Semi-automatic extraction of road networks by least squares interlaced template matching in urban areas

Abstract

Semi-automatic extraction of roads is greatly needed to accelerate the acquisition and updating of geodata. In fact, most roads are often seriously impacted by various types of noise, such as the occlusion of vehicles and the shadow of trees on very-high-resolution (VHR) remotely sensed imagery, which makes most of the existing road trackers ineffective. Fortunately, lane markings are less frequently disturbed than other parts of the road surface, and they provide a unique clue for road tracking on VHR images. In this paper, a semi-automatic method is proposed to extract roads with lane markings in urban areas. First, an operator detects a road segment and selects three seed points, which indicate the starting point, the direction and the width of a road, and lane markings near the seed points are automatically detected. Subsequently, an interlaced reference template of the selected road, composed of a few profiles of the road surface and the detected rectangular templates of lane markings on the road surface, is constructed. The reference template is then convolved with the image, and least squares template matching is employed to track the road axis. To complete the task, the process is then repeated. Various types of images are used for test, and the results show that our method is capable of robustly extracting roads from VHR imagery because the special configuration of the reference template can decrease side effects such as the occlusion of vehicles and the shadow of trees as much as possible.