Future image understanding systems must be able to respond to scene dynamics within a fraction of a second if they are to be useful in real-time applications. Current image understanding systems are not only very limited in capability, but they are painfully slow. One approach to achieving real-time image understanding is to build faster hardware. This paper presents a different approach. Strategies for implementing real-time image understanding systems are discussed which offer alternatives to traditional computational paradigms. Requirements for real-time operation are shown to constrain the selection of artificial intelligence (AI) methodologies. These strategies are currently being tested in an experimental prototype vision system. The topics discussed in this paper are applicable to other non-vision AI applications as well.