Research On Automatic Road Extraction Based On DPM And Level Set Evolution

Road information is an essential part in people’s daily life.Recently,with the development of science and technology as well as the growth in living standard,various applications related to road information ask for more efficient and accurate road extraction technology,such as navigation system and traffic management system.At the same time,the development of satellite communications improves the resolution of remote sensing images,which promotes the road extraction technology to become one of the hot issues in computer vision.This thesis studies road extraction technology in remote sensing images,proposes an automatic road extraction method based on DPM and level set evolution,and focuses on how to realize automation in level set evolution for road extraction as well as how to improve efficiency of the algorithm.The main work is as follows:1)This thesis makes improvements on traditional HOG feature and proposes a new image feature called HOG_Rot which can change regularly with the rotation of the object.Furthermore,an algorithm for judging rotation angle of an object was proposed.This algorithm judges the rotation angle of an object with respect to the controlled one by extracting their HOG_Rotfeatures.By this way,only horizontal road fragment samples are needed in training process of DPM,and in testing process,all the road fragment samples were rotated to horizontal direction to accomplish road fragment detection.If we don’t do this,road fragments should be classified to a number of directions,that way,we should run training and testing processes for each direction,which causes time consuming.So this improvement decreases the training time of DPM and improves the efficiency of the whole algorithm.2)This thesis introduces DPM to automatically initialize the zero level curve in level set evolution,for which reason we difine our method as automatic.In the past research,most level set methods are semi-automatic,zero level curves are initialized manually.This thesis introduces DPM to replace traditional manual operation.First,the whole remote sensing road image is cut into a number of small size patches,and the trained DPM model is used for detecting road part in the patches which is called road fragments labeled with a rectangle curve.Then,the rectangle curves are regarded as the rough initial zero level curves in level set method,after removing wrong ones,the final initial zero level curves are acquired.This method takes place of the traditional manual operation,and realizes the automation of road extraction in remote sensing images.To demonstrate the effectiveness of the above methods,experiments were conducted on real high resolution remote sensing images.This tesis selects 5 different kinds of road images as testing set,of which the number is 75 in total.Using HOG_Rot feature,71 of them were successfully rotated to horizontal direction,the accuracy is 94.7%.Finally,the road extraction results of two different images were given visually.And we compare our method with traditional semi-automatic method in iteration times and time consuming,the results show that our method is more effective.t.Meanwhlie,correctness,completeness and overall quality are employed to evaluate our method,and the results show that the method this thesis proposed can automatically extract road information from high resolution remote sensing images with high efficiency and accuracy.