Introduction
Feature extraction and shape representation are important
problems in computer vision, especially in robot vision and
pattern recognition applications. Such applications usually
impose difficult real-time requirements, as the complete
understanding of the relevant features of a scene has to be
performed in real-time, i.e. less than a second. Until now,
however, these requirements have only been fulfilled by
imposing several simplifying assumptions: the position and
orientation of objects must be tightly constrained, the con-
trast between the background and the foreground must be
sufficient to allow trivial segmentation using thresholding,
and the objects must not overlap. Although our understand-
ing of computer vision is sufficiently advanced to address
the general problems of feature extraction and shape repre-
sentation in unstructured environments, their real-time
implementation is difficult to achieve because of the large
amount of information to process. Typically, extensive
arrays of special-purpose units have been required, with
corresponding consequences in cost and bulk.
A promising research area in computer vision seeks to
exploit special-purpose architectures embedded in the
photo-detection plane in order to provide the processing
capabilities required for real-time applications. These
1077-2014/97/050307 + 11 $25.00/ri960066 © 1997 Academic Press Limited
Mixed-signal VLSI Architecture for
Real-Time Computer Vision
his paper presents the architecture of a computer vision system targeted for real-time robot
vision and pattern recognition applications. The proposed mixed-signal very large scale
Tintegration (VLSI) architecture integrates photo-transduction with low- and medium-level
processing such as multi-resolution edge extraction, scale-space integration, edge tracking, domi-
nant point extraction, and database generation. Its high performance stems from a custom CMOS
smart image sensor providing parallel access to illuminance data and a set of parallel analog filters
performing multi-resolution edge extraction. We have also developed a digital controller which
manages data flow between the processing modules of the system and which constructs a database
of the observed scene under the supervision of a digital signal processor (DSP) unit. This database
describes relevant object contours as a linked list of linear segments and circular arcs with precom-
puted local and global properties. Such a token description of the scene is suitable for robot vision
and pattern recognition applications, since it significantly compresses the amount of data to be
processed by further high-level algorithms. Experimental results obtained with the current proto-
type of the system are very promising, with the complete process, from image acquisition to scene
database creation, performed in less than a second.
© 1997 Academic Press Limited
Stéphane Dallaire, Marc Tremblay and Denis Poussart
Computer Vision and Systems Laboratory, Department of Electrical and Computer Engineering,
Laval University, Québec, Canada, G1K 7P4
Email: stefdal@gel.ulaval.ca
Real-Time Imaging 3, 307–317 (1997)