Research On Road Disease Recognition Based On Uav Continuous Road Image

In recent years,UAVs(Unmanned Aerial Vehicle)have been widely used in engineering geological surveys and road route selection due to their flexible manipulation and small size.By controlling the drone to collect multiple road images in a single flight,it can effectively improve the detection efficiency of road surface diseases and solve the problem that the detection range of the traditional detection vehicle is limited.This research takes the UAV as the road image acquisition device,and indicates the UAV continuous road image acquisition and stitching technology,automatic disease target detection of batch road images,pixel-level morphological feature extraction of road disease areas,and road disease size extraction.The above research results are comprehensively applied to form a set of road surface image collection and evaluation process.First of all,combined with the hardware parameters of the UAV acquisition device,the UAV flight height,UAV flight speed,sampling accuracy,sampling interval,shutter speed and other acquisition flight parameters are formally analyzed,and the UAV flight parameter calculation expressions and flight parameters are given.When performing the road image collection task,the flight parameters can be input according to the above results to improve the collection efficiency.On the other hand,based on the SIFT feature point matching algorithm,the serialized road images are stitched into continuous road images,and the redundant information of repeated area is removed under the condition that the diseased area is not damaged,forming a high-resolution unmanned aerial vehicle continuous road image database.Secondly,a disease target detection model of batch road images is established based on the deep neural network.The target detection model can automatically use the bounding box to demarcate the disease area for the detection set samples that have not participated in the training,including transverse cracks,longitudinal cracks,block cracks,maps Cracks,potholes.(1)the improved loss function formula is obtained by improving the position regression loss CIOU＿loss;(2)the original backbone feature extraction network CSPDarknet of Yolov5 is replaced with Efficient Net-B5 to form the optimization model of this research(Efficient Net-B5-Yolov5).Experiments reveal that the average precision,average recall and m AP: 0.5 of the optimized model in this study are 0.801,0.665,and 0.739,respectively,which are significantly better than the control model.Thirdly,the pixel-level segmentation of various types of diseases in road images is realized from two levels,namely automatic quantification extraction based on deep learning model and manual quantification extraction of variable parameter visualization.The first one is to establish VGG16-Unet and Resnet34-U-net based on the U-net semantic segmentation model,and conduct model training and test sample average accuracy test.The Dice and Jaccard of the VGG16-U-net model are better than those of the Resnet34-U-net model 16.75% higher,21.14% higher,respectively.The VGG16-Unet model outperforms the Resnet34-U-net model in the fine crack extraction ability.The second one is to develop a variable-parameter crack quantization and enhancement extraction interface based on the Opencv library in combination with the actual application conditions.Through mean filtering,bilateral filtering crack enhancement,USM image sharpening,Canny detection operator,morphological closing operation,and non-crack feature clusters Culling achieves crack extraction.By manually adjusting the slider,the parameter values of each crack extraction can be dynamically adjusted,the changing trend of crack extraction results can be observed,and real-time control can be achieved to achieve the optimal result of crack extraction.Finally,combined with the above-mentioned disease extraction model and the current road disease assessment specification,the actual size of the disease is determined according to the road size scale,and the DR formula of the road damage rate is adjusted.Form a UAV continuous road image disease library.