RGB-D Based Elastic Fusion Real Time 3D Reconstruction Algorithm Optimization

With the continuous progress and development of modern computer vision,3D reconstruction technology has attracted widespread attention from academia and industry.It has a very wide range of applications in the fields of model defect detection,intelligent robot vision,3D printing,etc.,especially for indoor scenes.Real-time dense and high-quality 3D reconstruction is the focus of attention in the fields of robotics and augmented reality.Three-dimensional reconstruction algorithms can be roughly divided into threedimensional reconstruction of binocular stereo vision and real-time three-dimensional reconstruction based on RGB-D sensor.The 3D reconstruction based on binocular stereo vision generally calculates the depth information of the object through multiple viewing angles of observation data frames and the disparity information between the frames,and then reconstructs the corresponding 3D model.This type of algorithm cannot meet the real-time needs.It also cannot solve the closed loop problem in the 3D reconstruction process.Although the 3D reconstruction based on Kinect sensor can meet the real-time requirements,it can only be reconstructed in a small area and cannot solve the closed-loop problem.The current point cloud model based on the real-time 3D reconstruction technology of RGB-D sensor contains noise,poor accuracy of the reconstructed model and local misalignment.There are two main factors that affect the quality of real-time 3D reconstruction,one is the accuracy of the pose solved by the point cloud registration algorithm,and the other is the accuracy of the key frame determination and registration in the loop detection process.Therefore,to improve the quality of the real-time 3D reconstruction model,we need to start from the above two aspects.Based on the elastic fusion algorithm,this paper analyzes in detail the joint pose tracking,elastic node and random fern real-time database based on point cloud and color images,and proposes a weight-adaptive pose model and a redefined random fern database model.Compared with the previous elastic fusion algorithm,the work of this paper not only improves the accuracy and stability of the pose,but also effectively improves the inconsistency of facets in partial scene reconstruction.The main work content and innovations of this article mainly include the following two aspects:(1)The optimization of the adaptive weight model of the camera pose.The model used in this paper to solve the camera pose is the joint tracking calculation of deep point cloud ICP registration and color image RGB registration.To this end,first calculate the depth error,and use the optimized ICP algorithm to perform point cloud registration;secondly,calculate the photometric error,use RGB registration to minimize the distance between the corresponding pixels;finally,use the weights after ICP and RGB registration Than adaptively adjust and jointly solve the final camera pose transformation matrix.The optimization model also considers the distribution of depth images and color images between two adjacent frames,making the solution of pose parameters more reasonable.(2)Random fern real-time database model optimization.Aiming at the loop detection problem,this paper proposes a new random fern database model based on the random fern realtime database model.The optimization model solves the key frames and registers the key frames by redesigning the decision function of the key frames and introducing new domain parameters.When a loop occurs,use the deformation map to optimize adjustments and update the scene.The optimized model can make the judgment and registration of key frames more accurate and effective.Experiments show that,compared with the original elastic fusion algorithm,the optimization algorithm proposed in this paper has a very stable effect in the reconstruction of general indoor 3D scenes,even in complex scenes and scenes where the camera moves slightly faster.Higher accuracy and stability.This algorithm can not only improve the accuracy of the pose and the accuracy of the reconstructed panel to a certain extent,but also repair the inconsistencies in the local model,making the reconstructed 3D model more in line with the real scene.