Road Detection Based On Aerial Image Of Drone

With the development of unmanned aerial vehicle technology,unmanned aerial vehicle has important practical value in military reconnaissance,rescue,geological exploration,transportation and pipeline inspection.In these applications,the visual sensor has been widely used by various UAVs because of its light weight,low power consumption,good concealment,and rich information.Using visual sensors to enhance its ability of autonomous navigation and scene understanding is a promising research direction about UAV,in which road detection is an important basis for visual navigation and scene understanding.For example,road information can provide important information for object detection and tracking when UAV is used in target reconnaissance.In this paper,UAVs are selected as the technology platform for acquiring aerial image data.Based on previous existing work,we focus on lane and road detection.After analyzing and summarizing the functions and limitations of the existing unmanned aerial images,this paper analyzes the road detection problem based on the unmanned aerial image.Availble technologies and solutions are proposed and simulations are executed to solve problems about road detection method from UAV images.The main contributions of this paper are as follows:(1)A simple method of detecting the lane line based on Hough transform is proposed in this paper.Firstly,the aerial image is preprocessed,and the edge image is obtained by changing the gray level near the roadway line.Then the Hough transform is used to extract the area of the lane line.Finally,based on the result of the test of the candidate of the lane line,this paper adds geometric verification and RANSAC line fitting method to further extract the lane line.Experiments show that this method can achieve better results in different environments.(2)A visual road detection method based on stroke width transformation based the width consistence of roads is proposed.Firstly,we use the stroke widthtransformation algorithm to calculate the road width and improve the SWT performance to make the road width image more accurate.Then,we extract the road shape and color features to find out the category that contains the road area automatically.Finally,the interference and burr in the road area category is removed to achieve accurate detection of the road.(3)Another road detector which based on superpixel clustering is proposed in the paper to deal with some complex scenes which the SWT detector does not work well.To be specific,our method uses two different schemes to generate proposal regions,which fully utilizes the effectiveness of the shape feature of the road and is good at dealing with complex scene.A random forest classifier is applied to select those regions which are long and consistent width from the proposal regions.Then the road segments are grouped and used for inferring the color distribution of different roads which coexist in a single image.Finally,a global graph based labeling model is applied to remove false positives from background clutters and to yield final detection results.