Object detection using perceptual organization and prediction/verification of hypotheses

Abstract

The authors describe an approach to perceptual organization and prediction/verification of hypotheses for detecting a polyhedral 3-D object in indoor scenes. This method is a sequence of different stages which allow a reduction in the space of research and to give semantic significance to each primitive at different levels of representation. Since in indoor environments most of the objects are man-made, it appears that their morphological characteristics are essentially linear and regular, and for these reasons straight line segments are the best primitives available in order to detect such objects. From these primitives, geometric relations allow to obtain different classes of pair of segments which represent a second level in their algorithm. Next a method based on prediction and verification of hypotheses with score assignment is performed in order to create the third level, constituted by different simple geometric shapes as squares, rectangles, triangles and other polygons. Last stage is dedicated to the grouping of these different shapes by perceptual organization for finding complex objects in the scene. After a more detailed description of their algorithm, the authors prove that the method is both efficient and robust in the presence of noise like occlusion. Finally, experimental results obtained in the laboratory are shown in order to demonstrate the method's effectiveness.