RGB-ONLY WORK-AREA RECOGNITION FOR FIREPROOFING SPRAY ROBOT

Abstract: Deep learning–based segmentation is widely used for construction work-area recognition. But it may lead to mask boundary misalignment, degrading work quality in processes requiring precise work-area recognition, such as fireproofing spray. To address this limitation, we present a pipeline for accurate work-area recognition of steel structures. It uses deep learning for segmentation and edge detection, followed by computer vision for line extraction and work-area segmentation. Experiments show mean Intersection over Union (mIoU) gains of 4.47 pp before spraying and 4.51 pp during spraying over the segmentation model. These results indicate our pipeline enables more accurate work-area segmentation