Environmental Perception And Modeling Based On Generative Networks

Environmental perception and modeling is a multi-disciplinary field and covers all kinds of hardware and software processing systems.As the core of the perception and understanding of the original data,the intelligent environmental perception processing algorithm has always been a research hotspot.With the development of various measurement methods and sensor technology,the demand for integrated,intelligent and automatic environmental perception technology is increasing.However,in the face of complex environment and scenes,the types,dimensions and accuracy of data all present various forms.In addition to the diversity of perception and processing tasks,the technical requirements for perception algorithms are constantly improving.On the other hand,with the advent of the era of artificial intelligence,intelligent algorithms have been fully proved their effectiveness in computer vision,natural language processing or other fields.In environmental perception and modeling,some artificial intelligence algorithms are gradually used to deal with and solve practical problems,but many aspects are not mature due to the difficulty of tasks and short development time.This thesis studies environmental perception and modeling technology based on generative networks.According to the different data types,this thesis can be divided into two dimensional representation and three dimensional representation of the environmental perception algorithms.According to the different data range,it can be divided into object-level and scene-level perception algorithms.According to the different perceptual tasks,it can be divided into structural modeling and analysis,model completion,abstract representation,vectorization representation,semantic segmentation and understanding or other perceptual processing algorithms.The main contributions of this thesis include:1.Road network structure perception in remote sensing images based on generative adversarial networks: Road often presents complex and changeable spectral characteristics in remote sensing images.Due to the appearance or change of shadow,shielding,light intensity and other conditions,as well as the influence of resolutions,road conditions and distribution areas,the road extraction effects of remote sensing images in different environments is quite different.This thesis proposes a road network extraction algorithm for remote sensing images based on multi-supervised generative adversarial networks and a road network topology optimization algorithm for remote sensing images based on multi-conditional generative adversarial networks.The two networks can simultaneously study the spectral and topological characteristics of roads to optimize the extracted road network structures,so as to overcome the difference of the extraction effects caused by the changing road spectral performance to some extent.2.Object point cloud structure perception based on three-dimensional generative networks: Due to the uncertainty of the mapping relationship between 2d pixels and 3d spatial coordinates,3d reconstruction based on a single image is a tough problem.However,for deep-learning-based 3d reconstruction,most methods require the input image with a pure background and a specific perspective,while the generated model is not fine enough and the local structure is not clear enough.In addition,most previous laser point cloud completion algorithms can only complete the point cloud models with small missing structures.This thesis proposes a single-image object reconstruction algorithm based on 3d generative networks and a data-driven point cloud object completion algorithm.The two networks can adaptively fuse two-dimensional and three-dimensional information and extract more effective features for reconstruction,so as to improve the effect of model reconstruction and completion.3.Object structure perception based on hierarchical generative networks: For semantic reconstruction of three-dimensional objects,previous single-view object reconstruction algorithms are difficult to finish the task,and most works are for large-scale scenes.In addition,due to the diversity of program expressions and the complexity of3 d spaces,the solution space is too large to obtain good results in 3d inverse procedural modeling.This thesis proposes an object semantic reconstruction algorithm based on recurrent generative networks and a recursive generative network for 3d inverse procedural modeling.Among them,the object semantic reconstruction algorithm based on recurrent generative networks uses the single-view images and the projection of different parts as input,to generate local structures of the model orderly,so that the reconstructed models have precise semantic labels.The proposed recursive generative network for CSG inverse procedural modeling,can analyze the objext voxel models hierarchically and obtain the CSG programs,which can be used to reconstruct the models directly.4.Large-scale 3d scene structure perception based on generative adversarial networks:It is difficult to perceive and understand the spatial structure of the large-scale3 d scene due to its complex structure and large scale.In this thesis,a 3d guided multiconditional residual generative adversarial network for point cloud contour extraction and a contour-guided large-scale semantic segmentation algorithm for meshes are proposed.Among them,for the 3d guided multi-conditional residual generative adversarial network,by representing and processing the line structures in the parameter space and guided by the initial contours to instruct the extraction process,we improve the accuracy of the results.For the contour-guided large-scale mesh semantic segmentation algorithm,we optimize the ”superfacet” extraction results with the contour guidance,perform semantic segmentation based on a graph network,and obtain more accurate segmentation results.