提出了一种基于角度纹理特征及剖面匹配相结合的高分辨率遥感影像带状道路半自动提取方法.该方法由用户输入道路起点、初始方向及宽度,使用角度纹理特征模型预测初始的道路中线点,以抛物线方程参数构建道路中线轨迹参数模型.使用计算曲率变化的方法验证道路轨迹点,对验证失败的中线点位使用剖面匹配算法进行重新预测并确定,最终提取出该道路中线轨迹.本文使用Visual C 构建了原型系统,对QuickBird及IKONOS影像中具有一定宽度的带状道路进行了提取试验,并与经典的基于剖面匹配的半自动道路提取算法和基于Snakes的半自动道路提取算法进行了对比试验.经试验验证,本算法取得了较为理想的结果.

In this paper, we propose a novel semi-automatic ribbon road extraction schemewhich combines angular tex-ture signature and profile matching. The angular texture signature was proposed byHaverkamp in 2002 for extracting straight road structure from IKONOS satellite mi agery. In thismethod, a rectangular template region rotates around a spe-cific pointwith certain angle intervals, and computes texture signature ofallpoints in this template region ateach position of rotation. The basic texture signature is variance and the entropy also to be tested in this paper. A profilemeans that in the centerline points, perpendicular to the direction of the road, we painta straight line segment, and extractgray values of various points along the line segment. And the straight line segmentshould be longer than the roadwidth. Profilematc-hing calculates the least square values between profile line model and the profile in the predicted points. To a certain extent, we offsetthe profile in the predicted pointofand calculate valuesofatthe leastsquare each offsetposition, we be-lieve the corresponding offsetwith theminmi um value should be possible pointofroad centerline, therefore, the final road centerline can be determined through predicted points and the offset least square values.In the scheme, user inputs the initial position, direction,and width of the road firstly. The initial road centerline points are predicted based on angular texture signature,and road center trajectorymodel is created using parameters of parabola. Themultiple linear regressionmethod is used for acquiring the road trajectory parameters, andwe utilize these parameters and the initial centerline points coordinates to compute the initialmean of the curvature. Thenwemake use of themethod of curvature change to verify the points in trajectory. By comparing the value and the given threshold,we re got following conclusion: if it exceeds the lmi it,the verification fails, a profilematching algorithm willbe used to re-predict and decide the position of centerpoints again;if the value doesn texceed the lmi it, the verification succeeds, we believe the centerline points predicted in angular texture signature is the finalposition. In the algorithm, themanual intervention is necessary especially for complex situation. After the verification, we can achieve the final centerline points and add them to the trajectory. If these two algorithms both fai,l we could decide the final centerline point completelymanually.In thispaper, we mi plementthe abovementioned scheme and build a prototype system usingVisualC++, and veri-fy the extraction resultofribbon roadswith appropriatewidth inQuickBird and IKONOS mi agery. In the expermi ent1, the road hasuniform texture, and the contrasttobackground isobvious, butno obviousobstructionson the road. In this situa-tion, the scheme extracts thewhole road automaticallywithout anymanual intervention. In the expermi ent2, the ribbon road extraction from IKONOS is generally successful and need a fewmanual intervention. In the expermi ent3,a rather complex situationwith several roads in an digitalaerial mi age, we extract themain roads automatically and some junctions of the roads needs semi-automatic extraction, but the final result is satisfactory. We also utilize expermi ents to compare our schemewith traditionalprofilematching semi-automatic extraction algorithm and Snakes algorithm. After these experi-mental verification, itproves this semi-automatic ribbon road extraction scheme can achieve good results.