Research Of Vehicle-Mounted Automatic Pavement Crack Identification Technology Based On Semantic Segmentation

In recent years,the scale of highway infrastructure construction in China has been expanding,and the mileage of highway traffic has increased rapidly.Both the original roads and newly built roads need to be properly maintained.Efficient and accurate access to pavement disease information is the basis for scientific decision-making by highway maintenance departments.Crack is the most common disease on the pavement.In order to meet the needs of road maintenance work and improve maintenance efficiency,the rapid detection of cracks has become an important research topic.Aiming at the need of rapid detection of pavemnet cracks,this paper studies the automatic identification technology of vehicle-mounted pavement crack based on semantic segmentation.Firstly,by analyzing the functional requirements of the pavement crack detection system,the overall design scheme of the vehicle-mounted pavement crack image acquisition system is proposed.The vehicle-mounted pavement crack image acquisition system is designed,and the hardware system of image acquisition,auxiliary lighting system and the software of pavement crack image acquisition are developed.The developed auxiliary lighting system is an LED modular spotlight that is illuminated in a tilted manner.Then,a network model for automatic pavement crack identification is designed based on the semantic segmentation and the Deeplab V3+ model.The basic network of this model is Resnet152.The group normalization method of grouping in the channel direction of the feature graph is used to make the model achieve higher accuracy with smaller batch size.Attention optimization mechanism is added to the model.Global context information is obtained by global average pooling.By calculating the weight vector,the target crack is assigned a larger weight,which makes the model more favorable for crack identification.Finally,the pave crack image data set is constructed.The data set contains interference information such as pavement marking lines,oil stains,brake marks and so on.The training set of the data set is used to train the pavement crack identification network model that based on semantic segmentation.The validation set is used to validate the training effect of the model.The validation results show that the trained model can segment the cracks and background well,and the miou is 81.31%.Experiments show that the proposed method based on semantic segmentation has better crack identification effect and can provide useful reference for automatic crack identification of pavement.