Research On Road Object Recognition And Extraction From High Resolution Remote Sensing Imagery

Attention has been paid extensively to obtain information of the earth's surface from High Resolution Remote Sensing Imagery(HRRSI)due to its economy and efficiency,which has become a research hotspot in remote sensing image information extraction.Road map is an important geographic information resource.Its accuracy and timeliness not only directly affect the effectiveness of spatial decision,but also have significant meaning with respect to promoting the precision of image registration,information fusion and change detection.The proposed methods and related algorithms on extracting road from HRRSI in presently available literature have some limitations in terms of applicability,accuracy and practicability.Against the above limitations,this dissertation deeply studies such key technologies as road element rough extraction,road element precise extraction,road network connection,intersection identification and centerline extraction so as to explore novel methods and ways of road object recognition and extraction.The main contents and achievements are as follows.(1)A road element rough extraction method based on multi-kernel learning and multi-feature fusing is designed after analyzing the main features of HRRSI.This method optimizes the weights of different features in the training stage by multi-kernel learning framework,and achieves effective fusion of spectrum,texture and direction information,which greatly improves the accuracy of road elements recognition.(2)A novel method for road element precise extraction is designed based on the analysis of road shape features and geometric features.This method combines such features as the slenderness of road shape,the compactness of ground objects,and the area of surroundings to build road shape indexes for automatically filtering out the interference of non-road noises.A series of morphological operations are also carried out to tackle such issues as holes in some road sections,loose connection between different pixels,and incomplete structures.(3)By integrating priori knowledge and topological characteristics of road network,a knowledge-based road element connection algorithm is presented which establishes a global road connection model and implements topological connections of road network by constructing a penalty factor and a penalty function on connecting road elements.(4)An automatic intersection recognition method combining global and local detection is proposed after analyzing the characteristics of different types of intersections.Intersection candidate points are obtained globally under the road skeleton overall constraint condition,and intersections are then detected locally through the divided rectangular block rotation model proposed in this dissertation.Experiments show that the presented method can greatly improves the accuracy of intersection recognition.(5)To overcome the drawbacks of "spurs" and "deviation" caused by commonly-used centerline extraction methods,a new centerline extraction method has been put forward based on linear structure enhancement and multivariate regression.Firstly,a set of multi-scale polynomial filters is employed to convolve the road binary skeleton to obtain a pixel set with linear structure so as to smoothen the irregular road skeleton.Then the road-centerline-extraction problem is converted into a regression problem which has been solved by utilizing self-adaptive multivariate spline regression.Experiments show that the proposed method can effectively extract smooth and accurate centerlines from irregular road skeletons.