Abstract
One of the goals in computer vision is to interpret scene objects and establish relationships between them. One of tire problems associated with this task is that the image to be interpreted and the objects to be recognised correspond to different levels of information. The image is, on the one hand, represented as a collection of pixels in which three-dimensional information is transformed into two-dimensional one under perspective projection dictated by the camera position as well as photometric parameters such as focal length etc. On the other hand, the object is represented as a collection of three-dimensional structures and relations between them. These rather different representations highlighted the need to construct an intermediate-level representation which can facilitate the accomplishment of the goal of establishing correspondence between image features and scene objects. The complexity of the interpretation task is further compounded by image imperfections caused by lighting, total reflectance, surface markings, accidental viewpoints and so on. The problems highlighted earlier motivated the development of a novel feature grouping framework which takes into account feature stability and the underlying noise. This work advanced the state of the art in perceptual group extraction as the existing techniques tend to be ad hoc. Built upon the framework that we have established we developed the computational representation of higher level features such as junctions, collinear line and parallel line groupings. The low level feature representation and extraction phases of the work were the necessary prerequisites for the extraction of intermediate representations using AI techniques. These representations serve as visual cues in our role-based system (RBS) to classify runways/taxiways in most of the DRA supplied imagery captured from unknown viewpoints. Complexity problems reported in previous work on RBS for low and intermediate level vision tasks are apparently overcome by identifying a set of prioritised feature cues, uncertainties are handled by hypothesis generation and hypothesis verification, and the method can be regarded as a constrained search through the space of candidate hypotheses.