| As the largest part of China’s transportation system,road transportation is of great significance to the development of the national economy and the daily life of the people.In the road transportation system,rural roads have the largest proportion,the worst pavement condition and the highest maintenance difficulty.Due to the low design strength and low construction budget at the time of paving,most of the rural roads have cracks,potholes and other diseases,which cause certain safety hazards to the local residents’ travel.Therefore,an efficient and convenient way to detect the diseases of rural roads can better evaluate the road conditions and complete reasonable and timely maintenance work.Therefore,in this paper,after reviewing a large number of references,it is chosen to apply UAV technology to highway pavement inspection and combine deep learning and image processing technologies to accomplish the task of pavement crack detection and crack parameter calculation.The main work is as follows:1.In this paper,two UAV devices,DJI Phantom 4 RTK and Mavic 2 Air,were used to complete the data acquisition.After several trials,85 m was chosen as the flight altitude,and the geometric correction of the images was completed using polynomial correction model and bilinear interpolation method to make the UAV images closer to the real scenes.2.Due to the high flight altitude and more interference options that can be easily identified as cracks in rural environment,this paper firstly carries out the segmentation and extraction of the road surface,and proposes a semantic segmentation algorithm based on the improved sobel operator by combining the respective characteristics of edge detection and semantic segmentation,which complements the semantic segmentation features extracted by the semantic segmentation sub-network with the edge features extracted by the edge detection sub-network.The average intersection ratio of 84.4% is achieved by using concat fusion of two features for convolution operation to obtain the final segmentation results.3.The extracted images containing only the road surface were used as the data set to complete the crack recognition and classification task.YOLO v5 m was selected as the recognition algorithm for crack recognition,and MPGA-BP,a BP neural network optimized by multiple swarm genetic algorithms,was used as the classification algorithm.83% weighted accuracy was achieved in the final experiment,with 87.2% accuracy for transverse cracks and 87.2% accuracy for longitudinal cracks achieved an accuracy of88.7%.4.To complete the segmentation task,convolutional layers were added to the HED network,and batch normalization layers were added to each convolutional layer,and guided filtering was applied to refine the segmentation results,so as to obtain a better and more accurate segmentation.86% or more segmentation accuracy was achieved in many publicly available pavement crack datasets,and then the segmentation task was completed after applying it to the a priori frame dataset generated by crack identification.The actual parameters corresponding to each pixel in the UAV image are obtained by placing anchor points at the experimental location for parameter calculation and corresponding them to the number of pixels in the acquired image,and then the specific parameters of the cracks are obtained,and the maximum error does not exceed 13% when comparing with the three cracks measured in the field,which proves the effectiveness and feasibility of this paper’s research.In summary,the model in this paper can better complete the task of detecting and calculating the parameters of pavement cracks,and the research in this paper proposes a new solution for the automatic detection and maintenance of rural roads using UAV technology and deep learning technology,which has high practical significance and research value.Figure [51] Table [7] Reference [99]... |
PDF Full Text Request