Research On Detection Algorithm Of Cement Pavement Disease In Agricultural And Pastoral Areas

Pavement inspection is an important part of maintaining road safety.Considering the low efficiency and poor accuracy of traditional pavement inspection methods,and that most of the inspection methods are for high-grade pavements,an efficient,safe and economical inspection method for agricultural and pastoral roads is urgently needed.At present,the pavement disease detection technology based on deep learning has high detection accuracy,high detection efficiency and good operational safety.In this thesis,based on the target detection technology in deep learning,the following research is carried out around the detection method of cement concrete pavement disease in agricultural and pastoral areas:(1)The pavement disease detection system is designed;and for the task of UAV-based pavement disease detection,the UAV,gimbal camera,acquisition card and other equipment are selected;the collected images are pre-processed,and the data sets of cement concrete pavement disease in agricultural and pastoral areas based on horizontal frames and based on rotating frames are labelled according to the relevant specifications and expert opinions,respectively.(2)To address the problems of inaccurate target detection frames,complex road area background under UAV view,and large changes in disease scale,an improved YOLOv5-based method for detecting cement concrete pavement diseases in agricultural and pastoral areas is proposed.The mosaic stitching and other methods are used to expand the data set in the training phase;the ECIo U loss function is constructed to improve the accuracy of the model prediction framework without increasing the model complexity;compared with CBAM and Transformer,the model with the addition of coordinate attention mechanism has higher detection accuracy and improves the ability of the model to accurately extract the disease in the complex background;by introducing the ASFF adaptive spatial feature fusion mechanism to further improve the ability of the model to detect multi-scale diseases.It is found that the m AP value of the improved YOLOv5 disease detection model reaches 0.857,which is 8.9% higher than the benchmark model,where the AP values of broken plates and cracks are 0.958 and 0.824,respectively,and the algorithm has higher accuracy and better detection speed than the mainstream target detection algorithm.(3)The Rep Ghost-YOLOv5 model was constructed to further improve the detection speed of cement concrete pavement diseases in agricultural and pastoral areas.Compared with the benchmark model,the detection frame rate was improved by 2.97 frames/s and the model complexity was reduced by 17.1% under the premise of 1.8% improvement in m AP.(4)In order to realize the dimensional quantification of pavement defects,a rotating frame-based target detection algorithm was constructed;a rotating frame-based UAV cement concrete pavement defect dataset was established using the rolabel Img open source tool;the mosaic method and the attention mechanism were used to improve the detection performance of the model,and the study showed that: both improvement methods had a positive impact on improving the rotating frame detection algorithm The relative error for quantifying the defect size is less than 5.42%,and the purpose of accurately detecting the pavement defect size is achieved.(5)To solve the problem of repeated disease detection in continuous-frame images and to achieve the quantification of disease size,a Deep SORT-based disease size counting method is constructed.The accuracy of the disease quantity statistics is verified by experiments to reach 90%,which achieves the purpose of avoiding the repeated detection of the disease,and the proposed pavement disease statistics method provides a reference for pavement maintenance.