Research On Color Enhancement And Semantic Segmentation Of Building Point Clouds For Realistic 3D

Buildings are important 3D real scene geographic entities and play a crucial role in many application areas such as smart city construction,smart transportation and digital twin.The pursuit of better visualization and richer semantic information has been an important goal in the field of real-world 3D application.However,the shadow caused by the vertical structure of buildings affects their visualization.At the same time,the point cloud data is challenging to semantic segmentation due to its disorderly data organization,spatial dispersion and the lack of topological information.To address the above problems,this paper visualizes and enhances the shadows of building point clouds,and adopts KNN to collect multi-scale information of point clouds to improve the accuracy of deep learning on semantic segmentation of point clouds:(1)Presented a new color enhancement method for dark parts of buildings based on building surface segmentation.It improves the RANSAC(Random Sample Consensus,RANSAC)algorithm for surface segmentation of buildings and proposed a ‘light vector’ to divide the segmentation results into dark and bright surfaces,then proceed with color enhancement for dark surfaces separately.The experiment result shows that the proposed method can effectively improve the visual effect of 3D point cloud buildings.The brightness and contrast of the building point cloud are enhanced while the buildings’ style is maintained,and the influence of building shadows on the visualization effect is eliminated partly.(2)Presented a deep learning network for semantic segmentation of building point clouds fusing multi-scale features.This network first extracts multi-scale features in the local range and inputs them into a MLP form the multi-scale feature extract module.Then take the max pooling at different scales for each scale feature and sum the results to form the multi-scale feature fusion module.Combined with an attention pooling module for multi-scale features,achieves fusion of features at different scales on a local scale,enhanced feature information representation in different scales.Through the experimental analysis,the overall accuracy ratio of proposed network is 83.5%,got higher semantic segmentation accuracy,which shows that proposed network extracts the local multi-scale features of the building point cloud more effectively.(3)The above research contents are integrated to build a point cloud semantic visualization 3D GIS system.The system is in the form of a 3D GIS web system with front and back-end integration,which firstly integrates all the data appearing in this paper to complete the multi-source data fusion,then visualizes the results before and after point cloud enhancement respectively to highlight the color enhancement effect of building point clouds,and finally extracts point clouds such as doors and windows based on the results of semantic segmentation of building point clouds with fused multi-scale features,and uses its spatially different separated windows and doors by using clustering algorithm,and the semantic monolithic display of the building point cloud windows and doors is realized by combining with OBB wraparound box.The system is an integration and visualization display of the research results of this paper.