Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The existing pavement crack detection and information extraction method mainly comprises the following steps: the pavement crack detection method based on the two-dimensional visual characteristic comprises the following steps: the pavement crack detection method based on the two-dimensional visual features mainly obtains pavement data through an optical camera, a video and the like, and analyzes and extracts the pavement crack through the edge and gray characteristic of the pavement crack and the difference of a pavement background. For example, the gray level of the crack is lower than the background of the road surface, a crack seed region is extracted, and then the shape characteristic of the crack is combined for judgment; and acquiring crack information and the like in the two-dimensional data by using an edge detection operator commonly used in image processing in combination with the edge characteristics of the crack.
The method for extracting the pavement cracks based on the two-dimensional visual characteristics has the following defects: the pavement crack extraction method based on the two-dimensional visual features mainly obtains pavement data through an optical camera, a video and the like, extracts the pavement data through the gray scale and the shape characteristic of the pavement crack and the gray scale or the shape difference between other indexes of the pavement, for example, extracts a crack seed area by utilizing the fact that the gray scale of the crack is lower than the background of the pavement and then judges the crack by combining the shape characteristic of the crack; and acquiring crack information and the like in the two-dimensional data by using an edge detection operator commonly used in image processing in combination with the edge characteristics of the crack. The method cannot overcome the influence of ambient light, shadow, road surface tire grinding marks, oil stains and the like on the marking detection, and has limited applicability.
The pavement crack detection method based on the three-dimensional pavement data comprises the following steps: such methods typically use vehicle-mounted line scanning three-dimensional data, combined with three-dimensional data accuracy (lateral resolution 1mm), to detect using the characteristic that the elevation of a pavement crack is typically lower than the background of the pavement. For example, at the level of cross section of the collected elevation data, acquiring a fracture seed area by utilizing the sharp V-shaped characteristic of the fracture area expressed in the elevation section data, and judging by combining the shape characteristic of the fracture; and extracting cracks at the elevation data point cloud level by using a local threshold or sparse expression method.
The three-dimensional road surface data-based road surface detection method has the following defects: such methods typically use on-board three-dimensional laser scanning data in conjunction with data accuracy (lateral resolution 1mm) to detect pavement crack location information. Although the method can overcome the defect that the traditional two-dimensional gray scale image method is easily influenced by illumination and shadow, the existing method for detecting cracks by using three-dimensional characteristics of the road surface also has the following defects: when the precision of the three-dimensional data is high enough (the transverse resolution is 1mm, and the elevation resolution is less than or equal to 0.5mm), the three-dimensional pavement elevation data contains more complex pavement scene information, not only cracks, but also pavement deformation, marking, repairing and pavement texture; in the high-precision data, different types of pavement diseases or indexes have certain influence on crack extraction. For example, the texture fluctuation of a pavement with thicker texture is similar to the depth characteristic of the crack, and the robustness and the practicability of the crack extraction method can be influenced by only utilizing the crack depth characteristic without considering the structural depth influence of the pavement.
In the cross section of the line scanning three-dimensional pavement data, cracks are interfered by various factors, edge clearness and V-shaped structures are not always presented, and more, due to the interference of a pavement texture background, a plurality of V-shaped structures in the data are not real crack positions. Due to the influence of factors such as driving posture, deformation diseases, pavement material abrasion, crack types and pavement background difference, cracks in the cross section data of the line scanning three-dimensional pavement data are easy to present asymmetric V-shaped or non-V-shaped structures with unclear edges. Therefore, for the crack extraction of line scanning three-dimensional data, the traditional template matching and edge detection methods are difficult to obtain ideal crack detection effects.
In addition, a mature method is provided for fully utilizing different source data samples so as to achieve the purpose of extracting the three-dimensional pavement data cracks by machine learning method online scanning. In practical application, as more and more systems can acquire three-dimensional data, the difference of the background of the pavement texture and the difference of the crack types contained in the acquired data are gradually highlighted, the existing machine learning method is difficult to extract the cracks of different pavement backgrounds and different crack types under the condition that enough labeled samples are not available, and the defect greatly limits the practical application of machine learning and even deep learning in online scanning of the three-dimensional pavement data.
In view of the above problems, embodiments of the present invention first provide a point-by-point multi-window accumulated difference feature, which utilizes the characteristic that the fluctuation characteristic of the road surface texture is stable in a certain range, and the section part with cracks has an elevation fluctuation trend to perform modeling, so that while overcoming the noise and background difference, the obtained templates can be clustered among homologous data or sample shared among different source data, thereby reducing the threshold setting or the dependence on homologous labeled samples in the crack extraction process.
