UAV Nap-of-the-object Photogrammetry For Bridge Surface Disease Detection Method

China’s infrastructure construction is changing rapidly.As an important part of China’s infrastructure construction,bridges demand increasing maintenance and repair.At present,concrete bridge surface damage detection still takes a lot of manual effort,which is inefficient and prone to safety accidents.With the continuous development of computer technology and intelligent control technology,aided bridge apparent disease inspection by UAV can significantly improve the inspection efficiency and liberate labor force.Based on mapping industry-grade UAVs,this thesis proposes an integrated inspection scheme for concrete bridge apparent damage inspection.According to the demand for rapid and refined bridge inspection,the proposal is designed and proposed for bridge nap-of-the-object photography data acquisition,3D fine model reconstruction,and bridge apparent disease detection and quantification,taking into account the complex structure and serious obstruction of bridges.By collecting high-resolution images of bridge surface through UAV nap-of-the-object photography,the deep learning target detection algorithm is used to quickly find images with surface diseases from the massive nap-of-the-object images,and then the deep learning semantic segmentation model is used to process the surface disease images to achieve accurate extraction and quantitative calculation of surface disease information.The practicality and reliability of the bridge apparent disease detection method proposed in this thesis are verified by examples.The research mainly includs:1.Firstly,the bridge is photographed by UAV tilt,the tilt image of the bridge is collected,and a rough 3D model of the bridge is established.Based on the bridge sketch model,the bridge girders,cover girders and piers are planned for nap-of-the-object photography routes respectively,and then the bridge UAV nap-of-the-object photography route planning file is generated and uploaded to the UAV for nap-of-the-object photography to automatically collect the bridge detail high-resolution images.In order to visualize the actual location and distribution of the apparent bridge damage,aerial triangulation,image dense matching,texture mapping and other operations are performed on the bridge detail images to establish the bridge refinement model.2.Based on the YOLO V5 target detection model,the problem of difficult detection and low detection accuracy is caused by the characteristics of multi-scale and small target of bridge surface defects,the BPS＿YOLO V5 target detection model is established by optimizing the C3 attention module,introducing the PAN weighted cross-layer cascade mechanism,and choosing SIo U to replace the CIo U loss function.Based on the constructed bridge apparent disease sample dataset,BPS＿YOLO V5 is tested and performance analyzed.The results show that the BPS＿YOLO V5 model has improved the detection accuracy of cracks,spalling,exposed tendons,etc.,among which he full category average accuracy(m AP)is 3.7% higher than that of YOLO V5 model,which fully proves the effectiveness of BPS＿YOLO V5 model.3.Based on Mobile Net V2,a lightweight Deeplab V3+ segmentation model is proposed for the pixel-level qualitative and quantitative detection of three types of diseases: cracks,spalling and exposed tendons.By replacing the backbone network responsible for feature extraction in the Deeplab V3+ model with the Mobile Net V2 model,and by fusing the advantages of the lightweight Mobile Net V2 and the high accuracy of the Deeplab V3+ model,the model significantly reduces the number of parameters and the computational effort while ensuring accuracy.The images with apparent diseases detected by BPS＿YOLO V5 are used as a new dataset to segment the disease images,and the experiment proves that this model has high segmentation accuracy for three diseases,namely,cracks,spalling and rust,and quantifies the diseases using the segmented images.4.Based on the bridge apparent disease image square information obtained by deep learning semantic segmentation and combined with the aerial triangulation results of UAV nap-of-the-object photogrammetry,i.e.,the image internal and external orientation elements,the apparent disease image square information in the image is mapped to the fine 3D model of the bridge,which facilitates the maintenance personnel to determine the actual location and distribution of bridge diseases more conveniently and provides the basis for subsequent maintenance.In this thesis,a bridge in Zaozhuang,Shandong Province,is used as the research object,and conducted field tests in order to verify the feasibility of this method.The bridge apparent disease detection method proposed in this thesis can greatly improve the detection efficiency of bridge maintenance personnel,reduce the bridge maintenance cost as well as the safety risk of maintenance personnel,which has important engineering practical application value.