A Structural Displacement Monitoring Method Based on YOLOv8-SAHI and Gradient Correlation Matching

To address deployment difficulties and insufficient measurement accuracy associated with traditional structural displacement monitoring methods in practical applications, this study proposes a non-contact visual monitoring method integrating deep learning and gradient correlation matching. A marker with multi-scale geometric encoding features was designed, and a dataset encompassing various interference scenarios was constructed. Utilizing the YOLOv8 model combined with Slicing Aided Hyper Inference (SAHI), precise localization of markers in small-target scenarios was achieved, reaching a mean average precision (mAP50-95) of 97.84 % after training. Furthermore, an innovative Gradient Correlation Matching (GCM) algorithm was proposed, establishing a subpixel displacement calculation model through the analysis of image gradient field characteristics. This effectively overcame the failures of traditional template matching under scale variations and partial occlusions. Comparative indoor and outdoor experiments demonstrated that, within a displacement range of 0-80 mm, the maximum RMSE of the proposed method compared with reference data was 0.064 mm, maintaining stability even under artificially induced occlusion conditions. The findings offer a novel technological pathway for long-term displacement monitoring of infrastructure such as bridges and high-rise buildings.