Semantic Segmentation Of Road Potholes Based On Deep Learning

Road pothole is a common road disease phenomenon,easy to induce traffic accidents in the process of transportation,in the wet road after rain and snow weather is more serious.Unmanned driving is the development direction of transportation and automobile industry in the future.In order to realize the rapid and accurate detection of road potholes in the process of automobile driving,so as to reduce the occurrence of traffic accidents and protect people’s life and property safety.This paper studies the method of semantic segmentation of road pothole images based on U-Net network model in deep learning,but the current research does not fully consider the problem of low classification accuracy caused by the shallow level of U-Net feature extraction network that cannot represent pothole features in complex environment.At the same time,the classified road pothole Mask image has some problems,such as misclassification,pothole feature leakage and uneven segmentation of pothole boundary.In view of the problems related to road pothole detection described above,the main research is as follows:(1)on the slippery road potholes data acquisition and data image used in this article are from onboard camera height reduction the self-driving car body surrounding environment information,the performance of the car while driving more real cameras can perceive all traffic information,and combining the data enhancement technique to make semantic segmentation data sets,improved the image expression ability,Avoid the phenomenon of over-fitting in the network training process.(2)It is proposed that U-Net network is improved based on ResNet-34 feature extraction network to build a deeper D-UNet semantic segmentation network model.The residual learning structure of ResNet34 feature extraction network can effectively avoid the phenomenon of gradient disappearance and gradient explosion when extracting road pothole features.More conducive to the extraction of potholes after rain and snow weather targets.(3)The segmentation results are incomplete due to the problems of small misclassification phenomenon,leakage phenomenon and uneven boundary of potholes in wet road pothole images.In this paper,full-connected CRFswas introduced to post-process the segmented road pothole Mask image,namely,D-UNet+CRFsnetwork model was built by combining D-UNet network structure with fullconnected CRFs,and the grid parameter search method was used to obtain the optimal parameters of fullconnected CRFs to achieve the best post-processing effect.The experimental results show that the improved D-UNET+CRFs semantic segmentation network model has an average detection accuracy of 90.33%,which is 5.99%higher than the improved U-NET network model.It can effectively perform semantic segmentation on wet road potholes after rain and snow weather.To achieve pixel-level identification accuracy.