Road Pothole Detection Based On Deep Learning

As an important part of the assisted driving system,road pothole detection technology can provide efficient and practical assisted methods for assisted driving.There are various potholes in the road,showing uneven lighting,different shapes,and blocking.The traditional algorithm has certain limitations in detecting road potholes,which requires artificial design of pothole characteristics.The algorithm is cumbersome and complex,and the algorithm is not robust enough to accurately detect potholes.In recent years,with the continuous development of computer vision,the application of deep learning on it has surpassed the traditional image method.Faced with complex and variable road potholes,deep learning can automatically extract the features of the image,overcoming the shortcomings of traditional methods.Therefore,for the above-mentioned problems related to pothole detection,the main research is as follows:(1)This paper proposes a target detection method based on convolutional neural network to detect road potholes.For the research in this paper,the data of various situations of potholes are collected through manual collection and the Internet,and the data augmentation method of potholes is enriched by image rotation,salt and pepper noise,and cropped data augmentation methods.The A dataset and B dataset for pothole classification and the C dataset for pothole detection were constructed and labeled accordingly.(2)This paper proposes a classification method based on convolutional neural network to classify road pothole scenes.On the two basic classification networks,the road pothole training of the A dataset is used to obtain the best classification network Inception_v1;and Combined with the multi-scenario classification application of road potholes,the classification network was improved accordingly to obtain the Inception_v1.1 network,and the dataset B was used to improve the training of the network Inception_v1.1,and finally the pothole scenes were divided into one of a certain kind of pothole such as no pothole,strong light pothole,low light pothole.The accuracy of the classification model trained by it can reach 91%.(3)In order to further detect potholes,on the basis of road pothole scene classification,the improved multi-scene road pothole classification algorithm Inception_v1.1 and Faster RCNN fusion,that is,IFNet,after training on the C dataset,The average accuracy of IFNet for detecting potholes under strong light can reach 75.6%,and the average accuracy for detecting potholes under weak light is 88.1%.Therefore,the model can not only accurately and quickly identify the potholes of the road,but also detect the potholes and obtain specific location information.