Research On Unstructured Road Change Detection Algorithm Based Remote Sensing Registration

Unstructured roads occupy a dominant position in urban and rural roads.Detecting their change information is an indispensable key content in natural disaster assessment,geographic data updating,accident risk detection,road damage assessment,urban planning and dynamic detection of military targets.However,the unstructured roads generally do not have clear lane lines and road boundaries,and are susceptible to environmental factors such as illumination,shadows,weather changes,and etc.Therefore,it is becoming more and more difficult to detect their change information.In this work,a novel framework for unstructured road change detection using small unmanned aerial vehicle(SUAV)remote sensing image registration method is proposed,which plays an important role in transformig multi-temporal images into one coordinate system and determines the effectiveness of the change detection for unstructured road.The proposed framework includes the following contributions.(i)Image enhancement and image clustering are used for image preprocessing to reduce the influence of external factors such as atmosphere,solar altitude angle,UAV flight angle,etc.(ii)A multi-scale deep convolutional feature descriptor using layers from a pretrained visual geometry group(VGG)network is generated and combine geometric structural shape context feature to form complementary features for extracting road features.In addition,VGG-16 network is used to extract feature points for obtaining uniformly distributed feature points in the image,thus reducing the impact of redundant points on matching accuracy.(iii)A strategy of dynamically adjusting the number of inliers is adopted to maximize the potential information of the image for improving the robustness in point set registration process.(iv)The global and local geometric constraints are introduced to constrain the cost of image conversion and the feature structure of feature set in the process of feature point set registration,and to prevent the ill-posed problem from mismatching in the process of image transformation.(v)A framework of mixed feature descriptor is constructed and can adjust the weight between the two features arbitrarily to steady the road feature description.Three-step experimental results for SUAV(the DJI Phantom 4 Pro)images from different viewpoints and temporals have verified the robustness and accuracy of the proposed framework,demonstrating its capability in unstructured road images.This paper also compare with five state-of-the-art image registration methods and two state-of-the-art change detection methods,where the proposed method shows favorable performances in most scenarios.