Risk Target Identification And Displacement Measurement Of Buildings And Structures Based On Computer Vision

The digital twin city is mainly characterized by object modeling,information collection,and process simulation.The digital twin city platform can provide applications in the field of disaster prevention and mitigation and urban safety with significant advantages and broad prospects.Perception is one of the most important contents of digital twin cities.As an important perception technique,computer vision can obtain massive image and video information of the urban built environment.This study focuses on the information acquisition of risk targets.The purpose is to use low-cost perception techniques based on computer vision,such as UAV aerial photography and surveillance cameras,to carry out the identification and displacement measurement of risk targets in the urban built environment,and thus support the applications in the field of disaster prevention and mitigation and urban safety with basic data.The main results are as follows:(1)A method for semantic segmentation of the aerial point clouds of the urban built environment is proposed,which combines 2D image and 3D point cloud semantic segmentation models.The accuracy of semantic segmentation of risk targets is improved.Low-complexity 3D models for urban wind environment analysis,including terrains,buildings,and canopy fluid volumes,are established.The practicality of the method is demonstrated through a case study of an urban area in the south of China.(2)The identification of risk targets for monitoring is performed based on computer vision.For point cloud data,a point cloud semantic segmentation model is trained to identify the target structural components in beam-column nodes and serve the purpose of object-based monitoring and comparison.For image data with sufficient samples and fine labels,an image semantic segmentation model is trained to identify and locate the target non-structural components in UAV oblique photographic photos.For image data with insufficient samples,an image classification model is trained by a virtually rendered training set to filter out the surveillance camera data containing target structures of interest.(3)A three-dimensional structural displacement measurement method using monocular vision and deep learning pose estimation is proposed.The method uses virtual rendering to construct the training set by combining multi-view images with random backgrounds and trains the deep learning model for detection and pose estimation of target objects.The three-dimensional structural displacement is measured based on the origin and destination poses or based on the origin poses and keypoint matching.The effectiveness of the method is validated through experiments.(4)A method for improving the accuracy of the vision-based displacement measurement using a GAN-based super-resolution model is proposed.By supplementing the high-resolution photos of the target object,the texture details of the low-resolution surveillance image and videos are improved,thereby improving the accuracy of displacement measurement.The effectiveness of the method is validated through experiments.(5)The shaking event of a super high-rise building in the south of China is taken as the case study to apply the proposed risk target identification and displacement measurement method,which proves the feasibility of this study.