Dynamic 3D Reconstruction Based On RGB-D Image Sequence

With the continuous development of computer vision,3D reconstruction technology has been widely used as its important research branch.In particular,the advent of RGB-D depth cameras provides a new implementation method for the field of 3D reconstruction.Dynamic 3D reconstruction based on RGB-D image sequence refers to the acquisition of RGB images and depth images of the scene taken by depth camera,and the reconstruction of dynamic object model changing with time from the acquired image pair sequence acquiring the shooting.At present,the 3D reconstruction technology of static scenes has become mature,and the 3D reconstruction of dynamic scenes has gradually become the research focus in recent years,and has important practical application value.For example,in robots and augmented reality systems,completing the 3D reconstruction of the dynamic objects in the scene is the basic technology for the perception and understanding of the dynamic targets in the scene,and it is also a necessary prerequisite for human-computer interaction,virtual and reality interaction.In this paper,the dynamic 3D reconstruction algorithm based on RGB-D image sequence is studied in depth,and the main techniques used in the algorithm implementation are discussed,and researches are made in the aspects of enhancing the robustness of the algorithm,improving the operation efficiency and optimizing the reconstruction effect.The main research results of this paper are as follows:(1)A depth map optimization algorithm based on MRF(Markov Random Field)model is proposed.Aiming at the problem that the traditional filtering algorithm could not remove the noise in the depth map well and the reconstructed model had poor accuracy,an improvement was made on the basis of the traditional joint bilateral filtering algorithm,and depth value information was added into the weight coefficient to solve the problem of edge pseudo-depth value.On the basis of the improved algorithm,MRF model is added to comprehensively consider noise,color,edge and other information.The experimental results show that the proposed algorithm is effective in depth image processing.(2)A point cloud registration algorithm based on CTF(Course-to-Fine)is proposed.In order to improve the registration accuracy and registration efficiency,after the rough registration based on FPFH(Fast Point Feature Histograms)features is used for the point cloud data,the CUDA(Compute Unified Device Architecture)accelerated EM-ICP(Expectation Maximization-Iterative Closest Point)algorithm is used for fine registration.Through the test of the experimental data,it can be seen that when using this algorithm for registration,the registration efficiency is greatly improved while ensuring the registration accuracy.(3)A point cloud dynamic fusion and surface reconstruction algorithm based on Surfel model is proposed.In view of the high occupied video memory of TSDF(Truncated Signed Distance Function)model for fusion,Surfel point element model is used for point cloud fusion.The point element model structure is convenient for updating the model according to the captured real-time frame data.At the same time,after the fusion,the point elements in the model are reconstructed by poisson surface reconstruction algorithm.The algorithm is validated on the open data set,and the results show that the proposed algorithm has a good reconstruction effect.