| With the development of automatic control and camera shooting,the values of UAV technology in national defense security,industrial production and consumer markets are widely recognized.In the area of??national defense security,UAVs can collect a wide range of data that ignores the impact of the terrain on patrols and are commonly used in such fields as fire early warning and resources monitoring.In industrial production,UAVs can quickly collect geographic information about a certain area and track the safety of the plant in real time.The widespread popularity of UAV also means that the number of aerial images has risen sharply.How to mine effective information from massive aerial images through computer vision and machine learning technologies has become the key to the further development of aerial images.The ways of using computer vision to obtain advanced semantic information from images include object detection,object recognition and image segmentation.Image segmentation refers to the technique of dividing the en-tire image into regions with different semantic meanings,and the pixel level classification and calibration of the image.Compared with the object detection,image segmentation is more suitable for the calibration of meshed objects like roads,rivers,forests.Without loss of generality,this paper discusses the issue of road segmentation in detail.First of all,it is a prerequisite for monitoring road congestion and abnormal con-ditions and has very high practical value.In addition,some of the methods used for road segmentation also apply to other object segmentation tasks.This paper summarizes the road segmentation algorithms based on aerial and remote sensing images and analyzes the advantages and disadvantages of different algorithms.After that,we introduce the tech-niques and methods we use in this article,including super-pixel,XGBoost classification model,high-order conditional random fields and Robust P~nPotts model.Since road seg-mentation is a pixel-level annotation problem,the super-pixel pre-processing can ensure accuracy while improving the robustness of the algorithm and computational efficiency.XGBoost is currently the most popular structured data classifier.It has had a significant impact in industry and academia.Conditional random fields are one of the mainstream algorithms for image segmentation.Compared with the binary conditional random field,the higher-order conditional random field can get the segmentation results with smooth edges and more complete details preservation.Finally,we design a road segmentation method based on XGBoost and P~n Potts models.First,we partition the road image based on SLIC algorithm.The posterior probability output by XGBoost is used as the unary potential of the conditional random field,and the gradient of the posterior probability of the adjacent pixels is used as the binary potential.By random sampling and the shortest path method,the road likelihood cluster is obtained as a high-order potential.Finally,the robust P~n Potts model is used to model and solve the conditional random field.Finally show the results of the run and analyze the parameter sensitivity. |
