Perceptual Strategies For Purposive Vision


Garvey, T. D. (1976). Perceptual strategies for purposive vision.


This report describes a computer program that approaches perception as a problem-solving task. The system uses information about the appearances of objects, about their interrelationships, and about available sensors to produce a plan for locating specified objects in images of room scenes. The strategies produced allow for cost-effective processing by utilizing “cheap’’ features in the early stages, reserving more complex operations for later in the process, when the content has been sufficiently restricted. The general strategy paradigm used by the system is to acquire image samples expected to belong to the target object; validate the hypothesis that the acquired samples do belong to the target; and finally to bound the image of the object by outlining it in a display picture. Sensors used by the system include a vidicon with three color filters and a master-scanned, laser-rangefinder capable of producing a “range image.’’ In addition, the range-finder measures the reflectivity at the laser wavelength, at each post, producing a gray-scale image in perfect registration with the range image. The primitive attributes of brightness, line, saturation, height, and local surface orientation at specified range locations can be computed from these images. These attributes represent the new data available to the system for recognizing and identifying objects. Object descriptions are provided initially by indicating the object to the system, and allowing the system to measure attributes. Other data, such as typical object relationships, are provided interactively by the user. When required to locate an object, the system computes a distinguishing features representation that will serve to separate parts of the target from those of other objects. These distinguishing features are combined into a strategy represented as a planning graph. This graph contains optimal subgoals for achieving the goal of locating the object. An execution routine selects the “best’’ subgoal, executes it, rates its effect, and selects the next best goal, continuing with the progress until either the object is located, or there are no options remaining. This approach offers several contributions to perception research. By capitalizing on the goal-directed aspects of the problem, the system is able to select relevant information from a mass of irrelevant data. The system is able to organize its processing in such a way as to optimize the use of sensor data. By generating strategies when needed, the program allows for the easy introduction of new objects and new sensors. The system allows for the logical introduction of new information deriving algorithms.

Read more from SRI