A template-matching based approach for extraction of roads from very high-resolution remotely sensed imagery

摘要

Automated or semi-automated extraction of road networks is a prerequisite to fast acquisition and update of geospatial data. Actually, line-shaped lane markings and/or median strips on road surfaces are less impacted by occlusions of vehicles or shadows of trees than the other parts of road surfaces on very high-resolution (VHR) remotely sensed imagery. These features provide a clue for road extraction. This article proposes an approach for semi-automated extraction of road networks by tracking apparent lane markings and/or median strips on VHR imagery. After preprocessing the raw image, three seed points on a short road segment are manually selected, which indicate starting point, direction and width of the road, respectively. Based on the manually selected data, a reference template of the road, which is composed of two components: a cross-section profile, rectangular templates of lane markings and median strips, is created. With the created reference template, automated road tracking is triggered. During the process of road tracking, a least-squares template matching is employed to search the optimal road centreline points, and a human operator is retained in the loop to guide the computer. The above operation is repeated until an entire road network is completely extracted. Tests of the above-proposed method are conducted on both aerial and VHR satellite imagery. The results show that the proposed method can successfully track over 94% of the highways and 81% of the arterial roads from the VHR images, and save the time of 26% when comparing to traditional methods.