Research On Vehicle-mounted Pavement Disease Automatic Recognition System And Method Based On Multi-class Deep Semantic Segmentation

Nowadays,the highway network makes people’s travel very convenient and also brings great economic benefits to the development of the automobile industry.However,as the total mileage of roads has increased,road maintenance tasks have gradually increased,Traditional manual inspection methods have been difficult to meet such huge inspection needs,Although data collection has been automated in actual road inspections now,but the collected data still requires a lot of manual processing,especially in the automatic detection of road disease images.Therefore,under the premise of ensuring the accuracy,the realization of fully automatic road disease image recognition is a problem that needs to be solved urgently.The existing automatic detection technology of road diseases has problems such as low recognition accuracy,insufficient positioning accuracy of road diseases,and incomplete detection categories of road diseases.To solve the above problems,this paper proposes a road disease automatic recognition method based on multi-class depth semantic segmentation combined with fully connected conditional random field.The main research content of this paper are as follows:Firstly,when the semantic segmentation method is used for the research of automatic road disease recognition technology,a large amount of high-quality road image data accurate to the pixel level and information such as station number and mileage corresponding to the image are required.So this article proposes the overall design plan of the vehicle-mounted road image acquisition system,system hardware such as camera,lens,encoder and lighting equipment was selected,and road image acquisition software was developed.Seconly,Most of the pavement cracks are long and narrow structures.As a kind of supervised learning,neural network is very dependent on the label information of data set.However,the shape of the reticulated cracks image is complex and changeable,and it is particularly difficult to label.This paper proposes a semi-automatic labeling method based on DBSCAN(Density-Based Spatial Clustering of Applications with Noise)density clustering and binary classification semantic segmentation network.On the one hand,this method can reduce human involvement and improve labeling efficiency,on the other hand,it can also improve labeling accuracy.Finally,The original UNet network model is improved in many ways by combining the characteristics of road image data sets such as similarity,sample imbalance,and gray-scale difference.It also introduces post-processing optimization based on Dense CRF(Dense Conditional Random Fields).While using the deep semantic segmentation network to obtain features,Dense CRF is used to obtain more spatial context information.In this way,noise is further suppressed,segmentation accuracy is improved,and training errors caused by manual labeling are eliminated.The experimental results show that in the test set of other road sections collected by the same equipment,the improved UNet network accuracy rate reaches 98.43%,miou reaches82.74%,the detection result is good,and the generalization performance is good.In summary,the thesis researches and implements the automatic recognition technology of vehicle road diseases based on multi-category deep semantic segmentation and Dense CRF,which lays the foundation for the project research and provides a certain reference for further road maintenance decision-making.