Research On 3D Reconstruction Algorithm Based On Spacial Voxel Fusion

With the improvement of modern sciences and technology,three-dimensional reconstruction technology have been widely used in industrial,military and medical fields.The main application areas include three-dimensional model acquisition,virtual simulation,non-contact measurement,scientific computing,mixed reality and battlefield environment perception etc.After several decades of research and exploration,domestic and foreign researchers put forward a variety of three-dimensional reconstruction theory and implementation methods.But for large-scale real-time three-dimensional reconstruction is still a very challenging research topics.For example,in the battlefield environment,we use visual sensors installed on the UAV to capture the target digital image for real-time generation of the target area of high-precision three-dimensional model.In order to achieve large-scale,high-precision and efficient three-dimensional reconstruction process,the reconstruction algorithm needs to be applied to the complex scene robustly.In this paper,the three-dimensional reconstruction algorithm in complex scene is studied deeply for real-time 3D reconstruction of the scene.The main work of this paper includes the following aspects:(1)In order to solve the computational complexity of the traditional layered volume fusion algorithm,a GPU-based spatial voxel fusion algorithm is proposed to improve the traditional volume fusion method in the rules of hierarchical grid datastructure.First,the surface data is stored only when it is observed,so that it is efficiently flowing into and out of the rendering table.Then,online large-scale scene reconstruction in the process of sensor movement is realized.Secondly,in order to achieve fine-grained details and large-scale interactive real-time reconstruction of the scene,several stages are completed,including the depth map pre-processing,camera attitude estimation,depth map fusion and surface rendering.Finally,it is shown through experiment,that the improved reconstruction algorithm improves the reconstruction accuracy compared with the traditional volume reconstruction method.The average reconstruction time is 60.6ms,the times of ICP pose estimation is 15 within 23.0ms,the reconstruction performance of algorithm is better.(2)In order to solve the problem that the traditional self-adaptive feature subdivision method has caused the surface rendering rate to decrease due to the over-subdivision of the local feature region,an improved on-line adaptive feature segmentation algorithm is proposed.Firstly,the feature block processing unit is constructed to generate the dynamic subdivision factor of the feature region.The individual subdividing depth of the feature block is calculated.Then,the voxel block of the feature region is subdivided according to the subdivision rule.The number of blocks with the same type at each subdivision level is parallelized by establishing the mapping relationship between the block buffer and the subdivision table.The reconstructed surface rendering speed is improved.Finally,the structure is expanded to generate new GPU schedules and render tables to support dynamic segmentation and real-time rendering of voxel blocks.Experiments show that the improved algorithm guarantees the reconstruction speed with the detail drawing for large scale scene at the same time.(3)In order to solve the problem of feature region being irrelevant in large-scale scene,a GPU fast target generation algorithm for repetitive feature region is proposed,which is more suitable for real-time rendering of a large number of recurring surfaces on sharp points or creases.Firstly,a unified double cubic B-spline based on octree is established for the cusp or crease edge region with the sametopology.The feature points on the adjacency vertex information are obtained,the feature points on the more accurate matching model are realized.The paper uses the breadth-first traversal strategy to encode the patches in the meta to generate the traversing element table,it drives the hardware processing unit of the GPU.Finally,the experiment shows that for a large number of surface features of the grid model,the algorithm allows for the seamless reconstruction of the model,it achieves the real-time rendering for large-scale scene in the repeated topology regions.(4)A CPU to GPU reconstruction optimization framework is implemented to estimate the drift phenomenon of the traditional online volume fusion system.Based on the robust camera pose estimation strategy,all RGB-D input values are fused to an effective hierarchical optimization framework.Each frame is optimized according to the global camera pose,which eliminates the serious dependence on the tracking aging.It allows for continuously tracking of the global Optimized frames.In this paper,the system optimizes the global optimal pose(binding adjustment)in real time,it supports robust tracking and reestablishment(re-positioning),it also re-estimates the large-scale 3D scene to ensure global consistency.It is a set of sparse corresponding features,geometric and ray matching functions in one of the parallel optimization framework.Experiments show that for the three-dimensional reconstruction of complex environment,this system allows for the accurate reconstruction as well as the real-time tracking,while the reconstruction scale is guaranteed in a better way.