Abstract: In order to improve the monitoring ability of slope hazards,this paper proposes a slope anomaly monitoring technology based on deep learning and image local feature extraction. By extracting the two-dimensional coordinates of natural features of the slope,this technology constructs the triangular target network. As the slope danger range is defined by the changing area of the triangular network,feature points with the same name are extracted within the change range,while the displacement of those feature points describes the slope change. The first step is to take images before and after the slope occurs,followed by identifying the natural features of the slope with the target detection model YOLOv5. In the semantic segmentation model DeepLabV3+,the extracted natural features are semantically segmented to obtain their binarized areas,and their two-dimensional coordinates are determined by determining the centre of the binarized area. As a next step,the triangular target network will be constructed by combining the two-dimensional coordinate lattices of all natural features,and the slope change range is delineated as the triangular network changes. After analyzing the image,the feature points with the same names within the change range are extracted using the image feature extraction technology,and their displacement distance and direction are used to evaluate the slope change. According to the test results,this technology is effective at monitoring slope changes,and it is a feasible tool for slope monitoring engineers.

Key words: slope, natural features, monitor, triangular network, deep learning, feature extraction