Research On 3D Reconstruction Technology Based On Aerial Sequence Images

Image-based 3D reconstruction is an important step for computer vision from 2D to 3D,and it is the inevitable trend of computer vision development.It has the characteristics of low sensor price,non-contact,wide measurement range,rich reconstruction details and so on.It has become the focus and difficulty of 3D reconstruction research in recent years.Aerial photographing is an important means to obtain large-scale outdoor scene image data,and its corresponding 3D reconstruction method is of great practical significance in the field of digital city and topographic mapping.In the stage of image feature detection and matching,repeated textures in the scene will lead to a large number of mismatches.According to the moving characteristics of pixels in aerial sequence images,this paper proposes a matching outer point elimination strategy based on motion consistency.The algorithm takes the line slope of the matching point pair as the observation object,and references the 3 ? criterion in the normal distribution,so as to build the confidence interval of the slope value and eliminate the mismatching point pair outside the interval.In this paper,test data are designed and collected to verify the improved effect.The results show that the selected matching points can provide more accurate pose estimation.When the scene is square,grassland,etc.,the number of feature points is small and vulnerable to noise.For this kind of weak texture scene,this paper extracts segment features based on point features as a supplement.In this paper,a global geometric constraint is proposed,which takes the angle difference of matching line segment,the slope of midpoint line and the shortest length as the observation object,and uses RANSAC algorithm to eliminate mismatched line segment.Experimental results show that more accurate and reliable line feature matching relationship can be obtained by this constraint.In this paper,the existing algorithms are integrated.The reconstruction from sequence image to grid model mainly includes three modules: sparse point cloud reconstruction,dense point cloud reconstruction and surface grid reconstruction.Sparse point cloud reconstruction uses incremental motion to restore structure method;dense point cloud reconstruction uses depth map fusion method;surface mesh reconstruction uses implicit function method.In this paper,a complete description of the algorithm for each module is given,and relevant experiments and analysis are carried out.Through the UAV route planning and aerial image acquisition of the scene,according to the 3D reconstruction system realized in this paper,the 3D reconstruction of the campus building is completed.