Research On Improvement Of YOLO Algorithm For Pavement Distress Detection

Since 2010,the total mileage of highways in my country has been increasing year by year.As of the end of 2019,the total mileage of highways nationwide reached 5.0125 million kilometers.Due to long-term complex weather conditions and heavy load environments,road pavement diseases frequently occur.In 2019,my country’s road maintenance mileage reached 4.9531 million kilometers,accounting for 98.8% of the total road mileage.At the same time,in the field of road disease detection,traditional human detection and digital image processing methods are costly,time-consuming,and low in accuracy.In extreme cases,they may even affect normal traffic driving,and are becoming increasingly unsuitable for road disease detection.In response to the above problems,a road disease detection method based on deep learning is proposed.It uses the one-stage target detection YOLO(You Only Look Once)algorithm that has both speed and accuracy in the target detection field,and designs and implements a network model.Detection of pavement diseases.In order to improve the accuracy of the YOLO model,first use the Kmeans algorithm to cluster the labeled bounding boxes to obtain a more suitable a priori box;the IOU(Intersection Over Union)index in the algorithm loss function cannot accurately reflect the two For the problem of the degree of overlap between the boxes,it is proposed to use the CIOU(Complete-IOU)loss function to replace the regression loss part of the original loss function.The CIOU loss function reduces the overlap area of the two boxes,the center point distance and the aspect ratio consistency.Taking into account,the regression process of the target box is more stable;for the problem that the Leaky Re LU activation function in the algorithm is not smooth enough and the information is difficult to flow in depth,the Mish activation function is used to replace it;finally,in order to better combine the local features with the global features Integration,drawing on the idea of spatial pyramid,embedding SPP(Spatial Pyramid Pooling)structure in the model,enriching the expressive ability of feature maps.Use the improved YOLOv3 algorithm to treat the most common diseases on the pavement,including horizontal cracks,longitudinal cracks,mesh cracks,loose subsidence,etc.,as well as special pavement facilities,including blurred white lines on pedestrian crossings,blurred white lines on road signs,and public Facility holes are detected.The experimental results show that the m AP value of the improved YOLOv3 algorithm on the road disease data set has increased from 50.53% to 57.41% compared with the original algorithm.The accuracy has been improved,which verifies the effectiveness of the improved strategy.The latter algorithm can better meet the needs of road disease detection.