Research On Bridge Concrete Surface Defect Detection Method Based On Image And Laser Point Cloud

China’s infrastructure mostly uses concrete structures,such as bridge girders and road surfaces,high speed rail girders,large chimneys of thermal power plants and other major infrastructures use concrete structures.Take bridges as an example,the health status of bridges is related to the safety of people’s lives and national property security,and the concrete surface defects of bridges will not affect the bearing capacity and operation status of bridges in the short term,but will lay a major hidden danger to the bridge safety service,if not timely detection and early warning will lead to major safety accidents.With the continuous development and improvement of miniaturized vision sensors and LIDAR technology,bridge inspection tends to develop in the direction of informationization and intelligence.The visual inspection platform has significant advantages in bridge inspection in terms of high efficiency,low cost and high resolution,solving the problem of high human cost and low manual efficiency.However,the current visual inspection technology mainly focuses on the identification of two-dimensional defect information of concrete structure images,and this method has low detection accuracy for the three-dimensional size of defects.Based on this paper,we propose the research of surface damage detection method of concrete structure combining 2D image and laser point cloud information,and carry out the research of surface defect recognition method based on deep learning and laser point cloud visualization based on three forms of defects: cracks,holes and exposed reinforcement,aiming to improve the efficiency and detection accuracy of concrete surface defects of bridges,with the following main research contents:(1)A deep learning-based method for detecting defects on the surface of bridge concrete structures,a defect detection scheme is designed,images of bridge concrete surfaces are acquired,and bridge concrete defect images are pre-processed to construct a dataset for network training.Based on the characteristics and performance requirements of the detection platform,the single-stage target detection model YOLOv5 s is selected.Based on the network structure of YOLOv5 s,a network structure lightweight improvement algorithm is proposed,and the lightweight improvement of the model is achieved by adding Mobilenetv3 network and Bi-FPN bi-directional weighted feature structure,and the total number of parameters decreases compared with YOLOv5 s by The number of total parameters decreases by 30.7%,the number of floating-point operations decreases by 51%,and the m AP improves by0.4% compared with YOLOv5 s.(2)For the problem of noise reduction of the point cloud data of the bridge concrete structure surface cracks obtained by LIDAR,firstly,the point cloud data of the bridge concrete structure surface is filtered to remove a large number of outliers and noise points and to thin the density of the point cloud;the point cloud is aligned by using the normal distribution alignment algorithm NDT combined with the nearest iterative point alignment algorithm ICP,and the NDT attitude obtained from the NDT coarse alignment link is used as the initial value of the ICP fine alignment link.The NDT pose transformation matrix obtained from the NDT coarse alignment session is used as the initial value of the pose in the ICP fine alignment session,and the Gaussian Newton method is used to complete the output optimal transformation matrix of the ICP fine alignment session.The alignment results show that the root mean square error RMSE of the NDT-ICP algorithm is reduced by 51.4% compared with that of ICP,which meets the requirement of restoring 3D information of bridge concrete surface defects.(3)For the 3D point cloud bridge concrete surface defect segmentation and measurement problem,the curvature and normal vector of the points in the region are calculated by searching the nearest neighbor points using k-d tree,and the target point cloud is segmented into two parts,surface defect point cloud and background point cloud,based on the curvature and normal vector of each point using the region growth algorithm;the segmented defect point cloud is solved to solve the pose transformation matrix,and then projected onto the XOY plane to transform the 3D information measurement problem into a two-dimensional problem of calculating the maximum contour dimension information of the defect area;the length and width of the two-dimensional defect area point cloud data are calculated by using the OBB enclosing box method.(4)A LIDAR detection system was built,and several sets of bridge concrete surface defect data were obtained through outdoor experiments,and the collected defect point cloud data were processed according to the process framework of this paper;the impact on the measurement results was analyzed by adjusting the segmentation parameters for different defect data;finally,by detecting several sets of bridge concrete surface defect sizes and comparing them with the measured real values,the experimental results showed that For defects of length below 20 cm,the measurement error is about 3.5-6 mm,and the detection method proposed in this paper meets the requirements in terms of accuracy for the rapid detection of defects on the bridge concrete surface.