Research On 3D Indoor Scene Technology For Video Sequences

With the continuous development of China's real estate industry,applications such as VR viewing and virtual interior decoration are used in the real estate industry.The key to these applications is how to construct indoor scene models.Under the assumption of the Manhattan world,indoor scenes often conform to "structured scenes".There are a large number of geometric primitives subject to geometric constraints in such scenes.These geometric constraint relationships such as vertical and parallel,etc.,can use these primitives to reason about indoor scene layout and three-dimensional models.This article takes the key technology of 3D indoor scene modeling for video sequences as the research content,attempts to infer the structure of the indoor scene by extracting effective geometric primitives from the indoor scene images,and builds on the relevant theoretical knowledge of 3D modeling Indoor three-dimensional scene.This article introduces the general indoor scene layout estimation method,and focuses on the analysis of key steps related technologies.At the same time,the relevant theoretical knowledge of convolutional neural network and multi-frame image stitching technology is introduced to provide theoretical knowledge for the design of indoor scene modeling system based on multi-frame images.The main research contents are as follows:1.Aiming at the problems of line segment detection error and incomplete detection in the existing line segment detection algorithm based on convolutional neural network,this thesis proposes an improved line segment detection algorithm based on convolutional neural network.First use the endpoint detection network to extract line endpoints from the image,and then use the stacked hourglass network to obtain a line heat map of the image.This thesis proposes a new method to generate straight lines by combining the end point information and the line heat map to reduce the problem of false detection.Experimental results show that the algorithm effectively improves the problems of line segment detection error and incomplete detection in the line segment detection algorithm based on convolutional neural network,and improves the line segment detection accuracy.2.This thesis proposes a method to reason about indoor scene layout using right-angle features.First,extract the line segments in the indoor scene,use these line segments to estimate the three vanishing points of the indoor scene,then classify the wall pairs to which the straight lines belong,and then form a right angle according to these line pairs.Finally,after performing spatial right-angle classification on right angles,an error function is used to select the optimal angle from all right angles as the wall angle to form an indoor scene layout.The experimental results show that the algorithm can successfully estimate the indoor scene layout,and the correct pixel percentage is higher than the previous method.3.This thesis integrates the line segment detection algorithm and indoor scene layout estimation method of the previous two chapters,and gives the design of a three-dimensional indoor scene modeling system for video sequences.By combining the method of image stitching,a three-dimensional indoor model is constructed.This method can meet the needs of applications with structured models with orthogonal relationships and can automatically texture maps.Experimental results show that this method can effectively construct indoor scene models from real indoor scenes.