Pavement Surface Condition Index Detection System Based On Deep Learning

Deep learning is an emerging intelligent algorithm with widespread application in the field of intelligent transportation.However,research on intelligent detection of pavement defects using deep learning is relatively insufficient,and existing studies are still in the early stages and lack the ability to be applied in actual engineering,greatly limiting the improvement of intelligent detection of pavement defects.In view of this deficiency,this paper adopts the YOLO v5 target detection network,incorporating deformable convolution and multi-level attention block,and systematically explores the impact of various improved modules on the accuracy of the network.Based on the detection results of road defect detection,repeated detections of defects are eliminated using Euclidean distance and cosine similarity to estimate the actual size of pavement defects.Finally,a road defect detection system is developed using PyQT5,enabling the calculation of pavement surface condition index,and the system’s efficiency and effectiveness are verified through field testing.The main content and conclusion of this study are as follows:(1)By incorporating deformable convolution into the YOLO v5 model and replacing the CIoU loss function with the PIoU loss function in the network,the YOLO v5-DCN model was obtained.The testing results of the RDD 2020 and UAPD road damage databases showed that the accuracy of the YOLO v5-DCN model was higher than that of the YOLO v5 model in pavement damage detection.To further improve the ability of the model to detect pavement damages,this paper proposed the Multi-level Attention Block(MLAB)and obtained the improved YOLO v5-DCN model.A comparison of the MLAB attention mechanism with the Squeeze-and-Excitation Network(SENet),Efficient Channel Attention Network(ECA),and Convolutional Block Attention Module(CBAM)attention mechanisms showed that the MLAB mechanism significantly improved the accuracy of the four types of pavement damage:longitudinal crack,transverse crack,alligator crack,and pothole.(2)A repeated pavement damage database(RPDD)was constructed based on the pavement damage images provided by the on-board camera,which provided data support for the subsequent determination and verification of the parameters of the repeated pavement damage elimination algorithm.Based on the improved YOLO v5DCN model,this study innovatively transformed the problem of repeated detection of pavement damage into a sequence similarity problem,and proposed a repeated defect sequence elimination algorithm based on weighted Euclidean distance and cosine similarity.This algorithm can eliminate most of the repeated detected pavement damage.(3)To establish a connection between pixel blocks and actual sizes,a calibration plate with three different sizes(5cm/10cm/15cm)was used to calibrate the pixel block sizes.Then,the crack length detection algorithm based on a double polynomial fit was used to estimate the pixel lengths of transverse and longitudinal cracks.The algorithm first used SSR algorithm to enhance the texture of the cracks,then used double polynomial fit to remove noise points,and finally calculated the crack dimensions using the Gauss-Legendre formula with five nodes.Comparison with manual measurement results shown that this method had a high accuracy.In addition,the area of the damage detection box was used as the size of potholes and alligator cracks,which met the practical engineering needs.(4)Based on the PyQT5 interface design tool,the interactive software system was developed by integrating three algorithms:improved YOLO v5-DCN,repeated damage sequence elimination based on weighted Euclidean distance and cosine similarity,and crack length calculation based on double polynomial fit.This system was designed for different users and could rapidly evaluate the pavement surface condition index through actual pavement testing trials.