Research On Key Technologies Of Indoor 3D Scene Reconstruction Based On RGB-D Images

With the rapid development of virtual reality technology,the reconstruction of indoor 3D scenes has attracted the attention of many scholars.At present,most 3D scene reconstruction is based on laser scanning data.Although it can achieve a more ideal effect,the equipment is expensive,RGB-The D camera has been favored by many scholars because of its low price,compact structure,and simple operation.It can obtain both color and depth maps.This article focuses on the research of RGB-D images,and works on key technologies such as depth image repair,point cloud registration,point cloud segmentation,and model construction,which are involved in the indoor 3D reconstruction process.The specific contents are as follows:(1)Detailed analysis of the generation principle of the depth image and point cloud obtained by the RGB-D sensor.By analyzing the causes of holes and noise in the depth image,a noise reduction method based on ray tracing rendering is proposed,which decomposes the color of each pixel into Direct illumination and indirect illumination are two parts,in which the pixel illuminance is obtained through texture information,and then the depth image hole is repaired,and the contrast information is combined with time and space to compensate for the lack of information caused by the lack of samples.(2)Research on point cloud registration.To solve the problem of Iterative Closest Point(ICP)which is easy to fall into the local optimal solution,this paper first coarsely matches the original point cloud data before fine matching.After preprocessing the original data,this paper uses 4PCS algorithm and PFH respectively The four coarse matching algorithms of algorithm,FPFH algorithm and NDT are used for experiment.By analyzing the operation efficiency and registration accuracy of the four coarse registration algorithms,the precise matching is completed in combination with the ICP algorithm.Experimental results show that different algorithms have different application scopes and operating efficiencies.Among them,the NDT algorithm can be well applied to the construction of indoor scenes due to its high operating efficiency and high registration accuracy.(3)Research on RGB-D point cloud segmentation and fusion based on super-voxel.In view of the fact that the traditional point cloud segmentation algorithm cannot split the target well and affect the interaction of single entities in the subsequent 3D scene,this paper proposes a super-voxel fusion segmentation strategy.First,the indoor point cloud data is voxelized,according to the feature description and distance measurement.Voxelized point clouds are clustered into super voxels;then,through a local convex hull connection algorithm,the super voxels are further fused to achieve point cloud segmentation.This algorithm can better separate the different objects in the scene point cloud separately,and the edge information between the objects is well-segmented,which provides a basis for the interaction between the objects in the subsequent 3D scene.(4)Three-dimensional scene reconstruction based on RGB-D images.The indoor point cloud data is obtained through RGB-D image conversion.With the help of Geomagic Studio software,the model reconstruction is completed.In order to make the model more beautiful,this paper resamples the original data based on the least square method,and obtains the processed data by comparison The constructed model is more beautiful,with smaller voids and more prominent edge contours.Thesis innovation:(1)In view of the shortcomings of the traditional depth image repair algorithm,this paper proposes a ray tracing rendering noise reduction method.The illuminance information of the depth image is filtered by time and space to compensate for the lack of information caused by the small number of samples,and the noise of the depth image is removed and the hole area is repaired.(2)The point cloud modeling effect generated by the original RGB-D is not good.In this paper,the original data is smoothed based on the least square method,and then the scene point cloud is reconstructed with the help of Geomagic Studio software,which reduces the network structure.The occurrence of holes in the process makes the constructed model more realistic and the edge information is more obvious.