Research On Pavement Crack Detection Algorithms Based On Computer Vision

As the lifeblood of national economic development,highway transportation plays an important role in the development of transportation,manufacturing,tourism and agriculture and other national pillar industries.In recent years,with the extension of highway service life and the increase of traffic load,the development speed of pavement damage is accelerating.Pavement damage detection,maintenance and management has become the main task in the field of highway construction.Pavement cracks are the early manifestation of highway traffic pavement damage.Timely detection and maintenance of pavement cracks can avoid further impact of cracks on traffic safety,which has important practical significance.At present,the identification and extraction of pavement cracks is a key problem to be solved urgently in the field of pavement crack detection,and there are great challenges.On the one hand,due to the complex pavement conditions,texture,noise,shadow and other interference factors,there are a lot of misjudgments and omissions in the current pavement crack identification tasks.On the other hand,when there are known cracks on the pavement,how to eliminate the influence of interference factors and realize the automatic and accurate extraction of pavement cracks is another key problem in pavement crack detection.Therefore,this paper focuses on the theme of pavement crack detection,improving the accuracy of crack identification and extraction algorithm as a starting point,using the method of computer vision to study the on-line and off-line recognition of pavement cracks and the extraction of pavement cracks.The specific research contents are as follows:(1)In order to ensure the efficiency of the on-line identification method of pavement cracks under low-power platform and improve the identification accuracy of pavement cracks in complex topological structures,a Min Max k-Means pavement crack recognition algorithm based on prior knowledge is proposed.The algorithm can assign automatically correctable weights proportional to the variance of the cluster in the clustering process,and avoids the sensitivity of the traditional k-Means algorithm to the initial location of the cluster center by introducing prior knowledge.Compared with traditional k-Means algorithm and Min Max k-Means algorithm on the same data set,the accuracy and validity of the proposed algorithm are shown.(2)In view of the extensive application of deep learning technology in the field of target detection and recognition,good results have been achieved,in order to further improve the recognition accuracy,this paper applies deep learning technology to the off-line recognition task of pavement cracks whose hardware platform resources are no longer limited,and an off-line road crack recognition algorithm based on transfer learning and depth convolution neural network is proposed.Firstly,in the process of training and validating the model,data enhancement technology is used to expand the data set to improve the size and diversity of samples in the data set.Then,on the basis of VGG-Net network,a convolutional neural network model(Crack-CNN)is designed for off-line road crack recognition.Finally,in the training process of convolution network,feature transfer learning in isomorphic space is introduced,which further improves the recognition accuracy and generalization ability of Crack-CNN.(3)Aiming at the complex topological structure of cracks in the task of road crack extraction and the interference of repair marks,oil pollution,lane lines and illumination in actual road images,an automatic extraction algorithm of road cracks in structured random forest is proposed.Firstly,multi-level road crack feature information including HSV color feature,gradient magnitude feature and gradient histogram feature is obtained by integrating channel feature extraction method.Then structured random forest model is constructed by using the integral channel feature and the structured label feature of the reference image.Finally,using the integral channel characteristics of the extracted test set image,combined with the trained structured random forest model,the preliminary road crack extraction effect map is obtained.In order to further eliminate the influence of noise and other disturbances,the appropriate threshold is selected,the preliminary extraction effect map is binarized and the morphological method is used to remove noise,and the final road crack extraction effect map is obtained.Qualitative and quantitative analysis with minimum path selection,Canny edge detection,crack tree and iterative threshold segmentation shows that the proposed algorithm can detect cracks with arbitrary complex topological structure,and its accuracy and recall rate are better than other road surface crack extraction algorithms.