Real-time Indoor 3D Reconstruction Based On ORB Algorithm And Loop Closure

Three dimensional reconstruction has always been the research direction of the field of computer vision,and it plays an important role in military,medical,entertainment,architecture and so on.With the gradual development of theory and the rapid increase in the level of computer hardware,3D reconstruction can be rapidly developed.Researchers have been unsatisfied with non-real-time 3D reconstructions and are moving toward real-time 3D reconstruction.Since twenty-first Century,the field of real-time 3D reconstruction has gradually matured.However,it will inevitably lead to the accumulation of accumulated errors during the reconstruction process,which will lead to errors such as faults,errors,and overlaps in the reconstruction model.Aiming at this problem,this thesis proposes a real-time indoor 3D reconstruction algorithm based on ORB and loop closure combined with SLAM algorithm.First of all,we researches the current mainstream feature algorithms deeply.Based on the experimental comparison and analysis,we selects the improved ORB algorithm as the feature algorithm.Second,based on the research of the framework and theory of the traditional monocular SLAM algorithm,this thesis uses the feature matching relationship between successive frames to estimate the corresponding camera pose.For the cumulative error problem,the loop closure in SLAM is introduced.Random fern algorithm was chosen as the loop closure algorithm to modify the pose trajectory.Finally,we combine the existing real-time reconstruction with SLAM algorithm,the point cloud obtained by the pose is stored in the TSDF model,and the reconstruction surface of the scene is restored by using the ray projection algorithm to complete the entire real-time reconstruction process.This algorithm has been verified by test set and comparative experiments,which proves that the camera pose accuracy obtained by this method is high,and solves the cumulative error problem of pose in real-time 3D reconstruction.It completely and accurately reconstructs the 3D model of the target scene,especially for the case of loopback,this thesis is better than other real-time 3D reconstruction algorithms,and there will be no erroneous reconstruction caused by accumulated errors.