Extracting Road Information Based On Unmanned Aerial Vehicle Video Analysis

Country disaster rescue is becoming more and more important and it requires a rapid response for disaster rescue.The key component for disaster rescue is to plan the optimal rescue path.Our work has two parts,one is extracting road from the video which is taken by UVA,the other is giving the geolocalization of the road which is damaged in disaster.As for the part of extracting road,Traditionally,the optimal rescue path seriously relies on the recognition on images of the damaged areas and the corresponding recognition algorithms are proposed for analyzing satellite images.However,due to its low updating frequency satellite images are not suitable for disaster rescue.Therefore,unmanned aerial vehicle is a good alternative approach to acquire real time images on damaged areas.Techniques are then needed to recognize the UAV images.To cope with this situation,we first extract UAV videos to images,segments these images into fixed size pieces,and manually labeled these data.We then study whether the conventional methods such as mathematical morphology,Hough transform and P-value segmentation approaches can be used to extract roads from UAV images.At last,we propose to adopt SVM and combine it with GA to improve the performance of this approach.Empirical studies are performed on data sets both collected by us and collected from the Internet.Experimental results demonstrate that our approach works well on these data sets when compared with conventional approaches.As for the part of geo-localization,most methods predict the location of the image by matching the images with known locations,such as ground-level images.Nevertheless,most of the Earth does not have the image with known locations.In this work,we localized a query image which has obstacle in road by matching the reference image in our database.We use our dataset to learn a feature representation in which matching views are near one anther and mismatched views are far apart.In order to find a good approach to extract the feature,we compared SIFT with Alex Net.And the result of Alex Net is better than SIFT.Therefore,we use Alex Net to extract feature at last.Based on the proposed method,in view of its important position,it is of great theoretical and practical significance to study robust and accurate road information extraction in video.