A Research On Hierarchical Indoor And Outdoor Scene Understanding Method Based On 3D Point Clouds

Environment understanding is the basic work in the field of computer vision and has been widely applied in various military and civil fields.Due to the complexity of spatial structure and the diversity of target types,it is difficult for a single-level environmental understanding to cross the shortage of “semantic gap” to infer a higher level of spatio-temporal semantic information model,thus exposing the deficiency of the representation ability of environmental modeling.In order to address this limitation,this thesis studies the hierarchical understanding method of""point-object-space" based on 3D point cloud according to the construction level of objects.Through the scene analysis methods such as machine learning,the performance of fine-grained point cloud classification,specific object recognition and spatial topology reconstruction is improved.Consequently,a hierarchical analysis,and understanding of environment content is formed in a coarse-to-fine manner.The main contributions of this thesis are as follows:(1)Aiming at the limitation of the insufficient discriminative capability of low-level visual features in the point cloud classification task,the semantic correlation of local visual features within each primitive is analyzed comprehensively,to excavate the middle-level semantic representation model of classification primitive,and a probability graphical model framework based on Bayesian network model and Markov random field model is constructed,to generate locally continuous and globally optimal results.(2)To address the challenge that the saliency in linearity of linear objects in the object recognition task,based on the relationship between the local geometric characteristics of linear target and its radius,a linear structure recognition method based on Laplace smooth skeleton structure analysis is designed,where the implicit shape of the object is mined by the idea of shape regression,to relieve the problem of weakening local linear characteristics caused by its own specific structure.(3)Aiming at the complex layout structure and diverse space constraints of indoor scenes in the space topology reconstruction task,based on the prior hypothesis that indoor subspace(such as rooms and corridors)and their topological relations are defined by architecture entity elements(such as walls,doors and windows),a new spatial topology reconstruction model of indoor scene based on the node-relation-graph model is designed,where the spatial semantic information is used to assist the adaptive partition of indoor subspace and the robust reconstruction of the topological relationship of adjacent subspaces.Moreover,the different states of door leaf are considered to optimize the rationality of the reconstructed topology.As a result,a noderelation-graph model is established to accurately express the topological relationship of indoor space.In this thesis,3D point clouds of street corridors,electricity corridors and indoor scene are used as data sources to carry out experimental analysis of point classification,object recognition and space topology from qualitative and quantitative perspectives.Experimental results show that the overall accuracy of point classification results obtained in this thesis method is on average 4%higher than other existing methods,especially for the classification of cars and pole-like objects;In the street and electricity corridors,the object detection results in this thesis exhibits a remarkable improvement compared with those from the non-skeleton structure analysis method with differences in quality of approximately 3.59% on average,compared with other methods with differences in quality of approximately 3.70%;In the indoor scene spatial topology reconstruction,the space partitioning method in this thesis is insensitive to parameters,with the average accuracy of approximately 85.7% and the average recall of approximately 87.6%.Moreover,the correctness of the spatial topology described by the generated indoor node relationship graph model is verified in the application of indoor optimal path planning.According to the characteristics of 3D point cloud data in specific application scenarios,this paper proposes a hierarchical understanding method framework to meet the needs of different applications,which will promote the in-depth application of 3D point cloud in the fields of autonomous vehicles,3D reconstruction,navigation and location service.