Study On Road Surface Damage Detection Based On Deep Learning And Binocular Stereo Vision

Road damage detection has important value in maintaining roads and reducing road maintenance costs and the incidence of traffic accidents.Traditional road damage detection methods mainly rely on manual identification of road damage.Although existing image-based road surface damage detection methods can improve the automation and efficiency of road damage detection to a certain extent,they are susceptible to complex backgrounds such as water stains,shadows,and chromatic aberrations,resulting in poor anti-interference performance in the detection process,which restricts the improvement of road damage recognition accuracy.This paper adopts a road pavement damage detection method that couples deep learning and stereo vision.The method is based on the advanced semantic feature extraction capabilities of deep learning and binocular stereo vision technology to realize the detection of road damage from 2D to 3D.The experimental results show that the method in this paper has good applicability and certain practical value for damage detection in road images.The main research contents of this paper are as follows:(1)Use the open source road data set to expand the sample set used to train the road damage recognition model by using data enhancement methods such as image random rotation,adjusting brightness and contrast,and labeling damage(potholes,cracks,etc.)and undamaged samples.Combined with transfer learning,the Alex Net,VGG16,and Goog Le Net pre-training models are used on the sample set for supervised evolutionary learning,all layer parameters are fine-tuned,and the road surface damage recognition model is constructed.(2)Using a multi-scale SGM binocular stereo vision dense matching algorithm,introducing a three-layer image pyramid matching strategy,by using the disparity range in the pixel neighborhood window of the upper layer image to constrain the disparity search range of the next layer image,it can Effectively improve the matching speed and accuracy.In addition,by performing quadratic polynomial fitting on the road disparity map,the original disparity map is flattened to identify the damaged area from the disparity map.(3)The experimental results in this paper show that although the road damage recognition model has a high damage recognition accuracy,the model may mistakenly recognize the image background(water stains,shadows,etc.)as damage.The damage detection can be performed by coupled binocular stereo vision technology.In excluding the influence of the complex background of the image,the road damage can be accurately detected,but in the wrong area of the disparity map,it will cause misdetection.Deep learning can accurately identify this type of road image.Therefore,the two detection methods are coupled,and the advantages and disadvantages are complementarity that can improve the accuracy of road surface damage recognition.