Road detection in aerial images plays a key role in visual navigation and scene understanding for unmanned aerial vehicles(UAV). In general,the color of roads is the most straightforward feature for automatic road detection in aerial images. How?ever,it is difficult to model color distribution of roads without some kind of color priors due to the complexity and variation of real scenarios. However,the shape feature of roads shows less variation since a road usually appears consistent width. In this paper,we propose a road detection method using improved stroke width transform,an robust shape feature extractor,to estimate the stroke width of each pixel in aerial images. Then color and extracted width feature are applied jointly to infer road area by using K-means clustering. Finally,the interferences and noises inside the road area are removed to yield final road detection results. Experiments on typical public datasets demonstrate that our approach can detect roads with higher accuracy compared with several conventional methods.%航拍图像的道路检测在无人飞行器导航和场景理解中起着重要的作用.道路的颜色、形状是最直接的道路特征,对颜色特征而言,只有得到准确的道路颜色模型,才能较好地实现道路的分割,但道路的复杂性和多样性导致自动获得目标的颜色模型非常困难;相比颜色特征,形状特征则较为稳定,这是因为道路区域有着固定的宽度.鉴于此,论文提出了一种基于笔画宽度变换(SWT)的道路检测方法.该方法利用改进的笔画宽度变换来获得像素笔画宽度这一形状特征;之后使用K均值聚类分离并提取出含有道路区域的类别;最后去除道路区域类别中的干扰和毛刺,实现道路的准确检测.在典型公开数据集上的实验证明论文方法适于检测多种道路,准确性较高.
展开▼
