Font Size: a A A

Monocular Depth Estimation And Reconstruction For Indoor Scenes Based On Deep Learning Methods

Posted on:2020-02-07Degree:MasterType:Thesis
Country:ChinaCandidate:Y WangFull Text:PDF
GTID:2428330575956416Subject:Information and Communication Engineering
Abstract/Summary:PDF Full Text Request
Three-dimensional scene perception and reconstruction has always been an important research topic in the field of computer vision,and depth information is an important content to understand the three-dimensional geometric relationship of the scene and reconstruct the three-dimensional spatial structure of the scene.Current depth estimation and scene reconstruction algorithms either rely on expensive depth acquisition equipment or are complex and time-consuming,which improves the difficulty of practical application of 3D scene reconstruction technology.In view of the limitations of different depth estimation and three-dimensional scene reconstruction techniques,this paper studies the depth estimation and reconstruction for indoor scenes based on deep learning methods.The main work of this paper is as follows:1.Based on the traditional deep convolutional neural network,this paper proposes a pyramid monocular depth estimation model.By introducing the spatial pyramid pooling module,the model deploys the multi-scale learning strategy to effectively obtain the global image context information.In addition,for the uncertainty of mapping between monocular images and corresponding depth maps,this paper proposes a pyramid loss function to improve the accuracy of depth estimation.2.We use feature-point based method to extract features from input images and perform feature point matching.Then,the process of camera poses initialization is finished by combining the geometric constraints to obtain the camera's rotation matrix and translation vector.Afterwards,the iterative-closest-point method is used to iteratively recover accurate camera motion traj ectory parameters.3.This paper uses the graph optimization method to optimize the estimated camera pose parameters,and finally splicing depth maps and two-dimensional images to build a monocular indoor scene reconstruction system that can truly reflect the three-dimensional spatial distribution.The proposed monocular depth estimation and reconstruction algorithm based on deep learning in this paper takes the accuracy and real-time performance of the algorithm into account,and is robust to environmental changes,scene material differences and other influencing factors.
Keywords/Search Tags:scene reconstruction, deep learning, monocular depth estimation, camera pose estimation
PDF Full Text Request
Related items