Research And Implementation Of Road Anomaly Detection On Expressway

Expressways greatly facilitate people’s travel,promote economic and social development.The existence of various road anomalies in expressways not only threaten driving safety,but also shorten the service life of roads.Timely and accurate detection of road anomalies has important practical significance.Due to the influence of background changes,traditional digital image processing techniques have weak generalization ability for target recognition,while convolutional neural networks can more effectively solve this problem.Therefore,the use of convolutional neural networks for road anomaly detection is a trend in the development of road inspection systems.By defining the road anomalies of the expressway,road cracks,deformations,reveling,markings and debris are taken as detection objects,and establish road anomaly detection and segmentation data sets respectively.In order to recognize various road anomalies accurately and quickly,the deep learning object detection algorithm YOLOv5 is used as the basis for road anomaly detection research.Aiming at the problem of insufficient accuracy in target recognition caused by the variable anomalies scale of expressways,as well as low detection rate of smallsized targets in drone videos,by adding DS-GAM attention fused with depth separable convolution to achieve more refined feature extraction to improve the ability to identify road anomalies,introduce Softpool pooling to optimize the spatial pyramid pooling network to enhance the detection effect of the model on multi-scale targets,and adopt the feature pyramid network considering the scale sequence to improve the robustness of small-sized road anomaly detection.After experiments,the improved algorithm obtains higher accuracy on the highway road anomaly data sets compared with YOLOv5,which is suitable for the classification and identification of road anomalies in the UAV highway inspection system.To realize the quantitative evaluation of road anomalies,on the basis of detection,the Deep Labv3+ algorithm combined with codec and feature pyramid structure in semantic segmentation is firstly used to achieve pixel-level segmentation of road anomalies.In order to solve the problem of poor real-time performance caused by the complexity of algorithm,and inaccurate identification of road abnormal edge,the lightweight Mobilenetv2 is used as a feature extraction network to improve the segmentation efficiency of road anomalies,and SE attention is introduced at the end of the backbone network to make up for the negative impact on segmentation accuracy caused by the reduction in the number of model layers,improve the representation ability of road abnormal features,in the atrous space pyramid pooling network,apply GSConv convolution to multi-scale feature transformation,further reduce the complexity of the model,at the same time,use the multi-level features of the backbone network as the input of the 1decoding structure,through enriching the edge information of the target,improve the accuracy of multi-size road anomaly segmentation.On the basis of segmentation,according to the characteristics of various road anomalies,different types of anomalies are evaluated by the pixel length of cracks,the integrity of marking,and the pixel areas of reveling,deformation and debris.After experiments on the road anomaly segmentation data sets,the improved algorithm achieves more accurate anomaly segmentation and further evaluation results.Based on previous research,the expressways road anomaly detection system software is designed,and use Qt Designer to develop the visual interface to realize the detection and evaluation of road anomalies in the input inspection video,at the same time,store the results in the database.After testing,the system software can effectively improve the accuracy of road anomaly detection and segmentation,and then obtain reliable evaluation results,providing technical support for expressways inspection and maintenance,which has certain application value.