ] [ Information about accessability ]

Computer Vision Laboratory

Information

Research

Scientific Output

Education

LiU - ISY - CVL > Objects and Scenes

Object and Scene Recognition

The core of vision system capabilities is to navigate in a world of realistic complexity and to recognize objects and events in this world to perform appropriate actions. This implies the ability to recognize and to handle objects, parts of objects, relations between objects, relations between objects and their context in what we call scenes, dynamic characteristics of objects, etc. For simplicity, we will in the ensuing discussion refer to all such entities of interest as objects. This is also due to a belief that all these entities can be handled in a generically similar manner. In fact, our belief is that if significant progress shall be made, it requires that generic methods can be found.

A structure for unrestricted recognition of 3-D objects is under development. By unrestricted, we imply that the recognition shall be done independently of object position, scale, orientation and pose, against a structured background. It shall not assume any preceding segmentation and allow a reasonable degree of occlusion.

The recognition structure utilises sparse features representing regions/blobs, lines and curvature/corners. Such features are computationally efficient, and they are suitable for channel representation.

The object representation uses a hierarchy of triplet feature invariants [gm2003], which are at each level defined by a learning procedure. The method uses a learning architecture [gfj03] employing channel information representation, [gg2000g].

This structure is intended for use together with controlled cameras or other sensors, where the focus-of-attention (FOA) can be switched quickly, and zoom controlled to provide details at a higher resolution.

The structure is also used in a 2-D version for navigation.

Navigation using blob features. Red lines indicate the image region used by the algorithm (ground truth) and blue lines indicate the estimated region. The red and blue crosses are the true and estimated camera position respectively. (click to view movie).

Last updated: 2010-05-31

Responsible for this page:
webmaster

Org.nr: 202100-3096
VAT nr: SE202100-309601
Plusgiro: 18 34 15-9

Computer Vision Laboratory
Linköpings universitet
SE-581 83 LINKÖPING
Tel: +46 13 281000
Fax: +46 13 138526