Research On Road Surface Condition Recognition Using Deep Semantic Segmentation Network

The enlargement of the scale of highway transportation system has brought tremendous development of community economy as well as increasingly serious traffic accidents which are caused by meteorological factors such as heavy rainfall,snowfall,road ice formation.Therefore,real-time monitoring and accurate identification of unfavorable road surface conditions under severe weather then warning early and response timely have great practical significance for ensuring the safe of highway system.The existing pavement state recognition technology can be roughly divided into two types: contact and non-contact type.The contact methods have limited effect in practical applications due to shortcomings such as high installation and maintenance cost and small detection range.Among the non-contact ones,the image-based method has become a hot issue with the widespread road monitoring system.Using segmentation algorithm to fine-grained partition of mixed pavement state area is a new research direction,which has good practical value and application prospects.The thesis aims to recognize the road surface condition by processing the image data,and uses the frontier deep learning semantic segmentation algorithm to identify and divide the complex mixed road condition regions accurately.Firstly,the basic principle and network of semantic segmentation algorithm based on deep learning are studied.Then,after comparing precision and performance of each algorithm on the self-made hybrid road condition recognition dataset,we choose U-Net which has high accuracy and real-time performance as the backbone of the algorithm.In the task of road surface condition recognition,the categories are highly similar so larger range and multi-level semantic information is needed.For this reason,we propose improvement schemes from the integration context information and multi-scale information fusion,and design the D-UNet network based on cascaded atrous convolution and the ASPP-UNet network based on the improved atrous spatial pyramid pooling module to further improve the accuracy of the algorithm.In the thesis,a series of experiments are carried out on the self-made dataset and appropriate parameter settings are found.Finally,the mean pixel accuracy and the mean intersection of union can reach 87.28% and 79.30% respectively.The effectiveness of the improved schemes is verified,which provides a practical reference for mixed pavement status recognition.