Research On Embedded Vehicle Detection System Of Urban Road Pavement Diseases

In recent years,with the fast improvement of pavement construction,the problem of pavement damage has become increasingly serious,such as pavement subsidence,cracks,network cracks,pits,etc.As an important part of road maintenance,pavement disease detection is one of the hotspots in this field.Because there are personal safety problems and hidden dangers of traffic jam in manual inspection,it is particularly important to develop an automatic recognition system of road diseases based on machine vision.Based on the analysis of the actual needs of urban pavement disease detection task,combined with the convolutional neural network technology,this paper adopts the embedded vehicle detection technology based on YOLOv5 to achieve the detection of pavement disease.The main work of this paper is described as follows:(1)This paper reconstructs a data set of urban pavement disease detection,including five types of diseases under different interference environments,namely,well cover,crack,road marking,pit and groove,and grid crack,totaling more than 8400 pieces.The data enhancement is performed on the basic data set to simulate the special situation of interference in the sampling process in the real scene,change the brightness of the image to simulate the special situation in bad weather,and mark five types of targets in the image,namely,uneven manhole cover,cracks,fade marks,grid cracks and pits.This data set will be mainly used for the target detection and recognition research in the field of pavement disease detection.(2)This paper summarizes the existing target detection algorithms,and proposes a method to replace the traditional convolution layer of YOLOv5 with the Ghost module,and uses Mosaic data enhancement to improve the detect ability of the model in view of the large number of parameters of YOLOv5 model and the large number of floating point operations.To verify that the algorithm proposed in this paper has higher detection performance,YOLOv5 s is used,and the model of replacing YOLO-v5 s backbone network with Shuffle Netv2 and Mobile Netv3 is used.Compared with YOLOv5 model,the average precision of the model proposed in this paper is 88.17%,increased by 4.01%,meeting the detection requirements of pavement diseases.(3)Combining the memory requirements of image computing and the ultimate goal of pavement disease detection task,a pavement disease detection system suitable for embedded device Jetson Nano 4G is designed.The device can realize real-time detection of pavement disease,disease location,disease information recording and other functions through the combined application of GPS positioning,image processing methods,depth learning technology,digital cameras,etc.The image computing memory is kept below 4GB,The detection speed reaches 12.51 frames/second,which is 184% faster than YOLOv5 model.Compared with other detection methods,it can achieve faster operation,better detection performance and better applicability.