Curtain wall frame segmentation using a dual-flow aggregation network: Application to robot pose estimation

Abstract

In the field of curtain wall construction, manual installation presents significant safety hazards and suffers from low efficiency, while automated installation is constrained by the limited localization capabilities of curtain wall installation robots. In this paper, an automated installation solution based on machine vision is proposed, and a detailed discussion of several steps involved is provided. To locate the installation area, DANF, a deep learning-based dual-flow aggregation network designed for curtain wall frame segmentation, is proposed. It employs Transformer for global analysis and CNNs for detailed feature extraction to handle curtain wall frame structures. On the dataset constructed in this paper, DANF achieves an IoU of 85.19 % with a parameter count of only 4.24 M, demonstrating higher accuracy compared to other algorithms. Additionally, a pose-solving method based on the semantic segmentation results of the curtain wall frame is designed to adapt to curtain wall installation scenarios.