Catalogue of Artificial Intelligence Techniques
Author(s): Andrew Fitzgibbon
Active Vision is the philosophy that many robotics vision problems are greatly simplified if a robot is allowed to collect and analyse a sequence of images as it undergoes some appropriately chosen motion. Examples of such problems include Structure from Motion, Stereopsis, object segmentation and low-level reactive behaviours such as tracking and collision avoidance. Active vision systems are generally characterised by their reliance on fast real-time hardware, relatively simple image-processing algorithms, and little or no calibration. Position control is often by visual servoing: converting image movement directly into motor signals, and the systems usually retain only a small amount of internal state (for example, the mean and covariance representations of the Kalman filter). This branch of computer vision research may be seen as a reaction to approaches that concentrate on single, static images of scenes and insist on a full interpretation of the whole picture. Active systems generally perform impressively on tasks which have proved difficult for the more classical approach, but have not yet scaled to more complex problems such as scene understanding.
- Special Issue on Active Vision International Journal of Computer Vision 1 (1993).