Jump to content

EU Centre of ExcellenceISO 9001

ERCIMW3C MemberFraunhofer Project Center

Strukturális információ az érzékelők mérési terében

Full name: Structural information in the space of sensor networks
Department: Machine Perception Research Laboratory
Start date: 2008. 10. 01.
End date: 2012. 09. 30.
External identifier: OTKA 76159
Cost: 22 MFt

Project manager

Tamás Szirányi
Tamás Szirányi
Address: 1111 Budapest, Kende u. 13-17.
Room number: K 409
Phone: +36 1 279 6106
Fax: +36 1 279 6292
E-mail: sziranyi.tamasEZT_TOROLJE_KI@EZT_TOROLJE_KIsztaki.mta.hu
Homepage: http://www.sztaki.hu/~sziranyi/




When we apply sensors to measure a workspace, we can meet some interesting questions about the information content of he examined system:
   * How can we estimate that something important happened in that space?
   * Can we estimate the degree of freedom of the space of events?
   * Measuring the events in the space, can we estimate the effect of an external impact?
   * If we directly manipulate some processes in the events’ space, can we measure the effect? Meaning that we want to measure the change in the events' process affected by the intrusion.
   * Without having a priori knowledge about the space, how much measure of geometrical structures or physical processes can be estimated through evaluating the sensor network?

These questions appear as exciting challenges in surveillance systems, traffic control, alarm systems, embedded medical devices and security applications.

We have a main question in this project, namely, when measuring a scene with a network of sensors, what is the measure of valuable information, which can be used as a feature-set for definite problems, like as indexing and retrieval.

In a given scene (surveillance, industrial testbed, medical supervision) the valuable information does not exist alone; it must be related to joint scenes and time instants. When asking about the importance of information, we should compare the measured parameters to that of the other scenes and time instants. However, events of other space/time instants may have a different set of significant parameters.

For this reason the task of comparison is not a simple indexing and retrieval, but chaining partly overlapping sets of parameters.

When climbing to a higher level of abstraction, we may have problems with the classification, which needs prior information. However, we must avoid just this turn: the question is not "what is it similar to?", but "Is it similar to anything so much that we should consider it?".

We can manipulate the measured space of events: by querying, tuning the setting, changing the relations among the levels of evaluation or rebuilding the network of sensors. This control can work as reinforced learning, or just a reinforced query can amend the evaluation. The control can affect the measured information content, resulting in a more structured association among objects.

The scene may contain geometrical, physical and causality relations. Their assessment can contribute to the better evaluation of the scene.

We may try with the next solutions:

* Detecting the important changes in a dynamic scene through stochastic optimization
   * Estimating the degree of freedom through reduction of dimension of the statistical model
   * Measuring the rate of the effect of intervention by estimating the change in the event-spaces’ dimensionality
   * Projecting physical processes and their models into the space of network of sensor-signals

Naturally, now we have more exciting questions than possibly useful solutions. On the basis of our previous research activity and the complexity of the raised problems we reckon on further challenges and interesting solutions. In the project we will face new theoretical problems in artificial vision, measurement theory, and information theory. The requested support may help in achieving several new applicable solutions and theoretical clarification.