This paper describes a vision processing architecture which, together
with a robotic camera platform and control subsystem, comprise a multi
-processor real-time system for research into active vision. Experimen
tation in time-critical aspects of the active vision paradigm over sev
eral years has demonstrated the flexibility of 'pipe-groups' used as t
he building block in the system. The design of the entire system is ov
erviewed, and the design of pipe-groups and their software is describe
d in more detail, An active method of determining the exact processing
latency, required for stable control, is explained. The paper illustr
ates the use of the pipe-group processing architecture and its real-ti
me processing capabilities, with an example of active stereo tracking
and structure recovery using the method of affine tranfer. (C) 1997 Ac
ademic Press Limited.