Research On Multi-view Information Extraction Technology For Target Reconstruction And Measurement

The imaging process is a two-dimensional projection of three-dimensional world information,which is information about the cropping and two-dimensional degradation of the objective world.In the process of visual measurement and remote sensing mapping,single view is often difficult to meet actual needs.It is limited by the limited field of view of a single viewing angle and the loss of three-dimensional information when the target scene is projected onto a two-dimensional plane.Multiple images from different perspectives contain far more information than a single view.By extracting and merging multi-view information,it is possible to reconstruct a planar image covering a large range of scenes,and also reconstruct a three-dimensional target area to be observed.In order to use the multi-view information extraction technology to achieve the reconstruction and measurement of the target,the basic problem to be solved is the registration relationship between multi-view images or the determination of the relative pose relationship between cameras at various angles.At the same time,there must be effective information extraction methods to obtain the required reconstruction and measurement results.In this paper,the splicing of multiple strip images,the measurement of three-dimensional surface parameters of large-sized targets,and the estimation of the volume of the collected lunar soil are used as background.Aiming at the problem of large-field image mosaic and 3D reconstruction measurement,this paper designs multi-view sequence image registration,camera orientation,3D point cloud registration optimization method for multi-view sequence images,rapid fusion algorithm for image stitching and 3D surface variation measurement.The innovations are as follows:(1)A grid-based global optimization method for image registration is proposed.In the process of registration,the problem of uneven distribution of image feature points may lead to errors in the calculation of the homography matrix,and the accumulation of errors may cause obvious misalignment of the stitching joints.The grid-based image registration method designed in this paper achieves the effective suppression of the accumulated error of multiple long-sequence aerial image images with uneven texture distribution by controlling the distribution of feature points participating in the optimization of the image on the image,and designing the global optimization function using the matching link.Then a method of image fusion based on background line is proposed.In order to realize the real-time splicing of the backhaul image of the aerial flight platform,a fast image fusion method is designed based on the image fusion idea of low frequency skeleton information combined with high frequency detail information.The detail information on both sides of the seam is retained,and the misalignment caused by the registration error is suppressed by the curve fitting.The splicing method is suitable for real-time splicing of near-plane scenes or fast and effective splicing of multiple long-sequence images under the condition that the aerial image acquisition platform cannot provide accurate real-time position information.(2)Two large-scale target surface parameter multi-view photogrammetry methods based on single camera and fixed dual camera were designed.In the single-camera multi-view mark point cloud reconstruction,the initial point pair selection method is designed to optimize the camera orientation accuracy.Using the dual camera as the image acquisition device,the high-precision orientation method of the camera with multi-camera multi-view image is designed,and the relative positional relationship of the fixed camera is introduced into the optimization process.Compared with the single-camera multi-view photogrammetry method,the dual-camera multi-view photogrammetry method is more suitable for application scenarios where it is inconvenient to arrange the scale reference near the target to be measured,such as the measurement of the top wall settlement of the subway tunnel construction site.The single-camera multi-view photogrammetry method designed in this paper can complete image acquisition and measurement with only one camera acquisition,and the measurement task can be executed at a lower cost,and its accuracy is on the same order of magnitude as dual camera multiview measurements.(3)A method for measuring the volume of lunar soil collected by lunar detector based on multi-view dense reconstruction and point cloud comparison is designed.Because the lunar detector is inconvenient to install an additional lunar soil volume measurement device,the multi-view sequence image acquired by a single surveillance camera attached to the acquisition arm is used to complete the three-dimensional reconstruction of the acquisition area.A point cloud registration method based on two-way alignment verification of reconstructed source image feature links is designed.The registration effect is better than the registration method based on 3D point cloud features.A method for measuring the terrain fluctuation based on point cloud comparison is designed and avoid additional errors introduced by point cloud surface meshing.The experimental results show that the accuracy of the proposed method meets the requirements and can be used to estimate the soil collection of the lunar detector.Multiple reconstruction and measurement tasks have been taken as the research background,extracts multiview information as an intrinsic core,and focuses on the optimization of large field of view stitching and high-precision acquisition of 3D information on target surface.The experimental verification algorithm is stable and reliable,and the results meet the actual needs.