Detection and segmentation of man-made objects in outdoor scenes: concrete bridges

Abstract

We present a new approach for the detection and segmentation of man-made objects in color images of natural scenes. The approach is based on detecting geometric structure in the image and combining the detected structure with color information to guide the segmentation. The central problem is the detection and semantic interpretation of large stationary man-made objects in color images of nonurban scenes. We describe the generation of color-based confidence functions for material selection in incremental segmentation. We focus on the segmentation of images of concrete bridges. These techniques are applicable to autonomous navigation, target acquisition, and several industrial computer vision problems. Large concrete objects often have rectilinear edge structures with many parallel relationships. We use these properties to guide our initial incremental segmentation toward concrete objects. The goal for our segmenter is to locate the representative faces of concrete material in the image as a starting point for the interpretation phase. These heuristics rely on the detection of straight line segments of the edge map of the gray-scale image. The straight line segments, once detected, are then grouped according to several perceptual grouping criteria. The straight line segments are then constrained further by region label restrictions. Finally, color cues are used to restrict the candidate artifacts further and to produce confidence measures of our initial belief in our estimation of the material of the identified rectilinear faces.