Research On Pavement Pothole Detection Based On Deep Learning

Pavement diseases such as cracks,potholes,etc.,especially potholes affect the aesthetics of the pavement,or reduce the service life of the pavement and destroy the safety and comfort of the pavement,and need to be found in time to protect as early as possible.Timely inspection and maintenance of the pavement can be found in the early stage of pavement disease and timely treatment of the disease,to prevent subsequent deterioration to bring greater maintenance expenses.However,pavement potholes are characterized by variable shapes,complex backgrounds and varying depths,making detection difficult.The current deep learning-based target detection technology has played a significant role in several fields and is suitable as a research solution for pavement pothole detection.However,target detection techniques based on deep learning networks require a large amount of data involved in training to ensure that the network fully learns the features of the detection target and has a stronger generalization capability in practical applications.However,in the existing research on pothole detection,the datasets are from a single source and are small in number,resulting in less than ideal accuracy and generalization performance of the detection method.In this paper,based on the above problems and the actual needs of pavement pothole detection,the following research work is carried out:(1)To address the problem of lack of resources in the dataset of pavement pothole images,this paper collects data by three ways: collating public datasets,web crawling pothole images,and taking pavement pothole images by oneself,and carrying out data enhancement to expand the number of samples in the dataset.For the current problem of consuming a lot of human and material resources for pavement potholes image annotation,this paper adopts a semi-automatic annotation method,which greatly improves the efficiency of dataset production.(2)To address the difficulties of pavement pothole target detection,such as irregular shape of the detection target,non-uniform road surface material,and complex environment of clutter interference,this paper conducts experiments to compare the training indexes of Faster R-CNN,YOLOv7 and SSD in the pothole dataset,and finds that the m AP of YOLOv7 algorithm reaches 89.3%,which is more accurate and faster among the three network models.higher and faster,which is suitable as a baseline network model for pavement pothole detection.(3)To address the problem that small target potholes on the road surface are prone to false and missed detection,we choose to add SE attention mechanism to ELAN and SPPCSPC of YOLOv7 network,and experimentally find that the m AP of YOLOv7＿A network reaches 91.7%.To address the limitations of the localization loss CIo U loss function in YOLOv7 and the imbalance of positive and negative samples in the pit dataset,the Focal-EIo U loss function is chosen to replace the CIo U function on the basis of YOLOv7＿A to obtain the improved network YOLOv7＿A＿EIo U.The YOLOv7＿A＿EIo U network m AP reaches 93.6%,compared with YOLOv7＿A network improved by 1.9.The problem of insufficient memory and insufficient computing power to implement the detection algorithm for edge device deployment is addressed.This paper introduces the Mobile One module based on the YOLOv7-tiny model to make the network lighter.The file size of the improved network model is 14.7 MB and the m AP decreases by 0.4 percentage points compared to 85.4% of YOLOv7-tiny,which is an improved solution that uses a reasonable loss of accuracy in exchange for a lighter network.(4)Finally,a series of experiments are designed to verify the detection capability of the algorithms in this paper for real-world situations,and a pavement pothole detection system is designed and implemented using Py Qt5.The system includes picture detection and video detection functions,and supports users to upload pavement pictures or videos for pothole detection and save the results to the specified location.