Attentive Fusion Global And Local Deep Features For Building Facades Parsing

As the most basic and main element in the city,the 3D building models are widely used in indoor and/or outdoor navigation,building energy modeling,3D visualization and generalization,building abstraction,among others.Since Li DAR can capture the complex structure of rooftop and facade components,such as windows,doors,balconies,point clouds are widely used to enhance 3D real scenes.However,turning the point clouds to semantically rich,geometrically accurate and topologically correct building models for semantic understanding and geometric representation is still a major challenge.Compared with the recognition of rooftop construction units,superstructures and other ornaments on rooftops,parsing building facade components such as windows,doors,and balconies is extremely challenging because of the complexity of facade components(diverse window/door,irregular arrangements,etc.)and imperfect facade point clouds generally corrupted by outliers,irregularity,and missing data caused by occlusion and/or selfocclusion.In recent years,deep learning has shown good performance in processing point cloud semantic segmentation tasks.Compared with the traditional manual design feature extraction method,this kind of method has higher segmentation accuracy.Therefore,it has important research significance to use deep learning to process building facade.Feature extraction from unbalanced category data is still a challenging problem in point cloud semantic segmentation task,and the lack of point cloud annotation benchmark dataset on large-scale urban building component segmentation scale makes it difficult to realize deep learning parsing and understanding of 3D urban buildings.Therefore,based on the Dublin buildings point cloud dataset,this paper first establishes a large scale 3D building facade dataset with accurate annotation and abundant component annotation,so as to automate semantic segmentation of building facade using point cloud deep learning algorithm.Secondly,AFGL-Net,a deep learning framework for semantic segmentation of door and window based on building facade point cloud,is proposed to improve the semantic segmentation accuracy of unbalanced door and window point cloud.Aiming at AFGL-Net framework,this paper first designs the local feature coding considering the direction and position coding,strengthens the local feature aggregation,enhances the edge feature of the door and window boundary point and the local feature of the facade point.Secondly,AFGLNet integrates the global Transformer module,which can capture the global characteristics of window and door context,deduce the geometric position structure and structural layout of building facade door and window,and correctly identify the door and window with no significant edge features from the building facade point cloud with uneven density,noise and outliers and missing data.Finally,the integration of global and local features based on attention mechanism can more effectively express the geometric structure of facade door and window,and balance the error of door and window extraction commission segmentation caused by Transformer.In this paper,the validity of AFGL-Net is verified by ablation experiment,contrast experiment and robustness experiment by using Dublin and Rue Monge2014 semantic annotation dataset.AFGL-Net predicted semantic segmentation accuracy m Io U of 67.02% and 59.80% respectively on two datasets,both better than the mainstream deep semantic segmentation framework.