Skip to main content

Signal and Knowledge Integration with Decisional Control for Multi-Sensory Systems

Objective

Thepurpose of the SKIDS project is to provide a basic generic approach, for both software and hardware, in the area of integration of sensory information and knowledge. "Sensory information" is understood as information coming from an outside, physical,real world, and "knowledge" as high-level symbolic representations and models of the external world and of the system's features and abilities. Such models are dynamically updated and partially acquired through learning.
The ultimate goal of the project is a perception machine represented by the SKIDS demonstrator prototype and realising:
-a unified perception of the observed world
-real-time reasoning, planning and adaptation of the whole software and hardware configuration to the actual observations strategy.
The purpose of the project was to provide a basic generic approach, for both software and hardware, in the area of integration of sensory information and knowledge. Sensory information is understood as information coming from an outside, physical, real world, and knowledge as high level symbolic representations and models of the external world and of the system's features and abilities. Such models are dynamically updated and partially acquired through learning. The demonstration environment where the prototype perception machine will run has been specified, in particular the sensor configuration. The functional architecture has been defined, and consists of 4 parts: the MMI the sensory chain, the interpretative chain, and the control and decisional chain. The last 2 parts are essential: the interpretation processs, which is data driven (continuous surveillance task) or goal driven (object recognition upon request) is segmented into elementary tasks which are driven by the knowledge based control system (KBCS). The KBCS selects the optimal interpretative path and manages the global resources allocation. The basic perception tasks that have been identified fall into 5 categories: detection, characterization, localization, tracking and identification. The sensors consist of fixed and pan and tilt cameras, microphones, optical barriers, a laser range finder, an ultrasonic belt, and an odometer, all mounted on a mobile platform. The hardware has already been specified and consists of a set of nodes linked via a ring bus. Basic tools for the software architecture have already been identified: they include inference engines and a rule compiler for achieving real time performance of the perception system. The objective is to achieve response time of a few seconds for indoors scene surveillance. The fusion of information from multiple cameras has been demonstrated successfully for single event tracking. Various tasks of detection, localization and recognition demonstrated t he soundness of the vision node architecture.
The demonstration environment where the prototype perception machine will run has been specified, in particular the sensor configuration. The functional architecture has been defined, and consists of four parts:
-the MMI
-the sensory chain
-the interpretative chain
-the control and decisional chain.
The last two parts are essential: the interpretation process, which is data-driven (continuous surveillance task) or goal-driven (object recognition upon request) is segmented into elementary tasks which are driven by the Knowledge-Based Control System (KBCS). The KBCS selects the optimal interpretative path and manages the global resources allocation. The basic perception tasks that have been identified fall into five categories: detection, characterisation, localisation, tracking and identification.The sensors consist of fixed and pan-and-tilt cameras, microphones, optical barriers, and a laser range finder, an ultrasonic belt, and an odometer, all mounted on a mobile platform. The hardware has already been specified and consists of a set of nodes (VME clusters) linked via a Capitan ring bus. Basic tools for the software architecture have already been identified: they include inference engines and a rule compiler (KHEOPS) for achieving real-time performance of the perception system. The objective isto ach eve response time of a few seconds for indoors scene surveillance.
The fusion of information from multiple cameras has been demonstrated successfully for single event tracking. Various tasks of detection, localisation and recognition demonstrated the soundness of the vision node architecture, which consists of a Datacubesystem connected to a transputer array and hosted in a SUN3.
Exploitation
The approach is basically a generic one, but is driven by two classes of application:
-mobile robots for public safety applications in nuclear plants, etc
-surveillance systems for offshore oil fields, nuclear plants, airports, etc.

Coordinator

MATRA SA
Address
3 Avenue Du Centre
78052 Saint-quentin-en-yvelines
France

Participants (7)

British Aerospace plc
United Kingdom
Address
Sowerby Research Centre Filton
BS12 7QW Bristol
CENTRE NATIONAL DE LA RECHERCHE SCIENTIFIQUE
France
Address
Avenue Du Colonel Roche,7
31077 Toulouse
CHR. F. ROVSING A/S
Denmark
Address
Adelgade 1
1304 Koebenhavn
Krupp Atlas Elektronik GmbH
Germany
Address
Sebaldsbrücker Heerstraße 235
28309 Bremen
MAPS INFORMATICA INDUSTRIAL
Spain
Address
Monegal
08023 Barcelona
THE CHANCELLOR, MASTERS AND SCHOLARS OF THE UNIVERSITY OF OXFORD
United Kingdom
Address
Parks Road
OX1 3PJ Oxford
UNIV POLITECNICA DE CATALUNYA
Spain
Address
C/pau Gargallo Apartado 30002
08028 Barcelona