3D Indoor Scene Generation Method For Virtual Navigation

Virtual navigation is widely used in real estate navigation,panoramic maps and virtual tourism.There are two main methods for scene generation method in virtual navigation: one is to use three-dimensional models to build the scene manually,the effect is real and the details are rich.But it has poor efficiency and low versatility;the other is to use images.Which can efficiently generate large-scale scenes,but there are also drawbacks such as limited view and bad interactivity.In this paper,an automatic generation method of indoor scene based on the floorplan and real images is proposed,which can efficiently generate a virtual 3D scene that is as close as possible to a real scene.This paper mainly carries out the following aspects of research.(1)This paper uses a floorplan analysis algorithm based on combination of deep learning and integer programming.First,deep convolution neural network is used to extract feature junctions of the floorplan.Afterwards,an integer programming algorithm is used to further solve the vectorized representation of the floorplan based on the extracted junctions feature.The algorithm has good robustness and achieves 90% accuracy.(2)An indoor objects layout estimation method based on prior knowledge is proposed.First a room layout estimation algorithm is used to obtain the bounding box of the room.Second our approach uses object detection to obtain the category and position of the indoor objects.Finally,optimization algorithm is used to estimate the spatial layout of the objects.In order to ensure the accuracy and rationality of the result,this method defines some constraints and assumptions based on the prior knowledge of the floorplan and indoor scene.(3)A 3D indoor scene generation system for virtual navigation is realized.Given the floorplan and several real images of the scene,our approach first retrieves the corresponding3 D model from the indoor model database,and places them in a 3D scene based on the indoor layout estimation result.Then the results are evaluated and optimized according to the object detection result.Finally,a virtual scene that is consistent with the structure of the real scene can be generated and virtual navigation for this scene can be realized.