Research On Extraction Of Quantity And Spatial Relationships Of Street View Ground Objects Based On Deep Learning

Geographic scene cognition is the identification of the structure and function of the geographic scene.The difficulty lies in the extraction and expression of the knowledge of the target shape,semantics,quantity and spatial relationship that constitute the geographic scene.Street view is a special geographic scene,and street view intelligent cognition can provide information services for government departments to implement supervision,decision-making,planning and other work.However,how to recognize and understand massive amounts of complex streetscape information and extract the spatiotemporal information required for supervision,decision-making,and planning is a major problem in the field of geospatial cognition.To address the problem of intelligent cognition of street scenes,this thesis proposes a method for extracting the number of street scene features and spatial relationships based on the joint attention mechanism of panoptic segmentation network,which realizes the cognition and expression calculation of street scenes,firstly,fuses the algorithms of panoptic segmentation and attention mechanism and spatial relationship calculation to extract,describe and express the street scene information formed by human in production and life,realizing the extraction of street scene entities,semantics and relationship extraction,and then employing semantic templates to generate image semantic and spatial relationship description statements to achieve accurate cognition and information expression of geographical scenes,and finally.The main research content of this thesis and its corresponding research results can be summarised as follows.(1)The attention mechanism module(Convolutional Block Attention Module,CBAM)of the convolution module improves the segmentation accuracy of the panoramic segmentation model and further accurately extracts street view information.This module reduces intra-class segmentation inconsistencies and unsegmented small objects by establishing the correlation between features in channels and locations,and further extracts more accurate target information.Experiments show that the Panoptic Segmentation(PQ),Segmentation Quality(SQ)and Recognition Quality(RQ)of our method in the Cityscapes panoramic segmentation dataset are 61.8%,81.6% and 75.4%,respectively.(2)This thesis proposes a method for extracting the number of features and spatial relationships in street scenes.As the target occlusion intelligence algorithm extracts the number of features and spatial relationships in street scenes often has the problem of error in judgement.Based on this,this thesis uses hindcast object extraction rules,spatial relationship extraction rules,and a panoptic segmentation algorithm depth estimation algorithm to optimise the panoramic segmentation prediction results in order to extract the accurate number of features and spatial relationships.(3)GUI software developed using Py Qt5.The software encapsulates a panoramic segmentation model and a method for extracting the number of features and spatial relationships,enabling the display of the results of the panoramic segmentation of the street scene,the visualisation of the number of features and the representation of spatial relationships in the street scene,and the retrieval of images of the categories required by the user,enabling the initial application of intelligent cognition of the street scene and assisting the user in decision-making.