Research And Implementation Of Road Pothole Detection System Based On Deep Learning

In recent years,with the development and promotion of deep learning technology,image-based target detection technology plays an increasingly important role in smart home,smart driving and other fields.Pot-hole detection is a technology used to determine whether there are pot-holes in images or videos and to accurately locate the target.It has a broad development prospect in intelligent driving and other fields.Due to the existence of pot-holed targets in darkness,background clutter,difficulty in marking and diverse shapes,this paper selected YOLOv3 network as the basic training network after comprehensive comparison of common target detection algorithms,and conducted performance analysis on the training model of the original network,so as to improve the problem.The main work of this paper is as follows:(1)By studying and analyzing the shortcomings of the original YOLOv3 network in pothole target detection,this paper proposes three parts of improved experiments for this experimental data set.Firstly,GIoU is combined to compare the training models of three loss functions,namely,cross entropy loss,Focal loss and TSE loss.The experimental results show that the cross entropy loss function has the best effect.After that,the performance of GIoU,DIoU and CIoU is compared comprehensively,and it is finally found that the model based on CIoU training has the best performance in this experimental data set.Finally,this paper attempts to add additional convolutional layer to the detection module of YOLOv3 network,and the results show that this method improves the mAP value of the experimental model by 0.8%,with significant improvement effect.(2)Based on the detection model of the final experimental training,a pot-hole detection system is designed and developed.This system can complete the basic functions of login,registration,selection and prediction,and complete the system test using this experimental test set.