Road Tracking by Real-Time Support Vector Data Description Classification from Very High Resolution Remotely Sensed Imagery

摘要

Semi-automatic road tracking is a practical way to speed up road mapping from remotely sensed imagery. However, the most of existing road trackers are frequently interrupted when they are employed on very high resolution (VHR) remotely sensed images. A road tracking method is proposed for VHR imagery herein. Initially, a human operator inputs three seed points on a selected road segment and then the road samples are obtained to train the support vector data description (SVDD) classifier. Subsequently, the automatic tracking is run. During the automatic process, the SVDD is employed to classify each sub-window into road class and none-road class, region adjacency graphs (RAG) is used to eliminate the small disturbing features on the road surfaces, and an improve angular texture signature is used to search the optimal road centerline points. Iterate the above automatic process until a whole road is extracted or resort to the help of human beings. Three VHR images were used to test our road tracker. The results show that our method is capable of extracting most of the main roads with an acceptable spatial accuracy, and it is robust to many image noises such as such as road markings, occlusions of vehicles and shadows of trees.