Work Area: Computer Vision
Keywords continuous computer vision, control of perception, integrated vision system, goal-directed scene interpretation
Start Date: 1 July 92 / Duration: 36 months / Status: running
[ participants / contact ]
Abstract The integration of basic techniques for the construction of a continuously operating vision system capable of interpreting a dynamically changing environment is studied. The main emphasis is on control of perception through the use of goal-directed focus-of-attention techniques. The approach exploits spatial and temporal contexts, multiple resolutions, and controlled motion of the sensor system. The work builds on the results of VAP (3038).
The VAP consortium aims to demonstrate that the paradigm of "vision as process" is basic to the function of a high-level vision system. Such a hypothesis can only be demonstrated within the context of a complete integrated vision system, where the potential benefits of continuous control of perception and of associated temporal context are evident. The project's goal is to refine existing vision techniques and to integrate them into a multi-purpose vision system.
VAP (3038) led to an initial integration of basic vision techniques. This integrated system will serve as a basis and test-bed for the continued study of techniques for control of perception at all levels of a vision system.
New techniques to interpret a dynamically changing, quasi-structured environment are being developed. These exploit goal-directed focus of attention involving controlled sensor motion. Processing is directed by goals which change dynamically in reaction to the needs of the task as well as to events in the scene. The motivation of this approach is to limit the computational complexity of the perception process by limiting the size of the internal models. These models, which must be continuously updated, describe the environment in terms of a number of qualitatively different phenomena, such as image phenomena, 3-D scene geometry, and symbolic interpretations of objects and events.
The necessary techniques are being developed in the context of integrated and continuously operating vision systems, which will serve as vehicles for testing the hypotheses. The research issues which are addressed include:
After a first integration of a "skeleton" VAP-system was demonstrated at the final review of VAP (3038), the year-one of this project has mainly concentrated on improving the individual elements and techniques of the integration.
The research is increasingly related to exploitation of active vision. Five controllable stereo camera heads with different virtues are now in operation at the partner's labs. Camera controllers and dynamic calibration systems have been developed. Ocular motor reflexes and a control strategy for dynamic fixation and explorative capabilities are in operation.
Real-time control strategies for space-variant sensors has been demonstrated for tracking, "time-to-crash" computation, and motion control based on qualitative optical flow. A new scheme of "Normalised differential convolution" has been developed for the filtering of missing and uncertain data. Colour information is being exploited, and a framework for recognition and matching of objects represented by attributed relational graphs has been developed using efficient probabilistic relaxation.
A system control approach based on the "Discrete Event Dynamic Systems" (DEDS) formalism from traditional control theory is under investigation and shows so far promising results.
VAP II will contribute to closing the gap between current vision approaches and techniques versus more general and "dynamic" multi-purpose vision systems.
The potential of such continuously operating systems as well as their facilities for actively acquiring data, opens for a new and large range of opportunities for pre-competitive research and industrial applications of machine vision.
The consortium has actively published the project results in various international journals and at a range of conferences and workshops. The manuscript of a book has been completed, and publication is due shortly.
An industrial affiliates programme for companies interested in developing and commercialising the results of the project has been established and held its first meeting in Stockholm in September 1992, shortly after the commencement of VAP II.
Project results are also disseminated through various other national and international projects and networks in which the partners participate.
Further information about VAP is available from the VAP home page <URL:http://www.vision.auc.dk/LIA/VAP>.
Ålborg Universitet - DK
Laboratory of Image Analysis
Fr. Bajers Vej 7, Bldg. D1
DK - 9220 ÅLBORG ST
I.N.P.G.-LIFIA - F
Università di Genova - I
KTH-Royal Institute of Technology - S
Linköping University - S
University of Surrey - UK
Prof. E. Granum
tel +45/9815 1133
fax +45/9815 4008
VAP II - 7108, August 1994
please address enquiries to the ESPRIT Information Desk
html version of synopsis by Nick Cook