About the author:

1. Shanghai Key Laboratory of Engineering Structure Safety, Shanghai Research Institute of Building Sciences Co., Ltd.,Shanghai 200032, China; 2. Tongji University, Shanghai 200092, China

Abstract:

With the continuous expansion of the scale of reinforcement and renovation projects for existingbuildings, the demand for automation of concrete surface pretreatment, a key link in reinforcement andrenovation construction, has become increasingly urgent. This paper conducts a research on the visionsystem of building grinding robots,which adopts the YOLOv8 algorithm to identifywall marking pointsbased on machine vision technology. Firstly, datasets are prepared under different environmentalconditions, and the Roboflow platform （a computer vision development platform） is used to complete dataannotation. Various data augmentation methods are applied to simulate the actual grinding constructionenvironment. After the completion of model training, an image post-processing module is used to judgeand classify the image validity,which provides a basis for robot grinding operations. The research resultsshow that the YOLOv8s model after a series of optimizations has achieved significant improvements indetection accuracy, robustness and generalization ability,with its key indicators reaching a high level,which can meet the basic requirements of real-time performance and accuracy for grinding robots inpractical engineering applications. However, the comprehensive adaptability and long-term operationalstability of the model in complex construction sites need to be further verified in practical engineering.