Most real-time vision applications have a sensing-perception-action loop which involves i) video capture at frame rate (25 Hz) ii) vision processing which is mainly compute bound and iii) robot control. Such systems often have hard real-time deadlines to enable small latency reaction to real world events and ensure stability of control. Real time vision applications are typically developed in reasearch environments using high-end visual workstations which provide out-of-the-box solutions for frame capture and sophisticated tools for image processing, display, multithreading etc. The operating systems in such workstations are general purpose with large overheads and often do not provide any real-time guarantees. Consequently, a programmer has to mainly rely on the fast processors available in such systems to meet real-time requirements and can seldom enforce them in a systematic way.

However, the end use of most such vision applications lie in special purpose embedded systems requiring low cost solutions. Such embedded systems are often required to be built around low cost processors like the Intel X86 family or around special purpose DSP processors. They must necessarily support high speed frame capture and facilitate control of robotic and other special purpose equipment. Rather than being general purpose programmable boxes, embedded systems are special purpose application systems which require the vision algorithms to be integrated with the operating system to provide one complete application.