On the basis of the point-by-point multi-window accumulated difference features, a line scanning three-dimensional pavement data crack extraction technical route is realized, and particularly, under the non-supervision condition (no sample mark exists), the three-dimensional pavement non-supervision crack information extraction is realized by utilizing the self aggregation of the point-by-point multi-window accumulated difference features and a typical non-supervision machine learning method, such as Kmeans. Under the condition that a small amount of crack labels exist, label information and the proposed point-by-point multi-window accumulated difference features are trained to form a feature template, different source samples can be shared, the requirement of a supervision machine learning method on sample labeling is reduced, and therefore a three-dimensional pavement crack extraction result is obtained more accurately and rapidly.
In addition, the provided point-by-point multi-window accumulated difference feature-based technical route has high adaptability to the existing unsupervised and supervised machine learning methods, and the characteristics of introduction of the efficient machine learning method and sample sharing capability are favorable for realizing large-scale three-dimensional pavement crack extraction under a small amount of labeled samples, so that a stable and robust method is provided for actual pavement crack detection.
Fig. 1 is a flowchart of a fracture extraction method with multiple windows accumulated difference according to an embodiment of the present invention, as shown in fig. 1, the method includes:
s1, acquiring three-dimensional road surface data of the target road surface after the attitude and deformation information is removed;
the method comprises the steps of firstly, obtaining initial three-dimensional pavement data of a target pavement, wherein the target pavement is the pavement needing crack detection, utilizing a line scanning three-dimensional measuring sensor to collect a series of section outlines of the target pavement along the measuring direction, and splicing the series of section outlines of the pavement to obtain the three-dimensional pavement data.
Due to the influence of noise, abnormal values generated by system abnormality and environmental abnormality need to be corrected, the cross section profile is converted from an image space to an object space through a calibration file, and cross section data of the object space is acquired so as to correct system errors in the measurement system, which is specifically as follows:
processing part abnormal zero-value noise points of the road surface section profile measured by the three-dimensional measuring sensor, which are caused by measuring environment interference, and acquiring an image space section profile; the calibration file is utilized to effectively correct system errors caused by sensor installation, laser line radian and uneven light intensity in the road surface section profile measured by the three-dimensional measuring sensor, the image direction and object space conversion is carried out, the real object space section profile information of the target road surface is obtained, and good data input is provided for subsequent crack detection and attribute information extraction.
In addition, in the process of collecting the road surface by the vehicle-mounted three-dimensional system, due to the influences of bumping and fluctuation of a travelling crane, deformation and damage existing on the road surface and the like, obvious low-frequency amplitude attitude fluctuation information exists in the collected original three-dimensional data, and cracks are hidden in the macro change information. In order to reduce the influence of data attitude fluctuation on crack extraction and the extraction of subsequent crack depth information, it is necessary to remove the attitude fluctuation contained in the data by using a correlation algorithm, and the method for removing the attitude adopts a three-dimensional road surface data component analysis method provided in the prior art: and removing low-frequency components in the three-dimensional cross section data, and only keeping the sum of the sparse components and the vibration components as the input of subsequent processing to obtain the three-dimensional pavement data of which the posture and deformation information is removed from the target pavement.
The processing reduces the influence of the driving fluctuation and the road surface deformation on crack detection on the premise of not losing crack information, so that crack display and elevation information extraction are facilitated.
S2, calculating the accumulated difference characteristics under each window based on the preset window number and the point elevation for each point on the section in the three-dimensional road surface data of the target road surface by taking the current point as a starting point to obtain the point-by-point multi-window accumulated difference characteristics corresponding to the target road surface;
in the cross section of the line scanning three-dimensional pavement data, cracks are interfered by various factors, clear edges and V-shaped structures are not always presented, and more, due to the interference of the background of the pavement texture, a plurality of V-shaped structures in the data are not real crack positions. Due to the influence of factors such as driving posture, deformation diseases, pavement material abrasion, crack types and pavement background difference, cracks in cross section data of line scanning three-dimensional pavement data are easy to present asymmetric V-shaped or non-V-shaped structures with unclear edges.
