Road Crack Segmentation And Width Measurement Based On Binocular Stereo Vision

Road crack monitoring is an important part of road disaster warning.The development of deep learning technology represented by convolutional neural network provides a more intelligent and efficient segmentation method for road crack monitoring.However,the interclass imbalance and diversity characteristics of road crack lead to the problem of low road crack segmentation accuracy by the existing segmentation networks.Therefore,the CA attention mechanism is integrated,the multi-scale image input mode is designed,and the VGG block structure is optimized for the segmentation network to improve the segmentation accuracy of road cracks.Meanwhile,the existing image processing-based crack width measurement technology uses the image pixel as the measurement unit,which is difficult to be applied in practice because of the lack of accurate correspondence between pixel width and actual width for different camera shooting heights.To solve this problem,a road crack width measurement method based on the binocular stereo vision is designed,which can measure the actual width of cracks at different camera shooting heights more accurately,as follows:(1)Clearly labeled crack labels are made by the drawing software Procreate with a capacitive pen in this project to address the current problem that crack datasets are difficult to make and labels are not standardized.The crack dataset used in the experiment is expanded to increase the diversity of road cracks to solve the non-uniform pixel width distribution problem of the road crack dataset used in the crack segmentation network.(2)A CAMII-UNet(CA and Multiscale image input UNet)crack segmentation network is proposed to increase the precision of the crack segmentation network.The CAMII-UNet is based on a UNet with the encoder-decoder architecture.CA attention mechanism is integrated into the UNet to assign higher weight to the crack class,so as to improve the poor overall crack segmentation effect caused by the imbalance between classes in the road crack segmentation.The small crack segmentation ability of the network is previously poor due to the multi-scale change of cracks caused by the road crack diversity.Therefore,the VGG block is improved by an add feature fusion operation and the perceptual field of the network is expanded by a multiscale image input mode,so that the ability of the network to segment small cracks can be improved.The crack segmentation accuracy of CAMII-UNet has been improved compared with UNet.The global accuracy of crack segmentation increases to 98.99%.The average accuracy increases to 90.62% by 0.85%.The average intersection ratio increases to 84.44% by 1.47%.And the F1-score increases to 82.30% by 1.99%.(3)A road crack width measurement method based on binocular stereo vision is proposed to address the problem that the current crack width measurement is stuck in pixel-based width measurement and cannot be put into practical use.Firstly,the dual-view images of the road crack are captured by the binocular camera.Then,two pairs of matching points are obtained by using the binocular stereo matching algorithm and convex packet hull algorithm.Next,the pixel resolution is calculated according to the pixel distance in the left-view image and the real distance in 3D space between the two pairs of matching points.Finally,the pixel width of the segmented crack is measured by a hybrid width measurement method,and the actual width of the crack is calculated according to the pixel resolution.Experimental results show that the calculated width of the crack by this method is of high accuracy,and can be applied to the practical scenarios.The road crack segmentation and width measurement method based on binocular stereo vision proposed in this paper not only achieves 98.99% accuracy in crack segmentation,but also achieves 94.34% width measurement precision in the actual crack width measurement.The proposed method effectively solves the problems of low crack segmentation accuracy and poor ability to put pixel width measurement into practical use in the road crack monitoring.