Scenario Reproduction Of Beichuan Building Seismic Damage Based On UAV Image Recognition And Regional Earthquake Disaster Simulation

The ultimate goal of earthquake engineering is to reproduce earthquake damage scenarios.The use of diversified technologies to study seismic damage is helpful for better pre-earthquake assessment and faster post-earthquake reconnaissance.A primary framework for UAV data collection,automatic image segmentation,damage recognition and regional earthquake disaster simulation has been established after solving massive technical problems.This topic focuses on the reproduction of the building damage scenario caused by the earthquakes,and the Beichuan earthquake site is selected as the study area.The main contents of this thesis are as follows:(1)Data acquisition.UAV oblique photography technology was used to collect seismic damage images of the study area.After two days and five flights,2,580 photos were obtained,and the photos were initially processed to form a three-dimensional surface model of the study area.Besides,151 buildings' data in the site area,including the basic information before the earthquake and the damage degree after the quake,was collected through various methods such as field observations,historical data.(2)Image segmentation.Based on the collected building data and UAV image metadata,a three-dimensional model of the pre-earthquake building in the study area was generated.Meanwhile,the frustum parameters at the time of UAV image shooting were determined.Blender software was used to form the pre-earthquake virtual scene of the Beichuan earthquake site,where different grayscales were added to the virtual buildings in the frustum as labels.Through programming,Blender traversed the photo parameters to output the virtual images composed of grayscale color blocks as the mask layers of the pictures taken by the UAV.Open CV was then used to segment all single-building seismic damage images,totally 15,921.(3)Deep learning.Based on the PEER building seismic damage database,three deep learning frameworks,Paddle Paddle,Tensor Flow,and Py Torch,were used to train the Res Net-50 model,and the effect of model training between different frameworks was analyzed.After labeling the Beichuan earthquake site's building image data,Paddle Paddle was further used to train the Res Net-50 model,whose impact is evaluated by the accuracy rate and Kappa coefficient.(4)Regional simulation.Two regional earthquake disaster simulation software,You Simulator and Sim Center r WHALE,were used to simulate the pre-earthquake buildings in the study area.The simulation results were compared with the actual values,and the accuracy rate and Kappa coefficient were used to evaluate the prediction accuracy.The second development of Sim Center r WHALE was carried out.The parallel operation was developed by dividing the input data,and the VTK tool was used to realize the dynamic visualization of the building displacement.Through the research of this thesis,it is found that the damage states of buildings between basically intact and collapsed is not easy to obtain through image recognition;multi-angle building damage images are of benefit to identifying the damage state of buildings correctly;regional earthquake disaster simulation software is subjected to the uncertainty of basic building information,whose accuracy of the simulation is not high.