Therefore, for the crack extraction of line scanning three-dimensional data, the traditional template matching and edge detection are difficult to obtain a relatively ideal crack detection effect. On the basis of the three-dimensional road surface data after attitude and deformation information is removed, the embodiment of the invention adopts the section point-by-point multi-window accumulated difference characteristic to carry out modeling representation on the cross section cracks and the fluctuation characteristics of the texture of the three-dimensional road surface data, the processed basic unit is the cross section of the line scanning three-dimensional data, which is more in line with the data acquisition principle, and fully considers the V-shaped and non-V-shaped structures presented in the cross section data of the crack in the actual situation, utilizes the characteristic that the fluctuation characteristic of the road surface texture is stable in a certain range, and the section part with the crack has the elevation fluctuation trend to carry out modeling, the method can be used for clustering the acquired template among homologous data or sharing samples among different homologous data while overcoming the difference between noise and background, and can reduce threshold setting or dependence on homologous labeling samples in the crack extraction process.
S3, carrying out unsupervised clustering on the point-by-point multi-window accumulated difference characteristics corresponding to the target road surface according to a Kmeans clustering algorithm to obtain all crack objects in the target road surface;
or, performing supervision classification based on different source samples on the point-by-point multi-window accumulated difference characteristics corresponding to the target road surface according to the trained supervision classifier model, and acquiring all crack objects in the target road surface.
Specifically, after the point-by-point multi-window accumulated difference feature corresponding to the target road surface is extracted, according to the self-clustering property of the point-by-point multi-window accumulated difference feature, the point-by-point multi-window accumulated difference feature corresponding to the target road surface can be subjected to unsupervised clustering by using a Kmeans clustering algorithm, and all crack objects in the target road surface are extracted. And classifying the point-by-point multi-window accumulated difference characteristics corresponding to the target road surface by using the trained supervision classifier model, and extracting all crack objects in the target road surface.
According to the crack extraction method of the multi-window accumulated difference, provided by the embodiment of the invention, the characteristic that the fluctuation characteristic of the road surface texture is stable in a certain range and the elevation fluctuation trend of the section part with the crack is used for modeling is utilized, so that the noise and background difference is overcome, the obtained template can be clustered among homologous data or the samples are shared among different homologous data, and the threshold setting or the dependence on homologous labeled samples in the crack extraction process is reduced. On the basis of the point-by-point multi-window accumulated difference characteristic, a technical route for extracting the line scanning three-dimensional pavement data cracks is realized. Specifically, the method comprises the step of extracting the three-dimensional unsupervised pavement crack information by using the aggregative property of the proposed features and a typical unsupervised machine learning method, such as Kmeans, under the unsupervised condition (no sample label at all). Under the condition that a small amount of crack labels exist, label information and point-by-point multi-window accumulated difference features are trained to form a feature template, different source samples can be shared, the requirement of a supervision machine learning method on sample labeling is reduced, and therefore a three-dimensional pavement crack extraction result is obtained more accurately and rapidly. In addition, the provided point-by-point multi-window accumulated difference feature-based technical route has high adaptability to the existing unsupervised and supervised machine learning methods, the efficient machine learning method is introduced, the characteristics of sample sharing capability are provided, large-scale three-dimensional pavement crack extraction can be realized under a small amount of labeled samples, and a stable and robust method is provided for actual pavement crack detection.
On the basis of the above embodiment, preferably, the obtaining of the three-dimensional road data of the target road after the attitude and deformation information removal further includes:
based on a plurality of three-dimensional pavement data samples with the posture and deformation information removed, modeling and representing the fluctuation characteristics of cracks and textures of the cross section of the three-dimensional pavement data by adopting the point-by-point multi-window accumulated difference characteristics of the cross section, and acquiring the point-by-point multi-window accumulated difference characteristics corresponding to each sample;
after the point-by-point multi-window accumulated difference characteristics of the section are obtained, various unsupervised and supervised machine learning can be carried out on the basis of the characteristics so as to achieve the purpose of extracting crack information in data, and in the actual selection, unsupervised learning or supervised learning can be selected according to whether the collected sample has label information.
The three-dimensional road surface data samples after the attitude and deformation information is removed are obtained, the samples are used as training samples, the point-by-point multi-window accumulated difference characteristics corresponding to each sample are obtained, the obtaining process is the same as the point-by-point multi-window accumulated difference characteristics corresponding to the target road surface, and details are not repeated here.
If the labeling information of all samples is unknown, performing unsupervised clustering through a Kmeans clustering algorithm by using point-by-point multi-window accumulated difference characteristics corresponding to all samples to obtain trained Kmeans clustering algorithm parameters;
specifically, under the condition of no sample labeling, the three-dimensional road surface unsupervised crack information extraction is realized by using the aggregation of point-by-point multi-window accumulated difference features and a typical unsupervised machine learning method such as Kmeans.
