Christensen, Henrik I., AI Magazine
The integration of AI and vision has been a long-term goal of both disciplines for more than three decades. This special issue illustrates some recent work on bridging the gap.
Forty years ago, the integration of computational vision and AI was considered in the thesis of L. G. Roberts (1963). Back then, the issue was anticipated relatively easily, and a full integration of the two fields was expected within a decade. Processing of images is known to result in noisy segmentation, and in general, data might not be consistent in space or time, making Al methods appealing. In early AI, the handling of noisy, partially inconsistent data was, at best, a major challenge. Nevertheless, there have been major efforts in knowledge-based interpretation and vision, for example, in Hanson and Riseman's (1978) seminal book and as part of the Defense Advanced Research Projects Agency's Image Understanding Program. One landmark was the ACRONYM system (Brooks 1983, 1981), which performed knowledge-based interpretation of three-dimensional objects. Despite these efforts, the emergence of fully integrated systems for complex tasks has been relatively limited. One of the limitations has been access to adequate computer facilities; another has been the lack of efficient methods for reasoning under uncertainty.
Because of the early efforts in integration of AI and computer vision, the available computer power has increased by more than 5 x [10.sup.7]; at the same time, there has been tremendous progress in reasoning under uncertainty (now an area with its own major conference). In addition, computer vision has also matured in terms of modeling, and today, there are well-developed models for description of features and assemblies of features, motion, and objects that use a sound Bayesian framework. Consequently, vision is gradually becoming a methodology that fits well with the corresponding methods in AI. Thus, there has recently been a major upsurge in research that integrates modern methods from AI and vision into systems. One of these research efforts is a major program run by the European Commission called the Cognitive Vision Program. This program, begun in 2000, currently involves 10 major research projects and a network for community integration called ECVISION. In addition, there is an effort in the United States to examine cognitive systems.
Scanning the Issue
In this special issue of AI Magazine, three different approaches to cognitive vision are presented.
In the first article, Ernst Dickmanns presents the results of a long-term effort to use vision for the autonomous control of cars. More than 20 years ago, Dickmanns embarked on an effort to build a vehicle that could drive autonomously on the road. The initial system could drive autonomously on a highway. The sheer need for computer power implied that the car had to be a truck; otherwise, it would have been too small to hold all the computers. Gradually, the methodology has been refined and extended into an impressive system for autonomous driving, automatic lane changing, interpretation of traffic patterns, and so on. The advances in computer power have at the same time resulted in a reduction in system size to a system that today can easily fit into a small rack in the back of a small car. The article illustrates how careful modeling of the imaging process, the target vehicle, and the domain of traffic allows design of highly competent systems that can be used not only on the highway but also in heavy traffic in urban environments. This study illustrates a holistic approach to the design of an advanced AI system for a target domain.
The second article, by Hans-Helmutt Nagel, is another example of the long-term effort to construct vision systems for a known hard problem. More than 20 years ago, Nagel initiated work on interpretation of traffic and dynamics in a scene. In vision, his Hamburg taxi sequence is still considered a landmark in motion interpretation. …