Research On Robust Reconstruction Of 3D Room Layout With Visual-Inertial Module

The goal of indoor 3D layout reconstruction is to reconstruct the structure and 3D layout of indoor scenes.It is important in the fields of virtual reality,augmented reality,Indoor navigation and so on.The reconstruction process can be divided into two steps:3D point cloud reconstruction and layout reconstruction.3D point cloud is often obtained by high-precision LIDAR scanner,consumer-grade depth camera,visual camera with corresponding algorithms.LIDAR scanner is very expensive.Consumer-grade depth camera is subject to constraints of environment.Visual camera is in great demand while the algorithm is difficult.Layout reconstruction often uses rule-based algorithm to get reasonable result.In recent years,with the development of deep learning,end-to-end solutions have emerged to reconstruct 3D layout from images.These methods are attractive,but the results are often lack of precision.In this paper,a robust method of indoor 3D layout reconstruction using visual-inerital module is proposed.First of all,the visual-inertial module is a combination of a visual camera and an inertial measurement unit.In recent years,it has been widely used for real-time localization and mapping.The visual-inertial module is mature and low-cost,and has obvious application advantages.Secondly,we propose to combine the real-time reconstruction result of visual-inertial odometry with the offline reconstruction method to solve the challenges in feature matching and incremental reconstruction under the challenge indoor scene of weak texture and repetitive structure.In addition,thanks to the visual-inertial module,witch can observe the direction of gravity,a robust method for determining the floor and ceil of the room is proposed.Finally,based on the Manhattan world hypothesis,this paper designs a rule-based bottom-up reconstruction algorithm for typical apartment rooms.Based on several planar structures extracted from point cloud data,we design an energy function,then the reconstruction problem are formalized into an optimization problem,and an efficient algorithm is designed to solve it.We used collected data and synthetic data for experimental verification.The experimental results show that the 3D point cloud reconstruction method of indoor scene combined with real-time reconstruction and offline reconstruction is better than the method of real-time reconstruction and the method of offline reconstruction.The layout reconstruction method we proposed can get correct and reasonable results in simple room data and complex room data.