If the labeling information of a part of samples in all samples is known, training the supervised classifier model by using the samples with known labeling information to obtain the trained supervised classifier model.
Specifically, under the condition that the labeling information of a small number of samples is known, the labeling information and the point-by-point multi-window accumulated difference features are trained to form a feature template, so that different source samples can be shared, the requirement of a supervision machine learning method on sample labeling is reduced, and the three-dimensional pavement crack extraction result is obtained more accurately and rapidly.
In addition, the provided point-by-point multi-window accumulated difference feature-based technical route has higher adaptation degree to the existing unsupervised and supervised machine learning methods, the efficient machine learning method is introduced, the characteristics of sample sharing capability are provided, large-scale three-dimensional pavement crack extraction can be realized under a small amount of labeled samples, and a stable and robust method is provided for actual pavement crack detection.
On the basis of the foregoing embodiment, preferably, the unsupervised clustering is performed on the point-by-point multi-window accumulated difference features corresponding to the target road surface according to the obtained parameters of the Kmeans clustering algorithm, so as to obtain all crack objects in the target road surface, and specifically includes:
carrying out unsupervised clustering on the point-by-point multi-window accumulated difference characteristics corresponding to the target road surface according to a Kmeans clustering algorithm, respectively obtaining the average depth of each clustering category according to the obtained clustering categories and the point sets with consistent labels of the clustering labels, and taking the clustering category with the minimum average depth as a suspected crack category;
forming a suspected crack object according to the connected domain by using the point set binary image corresponding to the obtained suspected crack type;
and finally selecting all crack objects in the target pavement from the suspected crack objects.
Specifically, unsupervised clustering is carried out on the point-by-point multi-window accumulated difference features of the road data by combining Kmeans clustering on the basis of the point-by-point multi-window accumulated difference features obtained by the three-dimensional road data after the attitude and deformation information is removed.
And respectively acquiring the average depth of each cluster type according to the acquired cluster type and the point set with consistent clustering label pairs (namely, the cluster labels correspond to the cluster types), and taking the type with the minimum average depth as the suspected crack type.
After the crack post-treatment, selecting a relatively determined crack object from the suspected crack objects as an extraction result, wherein the steps are as follows:
the category with the minimum average depth is the suspected crack category Ks, the other categories are set as 0, and the point corresponding to the Ks category is marked as 1. And forming an image object according to the connected domain by using the obtained point suspected crack point set binary image as a crack post-processing basic unit. And respectively obtaining an area average value A of the suspected crack object, a minimum external rectangular aspect ratio average value R and an area average value SA of the adjacent object. And (4) deleting the object with the object area smaller than A and the aspect ratio of the minimum external moment smaller than R by the crack post-treatment, and deleting the object with the object area smaller than A and the area of the adjacent object smaller than SA. According to the crack post-processing method, a relatively determined crack object is selected from the suspected crack objects as an extraction result.
The method is an unsupervised implementation example of a three-dimensional pavement crack information extraction technical route based on point-by-point multi-window accumulated difference characteristics.
On the premise that no marked information is available, accurate and rapid crack information extraction is completed by utilizing the point-by-point multi-window accumulated difference characteristics and the characteristic that the three-dimensional pavement crack depth is low, no complex threshold setting is needed, and simple, effective and universal line scanning three-dimensional pavement crack extraction is realized.
On the basis of the foregoing embodiment, preferably, the training the supervised classifier model by using a sample of known labeling information to obtain the trained supervised classifier model specifically includes:
randomly selecting crack sample points and non-crack sample points from marking data sets of different source three-dimensional pavements as training samples, wherein the training samples are samples of known marking information;
carrying out supervision classifier model training according to the supervision classifier model and the point-by-point multi-window accumulated difference characteristics of each training sample to obtain a trained supervision classifier model;
wherein the supervised classifier model comprises a Support Vector Machine (SVM), a K-nearest neighbor classifier (KNN) or a random forest RF classifier.
In the embodiment of the invention, a supervised classifier training model (such as a typical random forest RF classifier, a Support Vector Machine (SVM) or a K-nearest neighbor classifier and the like) is combined with a small amount of labeled samples of the three-dimensional pavement of the source A to train the supervised classifier training model for the point-by-point multi-window accumulated difference characteristics acquired by the three-dimensional pavement data samples after the attitude deformation information is removed.
