Research on Road Pothole Detection Based on Semantic Segmentation
Road pothole detection holds significant importance for driving safety,travel comfort,and the overall appearance of roadways.Traditional methods of pothole detection typically only delineate the general location of potholes without clearly defining the extent of the affected areas,often resulting in poor visual representation and ineffective detection.In response to these challenges,a novel detection network based on semantic segmentation is proposed in this paper.This network enhances the original U-Net architecture by replacing its down-sampling component with VggNet,thereby deepening the network and introducing additional nonlinear factors.This modification allows for a superior fitting of real-world data distributions,enhances generalization capabilities,and simplifies both the model structure and computational processes.A new combined loss function is employed to accelerate network convergence and improve prediction accuracy.Experimental results indicate that,in comparison to traditional U-Net,PSPNet,and Deeplab networks,the proposed network achieves an improvement of more than 5.12%in the average Intersection over Union(IoU)for pothole segmentation images,an increase of over 4.95%in average pixel accuracy across categories,and enhancements of more than 1.2%in precision and 4.95%in recall.These improvements substantiate the efficacy of the proposed approach,providing a valuable reference for road pothole detection.
semantic segmentationroad pothole detectiondeep learningcombined loss function
国家科技期刊平台
NSTL
万方数据
