In conclusion, there are many new directions that seem promising for this rapidly expanding subfield. We have seen that the incorporation of probabilistic reasoning is being used to provide effective integration, allowing representation of context, control, and even learning. The use of iconic representations, which are easily acquired from example images and can be used in subsequent recognition of the objects and behaviours, is also a major new direction. In a very different direction, there is a requirement to formalise reasoning to provide provably correct behaviour in many applications. This requires close interaction of specialised subfields using logic in AI and high-level vision research. Another move in this direction is the integration of work in vision and language for many application areas in advanced surveillance, medical analysis systems, and multimodal interaction. Work on such interactive systems forces the developers to use frameworks with a common semantics and to adopt cognitive models of the system users.
In addition to the directions above, there are new themes that are beginning to influence work on knowledge-based vision. In particular, the combination of deformable models with dynamic learning of their statistical properties seems set to grow rapidly. As we have seen, there are many applications of deformable models in biomedical image analysis [17], face recognition [42], and tracking of people [3]. The acquisition of these models by training can allow the development of generative models and these are now starting to model physical forces for visual understanding. In addition, situated cognition is now being taken seriously in the interpretation and understanding community, although there is a rejection of strong anti-representational positions like that of Brooks [11]. For example, work by Howarth and Buxton [13,30] on situated behavioural analysis and work by Mackworth [43] on situated agents for robotics. These examples are just a part of the groundshift over the last two decades from the traditional approach based on symbolic reasoning in Good Old-Fashioned Artificial Intelligence (GOFAI) to simpler, behaviour-based approaches.