Detection And Risk Assessment Of Submerged Vehicles In Urban Waterlogging Scenarios Based On Deep Learning

Urban waterlogging has always been a worldwide problem in urban construction and development.How to solve urban waterlogging and its derivative problems has been paid close attention and studied by scholars and research institutions at home and abroad.In recent years,the risk of waterlogging in many cities in China has increased greatly under the influence of some sudden weather events,such as the July21 incident in Zhengzhou.The occurrence of urban waterlogging damage to the city and people life and property damage are huge,especially when waterlogging flooding in the case of vehicle plays a very passive role,only on the detection of waterlogging recede assessment,subsequent to submerged vehicle evaluation,conducted under the statistics and track in human,and the cycle is long,the disadvantages of low efficiency,big spending.Compared with manual detection methods,target detection based on deep learning is more efficient,accurate and fast,and greatly liberates manpower and reduces costs.This paper proposes a deep learning-based hazard detection rating and counting tracking method for urban waterlogging vehicles,which provides more choices for urban residents and management decision-makers,as well as ideas and content-based research methods and models for relevant urban waterlogging researchers.The main work and achievements of this paper are as follows:Firstly,the demand analysis of submerged vehicles under urban waterlogging is carried out to determine the practical significance of the study.In terms of the risk classification of flooded vehicles,we refer to relevant theoretical exploration and experimental verification,and finally divide the risk of flooded vehicles into three levels.Secondly,the experimental data set is made according to the classification level.The original data is preprocessed by crawling pictures and videos on the Internet,and the data set is marked one by one and organized according to a certain format by using annotation tools.Then,YOLO,SSD,Faster R-CNN,Center Net,Retina Net and Efficient Det algorithms were used to train and optimize the flooded vehicle risk dataset.A variety of target detection models are obtained and the actual detection effect is analyzed.The optimal detection model YOLOv5 is selected based on the consideration of speed and accuracy,and the detection speed can reach 60 FPS in the local environment.Finally,the target detection and tracking model of urban waterlogging vehicle is built,which is YOLOv5 as target detector and Deep Sort as tracker.The excellent target detection ability of YOLOv5 S and the good target tracking effect of Deep Sort are used to realize the target tracking and counting of flooded vehicles in flood scenes,and maintain the tracking efficiency at the same time.Real-time processing capability is guaranteed.