Research On Dense 3D Point Cloud Reconstruction Based On Sequence Image

In recent years,computer vision technology has developed rapidly.How to use sequence images to efficiently and accurately reconstruct the 3D structure of the main part of image has become one of the research hotspots.The 3D reconstruction technology of realistic scene has broad application prospects in the fields of object detection and recognition,digital city construction,medical diagnosis,terrain exploration,virtual reality and augmented reality.Compared to other active or passive methods for 3D information acquiring,image-based reconstruction technology has an advantage of simple implement,flexible operation,low cost,and strong adaptability to the environment.With the popularity of the electronic devices such as mobile phones and tablets,people are paying more and more attention to the technology of capturing images with a handheld device and completing 3D reconstruction within it.In this paper,we study the sequence images taken from different perspectives of the same target,and realize the dense 3D point cloud reconstruction from the sequence images.The main research contents are listed as follow:1.At first,we analyze the basic principle of camera imaging,and introduce the mutual conversion relationship between the four coordinate systems in the camera imaging process.In addition,through the analysis of epipolar geometry,we introduced the internal and external parameters of the camera used in the 3D reconstruction process and given a specific method.2.After selecting the SIFT feature points to achieve image registration with more accurate positioning and higher matching accuracy,we use the incremental sparse reconstruction algorithm SFM to restore the camera pose and reconstruct the sparse point cloud through triangulation.In the process of incrementally adding images and reconstructing,each time an image is added,the LM algorithm can optimize the beam using adjustment method,in this iteratively way if adding images the optimal reconstruction result can be ensured.3.Compared with the existing feature point matching method,we introduced a doubleconstrained feature point matching algorithm based on grid correspondence.First,we initialize the screening of the matching completed by analyzing the matching constraints in the corresponding grid under different thresholds.Then we introduce the random sample consensus(RANSAC)to verify the matching performed by using the epipolar constraint to eliminate the wrong matches.Experiments show that the algorithm improving the matching quantity and precision without increase running time.4.At last,we research the algorithm of dense 3D point cloud reconstruction in the aspect of sequencing image after the camera pose has been restored and introduce a dense reconstruction algorithm based on PMVS.Under the premise that the fundamental matrix is known,the image can be re-matched.At this time,multiple feature points can be extracted and matched simultaneously,and the number of spatial seed points can be increased as much as possible,while in the weak texture region,the feature points are sparse or even not distributed,we use the strategy of seed patch expansion to reconstruct.After three times of expansions and filtering,the algorithm completes the dense 3D point cloud reconstruction of the sequence image.