3D Indoor Scene Modeling Based On Deep Learning

Today,robots are beginning to be used in more and more scenes,such as service robots in homes,shopping malls,and banks,and functional robots in factories and hospitals,as well as combat robots in military surveillance.In order to accomplish different tasks well,robots should have the ability of environmental awareness,interaction and combination between perception and interaction.However,the level of intelligence and perception of the robot are still in the initial stage of development.Therefore,the core of current robot research mainly focuses on two aspects: one is to improve the autonomy of the robot,so that the robot can automatically complete the task instructions given by the operator;the other is to improve the adaptability of the intelligent robot,so that it has the ability to adapt to changes in the environment.In two aspects of research,accurate modeling ability is the premise and foundation for autonomy and intelligence.Therefore,how to understand the indoor scene and understand the structure of the scene layout,and then achieve fast and refined 3D reconstruction is the key to completing various interactive tasks quickly and accurately.Research on indoor scene modeling has made a lot of progress,especially the proposal of the modeling framework based on multi-view fusion and the modeling framework based on single view,enhances the environmental perception ability of robots.However,the above work still has the following shortcomings :(1)the pre-processing based on multi-view fusion takes a long time,and offline optimization process is required after the modeling,which cannot meet the modeling requirements under specific conditions(2)the indoor scene of the modeling framework based on the single perspective is mainly represent by voxel,so the modeling quality is low and the information is seriously missing.The details of the objects in the indoor scene always may loss and it is difficult to meet the requirements of robot interaction.In view of the above deficiencies,this paper aims to analyze the influence of the number of input views on indoor scene modeling,and to explore the relationship between scene color map and depth map and different view's depth maps.Firstly,a modeling algorithm based on convolutional neural network for database replacement is proposed to model the scene.Then,a multi-view depth maps generation network and a fused network are built to predict depth information.Finally,with the aid of the database model modeling scenario,the depth map generation algorithm is used to generate depth maps of multiple views and merge them to achieve high-precision scene reconstruction,thereby greatly improving the environment awareness and intelligence of existing intelligent robots.The main contributions of this paper are as follows:(1)We constructed a scene modeling method based on database template replacement and the method use convolutional neural network to learn the high-dimensional features of objects.The scene reconstruction is completed by replacing the target object with the model in the database.(2)We constructed a depths generation network based on multi-view fusion,the network accurately predicts the depth map of the color map under a single view by summarizing the relationship between single-view color map and the depth map and the relationship between depth maps in different views(3)We constructed a single-view indoor scene modeling framework.The framework uses the modeling framework and the generated network proposed in(2)to generate depth maps of different views of indoor scenes,and finally complete the indoor scene modeling task through multi-view fusions.