TransWallNet: High-Performance Semantic Segmentation of Large-Scale and Multifeatured Point Clouds of Building Gables

Abstract

Intelligent recognition of bulges, windows, and other features in building gable point cloud data is a prerequisite and critical step for the implementation of automated spray-painting in construction. Gable point cloud data exhibit characteristics such as large scenes, orthogonal structures, color degradation, and feature imbalance. Addressing these attributes, this paper proposes TransWallNet, a point cloud semantic segmentation model based on the attention mechanism. To alleviate the computational load from large scenes, the model employs random sampling. For the orthogonal nature of the gables, it innovatively utilizes Chebyshev distance to query neighbors, incorporating an attention mechanism to effectively aggregate local point cloud information. This allows for the reliance solely on positional information of point clouds to identify various features, addressing the issue of color feature reliance. The combination of local feature aggregation and a global attention module attends to both local point cloud details and their contextual relationships, accurately segmenting various gable elements. Compared to other leading methods, our approach achieved the highest macroaverage accuracy and macroaverage F1-score on a building facade data set, increasing by 9.81% and 4.55%, respectively, over other methods. This research provides high-quality environmental information and perception methods for the construction of gable spray-painting robots.