Research On Generating And Annotating Point Cloud For Indoor Scene

As a high-level visual technology,indoor scene understanding is a popular field of artificial intelligence and computer vision.It has a wide application prospect in the fields of robot automatic navigation,robot arm control and unmanned vehicle.Indoor scene understanding technology can be divided into scene acquisition and scene environment perception two aspects,including 3D environment acquisition,object positioning,object recognition and other tasks.This paper focuses on the two aspects of the understanding of the scene:In view of the scene acquisition,this article deeply studies the three-dimensional reconstruction technology,mainly includes structure from motion(SfM)and simultaneous localization and mapping(SLAM).Aiming at the characteristics of indoor scene lighting changes,texture features less,a large number of objects overlapping occlusion,combined with the popular RGB-D data in recent years,the 3D scene reconstruction program based on Kinect equipment was developed.According to the environmental perception,understanding the research status at home and abroad after learning to understand the semantic annotation and annotation method of conditional random fields,the traditional CRFs is slow,the difficulty of solving the design features,combined with super pixel and point to the CRFs semantic annotation scheme.The main work and innovation of this paper are as follows:1)In depth study of the motion structure restoration method,and a lot of three-dimensional reconstruction experiments of different scale scenes are carried out.2)According to the processing of data frame,the process of motion structure recovery is improved by combining the map location and the immediate reconstruction technique,and the method of off-line motion restoration based on RGB-D data is designed and implemented.The method has the characteristics of fast reconstruction and uniform distribution of point cloud.It has laid a good foundation for the further research of plane extraction and point cloud segmentation.3)Based on the motion structure restoration method based on RGB-D in this paper,the key steps of camera calibration,depth map preprocessing,real scene data acquisition and reconstruction are verified.4)Based on in-depth research of current superpixel segmentation algorithm,improved RGB-D data for the classical SLIC algorithm,the algorithm experiments show that in the case of high precision depth map,the method in this paper can obtain better segmentation results for semantic segmentation comparing with the other recent algorithms.5)In combination with the superpixel segmentation algorithm,the pixels are replaced by superpixels as nodes.In addition,by using one-vs-all support vector machine to predict the class probability of superpixel blocks,the results of the semantic annotation of the edge preserving are achieved by solving the conditional random field.According to the research results of 3D reconstruction and semantic annotation,this paper provides a new idea and method for the point cloud reconstruction and semantic annotation based on semantic annotation sequence.