Research On Recognition And Classification Of 3D Point Cloud Targets In Urban Road Scenes Based On Semantic Information

In recent years,the vehicle-mounted scanning system that integrates GNSS,IMU,RGBD cameras and laser scanning systems with vehicles as the carrying platform has become an important means of three-dimensional information acquisition.Its data has the characteristics of high accuracy,high density,and real-time.Obtaining threedimensional surface model information of buildings and rods in the scene has great advantages in the production of high-precision maps,indoor and outdoor navigation and positioning,and smart city construction.However,three-dimensional point cloud data has the characteristics of target occlusion and overlap,incomplete data,complex and diverse scenes and object categories due to the impact of the platform,scanning method,and data collection environment of the collection equipment,resulting in insufficient geographic entity recognition and classification and extraction Characteristic difficulties and a series of problems.In response to the above problems,this paper proposes a semantic informationassisted three-dimensional point cloud data recognition and classification method for urban road scenes.The main research contents are as follows:(1)In view of the complex and highly discrete characteristics of the vehiclemounted laser point cloud scene,in order to reduce the influence of the ground points and noise points in the original point cloud data on the subsequent experimental results,statistical filtering is used to remove the sparse point clouds that deviate from the scene.Then filter out the ground points based on cloth simulation filtering(CSF)filtering and based on the elevation variance optimization algorithm;for the filtered point cloud data,first generate a single-scale supervoxel based on the local spatial characteristics of the point cloud and the echo intensity information,and then adopt the adaptive resolution the super voxel generation algorithm optimizes the result of super voxel segmentation,and obtains the over-segmentation point cluster with the boundary details of the ground object preserved intact;calculate the dimensional characteristics of the obtained different super voxels,these features provide the multi-regular region growth algorithm Data foundation,and improve the time efficiency of the algorithm.(2)Construction of semantic rule set for road scene.A conceptual framework for semantic feature recognition of urban road scenes is proposed.The automatic feature recognition and object classification tasks of urban road scene point clouds are divided into six steps: rule set construction,target detection,target recognition,segmentation,feature recognition and semantic-based The object classification of the information.Based on the semantic object classification of urban road scenes in different dimensions,the semantic object class of urban road scenes in this experiment is determined.Analyze the spatial metaphor rules between semantic objects from the three aspects of direction relationship,topological relationship and distance relationship,analyze the non-spatial metaphor rules of semantic objects from two aspects of shape attributes and semantic attributes,and extract and formalize the rules of semantic attributes.According to the semantic rules extracted from different features,different semantic objects are analyzed and expressed.The object-oriented method is adopted,and the production consistency rule expression method is adopted to express the semantic information as a computer executable language.(3)Semantic information-assisted 3D laser point cloud target recognition algorithm.Combining related algorithms such as multi-regular area growth,normalized cutting,feature extraction of feature target geometry and spatial distribution,the target recognition is carried out from the two directions of the recognition of building facade and detailed structure and the recognition of other features,thereby realizing the building Accurate identification of objects,trees,vehicles,power lines and other ground objects.Through detailed experiment comparison with point cloud target recognition and classification related algorithms,the results and evaluation indicators show that the semantic information-assisted three-dimensional point cloud target recognition algorithm proposed in this paper takes into account the attributes of semantic objects and spatial context information,and can effectively improve the efficiency of extraction and the accuracy of segmentation of overlapping features.