Research On 3D Modeling Of Indoor Scene Based On Multi-source Data Fusion

In recent years,indoor 3D model reconstruction has played an increasingly important role in new application fileds such as service robots,house decoration design and semantic localization and navigation.The point cloud of RGB-D camera with texture information is particularly suitable for three-dimensional reconstruction of indoor scenes,but reconstruction model will have holes and bending deformations.In order to improve the reconstruction quality,point clouds of sequence images and camera are captured separately,and then these two types of point clouds are registered to achieve triangulated mesh reconstruction and texture mapping.Experiments indicate that the point cloud reconstruction model is more delicate and closer to reality than that reconstructed by a single source of point data.The main research contents of this thesis are as follows:1.Design a RGB-D camera acquisition system and analyze its depth data acquisition principle and error sources.The camera calibration method of Zhang Zhengyou was used to calibrate the internal and relatively extrinsic parameters of the texture camera and depth camera,which make a good foundation for point cloud acquisition.Afterwards,an overview analysis of the combination of camera point cloud and sequence image point cloud is performed to determine the overall technical framework.2.Research on multi-source point cloud data collection method.For the purpose of depth camera point clouds acquisition,various real-time dense point cloud reconstruction systems are analyzed and compared firstly,Elastic Fusion algorithm with the highest reconstruction quality is chosen as the framework basis;for the point cloud holes and random noise of the depth image,frame weighting and bilateral filtering is used realizes the depth map of a single frame,improves the reconstruction algorithm framework,and improves the accuracy by at least 10%.For the color point cloud model generated by triangulation sequence images,the model is marked with points,and scale constraints are imposed to reduce the relative error below 1% when convert to true scale standard.3.The point cloud fusion registration and model reconstruction algorithm are studied,in which the point cloud registration is divided into two stages: coarse registration and fine registration.Before point cloud registration,preprocess of the scene point cloud is performed,including target point cloud extraction,adaptive voxel grid downsampling and statistical filtering.In the coarse registration stage,the texture information of the color point cloud is included in the distance error function,and the improved sample consistency initial registration(SAC-IA)is used to find the optimal coordinate transformation matrix,which improves the efficiency by 15% to 25%.In the fine registration stage,the SPCA algorithm is firstly used to accurately calculate the normal vector,and the traditional consistent point drift algorithm(CPD)is improved to extend the error distance measurement from point to point to point to surface,and a multi-level consistent point drift Algorithm(MCPD)is proposed,and then combined with iterative closest point(ICP)fine registration algorithm tuning to ensure accurate and stable point cloud fine registration.Experiments show that the registration accuracy and efficiency of this algorithm are better than other methods.Finally,the three-dimensional reconstructions of indoor scenes are carried out by means of grid reconstruction and texture mapping of the fusion point cloud.