Research On Key Technologies Of 3D Reconstruction Of UAV Panoramic Videos

The trend of smart cities affects various industries in society,the precise expression of spatial information for cities and surface objects can provide effective auxiliary services for management and decision-making in various fields.The threedimensional(3D)model based on the real-scene sensors can realistically and objectively express the structure and appearance of the target objects,and provide accurate geographic reference information.The core problem of 3D scene reconstruction is to recover the 3D structure information of scenes from inputted images.The coverage of the scene and objects directly affects the accuracy and effect of 3D reconstruction.In the current strategy of 3D reconstruction using binocular/multi-view images,even professional five-lens oblique photogrammetry may also have challenges such as limited fields of view,limited overlaps,and mutual occlusions of scene objects.At present,the panoramic cameras widely used in augmented reality AR image the scene in a continuous 360 ° full-view and high-definition video stream,and increase the multiview coverage of scene and object in two dimensions from a larger angle of view in space and a higher density in time,which provides a new possibility for 3D scene reconstruction.This paper manages to utilize aerial panoramic video for 3D reconstruction,and researches on the problems of over-overlaps of video panoramic image sequences,difficulty in modeling panoramic stereoscopic geometric relationships,and insufficient integrity of multi-view dense matching.Firstly,in view of the redundancy of panoramic video sequences,a key frame extraction method for UAV panoramic video is proposed through learning key frame concepts and strategies used in image retrieval and surveillance video technologies.The method is based on carrier motion state identification and baseline analysis,which reduces data redundancy while ensuring sufficient matching and intersection between key frames.The experimental results prove that the method in this paper can effectively extract the key frame sequence of the scene panoramic videos.Compared with the violent frame extraction method,the redundant frames are removed and the appropriate baseline is guaranteed,and the scene3 D model is constructed in a shorter time.Then,a strategy of panoramic photogrammetry with the concept of panoramic multi-view projection is proposed,the image pair sequence suitable for multi-view photogrammetry processing can be obtained by designing a panoramic multi-view re-projection model.The panoramic multi-view reprojection meets the advantages of stereo vision baseline,intersection angle,and image overlapping degree,and realizes the photogrammetric processing of panoramic images.The experiment proves that the strategy and model proposed in this paper can effectively perform panoramic 3D reconstruction,and obtain high matching accuracy between multi-view images.Reliable positioning accuracy and measurement accuracy can be obtained with the help of control points in the adjustment optimization.The strategy based on panoramic multi-view reprojection can reconstruct a more refined effect in a shorter time compared with using panoramas directly.Finally,for the problem of the incompleteness of the depth map and point cloud reconstruction integrity of Colmap multi-view dense matching algorithm,the theoretical basis and specific algorithm implementation are studied,and locate the key steps that affect the reconstruction integrity.The degree of influence of parameter sets corresponding to different steps are analyzed through detailed experiments,and a more complete depth maps and dense point cloud results are reconstructed in a reasonable time through parameter relaxation.