Research On Key Technology Of Pavement Crack Survey Based On 3D Point Cloud And Image Data

The pavement crack survey has gained much attention among the transportation departments.With the rapid development of computer vision technology,the crack detection based on image technology has been widely used in intelligent transportation system detection.In order to improve the crack survey accuracy and achieve satisifying results,this reseacrh studies on the key technology of crack survey based on 3D point cloud data(PCD)and image data.This research mainly contains the following aspects:(1)Focusing on the noise characteristics in the pavement PCD data,a filtering processing method is put forward in this paper.The different filtering methods are applied to remove the large-scale and small-scale noise respectively.For the large-scale noise in the 3D data,the proposed method applies the statistical analysis and comparison on the distance among the related neighborhood to remove the outliers.For the small-scale noise in the 3D data,a filtering method based on bilateral filtering is put forward.According to the PCD surface information(x and y axis)and the deep information(z axis),the 3D filtering factors for pavement data can be obtained based on the plane proximity factor and deep similarity factor.Experiments show that the proposed filtering method can eliminate the noise influence which lay the foundation for subsequent crack extraction and measurement.(2)Aiming at the edge feature of crack information in pavement PCD data,a 3D crack extraction algorithm based on fractional differential is proposed.Different from the traditional edge detection based on integer order differential,the fractional differential operator can effectively extract high-frequency information while preserving low-frequency information.The proposed 3D crack extraction algorithm is based on the Tiansi template of G-L fractional differential definition,which could extract crack and retain texture information effectively by a nonlinear stretching operator on the differential coefficients.Compared with the classical edge detection algorithms and other popular crack extraction algorithms by Precision-Recall system,the experiment verifies the proposed algorithm can extract 3D crack information effectively.(3)On the basis of the extracted cracks,a crack form analysis method is proposed to calculate the crack 3D parameters including the length,width and deep.Firstly,the best fit rectangle method based on sectional measurement is proposed to calculate the length and width.Secondly,a plane fitting algorithm based on RANSANC is proposed to calculate the crack depth information,which can accomplish plane calculation with errors in three directions.By fitting the up and down planes of each segment respectively,the deep information can be estimated by the distance between the up and down planes.Finally,a crack classification model is designed after analyzing the feature of different crack types.This model contains the crack area ratio and the crack distribution density to classify the cracks into simple cracks and complex cracks.(4)A classification method based on convolutional neural network is designed and implemented for pavement crack images.The proposed model is composed of one input layer,two convolutional layer,two pooling layer,one full connected layer and an output layer.In this model,the ReLU function is used as the activation function,the maximum pooling method is used for dimensionality reduction,and the Softmax classifier is used to divide the crack images into three types: transverse,longitudinal,complex cracks(including reticulation and block).In the experiment,4000 crack images were selected for training,and 1000 crack images for testing.The experimental results show that the the proposed method can get good classification results.