Research On Road Disease Visual Detection Method For Road Inspection UAV

Road disease poses a serious potential threat to the performance,life and driving safety of the road.Therefore,timely and accurate detection of road disease and implementation of repairs are of great significance to ensure the quality of transportation and reduce road maintenance costs.Existing road defect detection algorithms are mainly based on machine vision technology,and most of the algorithms are applied to image collection by manual or vehicle-mounted cameras.However,with the emergence of road inspection UAV acquisition methods with high degree of automation and low power consumption,due to the differences in shooting height and viewing angle,existing algorithms cannot be applied to the task of disease detection in aerial road images.Aiming at the road images taken by road inspection UAV,combined with conventional image processing and deep learning technology,two disease detection algorithms for aerial road images are proposed,and visual road disease detection software is designed based on the algorithm with better detection effect.The main work of the paper is as follows:(1)Firstly,UAVs are used to collect real road disease images including light,stains and vehicles in different regions at different time periods.Then,LabelMe is used to label the disease at pixel level.Finally,the disease images are cropped using the overlapping sliding window method,combined with public datasets and data augmentation methods are applied to create a complex and diverse aerial road disease dataset.(2)Aiming at the problem that the complex background in the aerial road image is easy to interfere with the disease detection,a disease detection algorithm based on the extraction of lane area and lane line area is proposed.In order to avoid the interference of lane lines on the detection of disease in the lane area,firstly,using the color prior information and complete edge information of the lane area in the aerial road image,a method of integrating color and edge information is proposed to realize the lane area extraction with smooth edge and complete area.Then,with the aid of the lane line extraction algorithm,the road disease detection is completed in the lane area and the lane line area by using the adaptive threshold method.The algorithm is verified by using 100 aerial road images in different scenarios.Experimental results show that the accuracy of the algorithm for disease image screening is 86.00%,and the accuracy of disease detection is 58.72%.However,the algorithm needs to reset the detection parameters for different road sections,and its robustness is not good.(3)In order to solve the problem of poor robustness of conventional image processing algorithms,a road disease detection model based on improved DeepLabv3+ is proposed.In order to reduce the amount of model parameters and improve the efficiency of model training and inference,MobileNetv3 is used as the backbone feature extraction network of the model,and the Ghost module is used to replace the ordinary convolution in the atrous spatial pyramid pooling(ASPP)module.To avoid degrading the model accuracy by replacing the backbone network,the following measures were adopted.Firstly,the strip pooling module was used to replace the global average pooling in the ASPP module.Secondly,a lightweight attention mechanism ECA(Efficient Channel Attention)module is introduced before the ASPP module and in the decoding part,and a shallow feature fusion structure is designed after the shallow feature map.Finally,construct a hybrid loss function and adopt a training method based on transfer learning.Compared with the traditional DeepLabv3+ model,the parameters of this model are reduced by 194.93 MB,which is reduced to 1/14.20 of the original,and the mean intersection of union and average frames per second in the test set are increased by 0.54% and21.54 frames per second,respectively.The model is used to detect 100 aerial road images.The experimental results show that the accuracy of the model for disease image screening is 96.00%,and the accuracy of disease detection is 80.73%.It has good adaptability to different road sections and is better than conventional image processing algorithm.(4)Based on the improved DeepLabv3+ road disease detection model,a visual road disease detection software was developed.And the comprehensive test of the software has laid the foundation for the actual engineering application.