Wherein both the fractured sample points and the non-fractured sample points are randomly selected from the annotated dataset. The crack labels required by training are not limited to crack types, pavement backgrounds and acquisition systems, and three-dimensional pavement data containing the crack labels with the transverse sampling interval of 1mm can be used as a training set. The cross section point-by-point multi-window accumulated difference features and the label set can provide classification models for a large amount of three-dimensional road surface data of different sources by the obtained supervision classifier training model.
And (3) using the random forest training model M obtained by the data training of the source A as a means of carrying out supervision and classification on different source samples on the point-by-point multi-window accumulated difference characteristics of the three-dimensional pavement data section of the source B. For test data B, marking and presetting conditions of the data are not needed, only the point-by-point multi-window accumulated difference features of the section are extracted, then a trained model M is marked by data A, and the crack extraction of the data B can be completed through a random forest classifier.
The method is an implementation example of a supervision classification mode of a three-dimensional pavement crack information extraction technical route based on point-by-point multi-window accumulated difference characteristics, and can also be applied to other typical supervision classifiers such as a support vector machine and the like.
Based on section point-by-point multi-window accumulated difference characteristics, the data B classification of different sources is realized by utilizing a model trained by data A, the sample sharing of a three-dimensional pavement crack marking data set is realized, and the limitation that the traditional supervision classifier can only realize the supervision classification of the same data on the same data sample can be overcome. The two specific technical route implementations both show the effectiveness of point-by-point multi-window accumulated difference characteristics, and for data which are not marked, clustering is carried out on the data according to characteristic aggregations to obtain crack information; and labeling information of different source data can be utilized to realize different source supervision and classification results and realize accurate and rapid line scanning three-dimensional pavement crack extraction.
On the basis of the foregoing embodiment, preferably, the calculating, based on the preset number of windows and the point elevation, the accumulated difference feature under each window to obtain the point-by-point multi-window accumulated difference feature corresponding to the target road surface specifically includes:
for each acquisition point of each cross section in the three-dimensional road surface data of the target road surface, taking the current acquisition point as a starting point, and acquiring point-by-point multi-window accumulated difference characteristics for each cross section based on the elevation of the acquisition point of the cross section and the number of preset windows;
specifically, the point-by-point multi-window accumulated difference characteristic is calculated by the following expression:
DN=[d1,d2,…,di,…,dN],i∈[1,2,3,…,N],
wherein D isNRepresenting a point-by-point multi-window accumulated difference signature, i representing a window range, diRepresenting window range threshold, e representing elevation of acquisition points of a cross section, epAnd (4) representing the elevation corresponding to the p-th point, wherein N represents the preset window number.
Fig. 2 is a schematic diagram of a line-scanning three-dimensional pavement crack extraction route based on point-by-point multi-window accumulated differential features according to an embodiment of the present invention, fig. 3 is a schematic diagram of two typical three-dimensional pavement data including cracks and corresponding crack labeling schematic diagrams provided in an embodiment of the present invention, fig. 3(a) shows a schematic diagram of a three-dimensional pavement including cracks, fig. 3(b) shows a corresponding crack labeling schematic diagram in fig. 3(a), fig. 3(c) shows another schematic diagram of a three-dimensional pavement including cracks, fig. 3(d) shows a corresponding crack labeling schematic diagram in fig. 3(c), it can be seen that posture deformation information in original data has an influence on crack extraction, fig. 4 is a schematic diagram of two cross-sectional cracks and partial display thereof in an embodiment of the present invention, and fig. 4(a) is a schematic diagram of one of the cross-sectional cracks and partial, fig. 4(b) is another schematic diagram of a cross-sectional crack and a partial display thereof, from which it can be seen that an actual road surface crack does not always present a simple V-shaped structure, so that a typical method for extracting cross-sectional crack information by using a template matching or edge method is likely to cause crack omission or an extraction result with a large background noise influence. The characteristic of the crack in the cross section in the three-dimensional data is scanned through an observation line, fig. 5 is a schematic diagram of calculation of point-by-point multi-window accumulated difference characteristics of the cross section and a basic abstract model in the embodiment of the invention, the crack in the cross section is abstracted into the model shown in fig. 5 in the embodiment of the invention, and the crack/non-crack point is modeled by using the local waveform characteristics of the crack and the texture on the elevation of the cross section.
Specifically, on the basis of the three-dimensional road surface data from which the attitude has been removed, the point-by-point multi-window integrated difference features are calculated for each point of the cross section in the manner described above, taking the current point as the starting point and taking windows 1, 2, 3, …, i, …, and N, respectively.
N represents a window range threshold, which may be set according to actual needs, and the value in the embodiment of the present invention is 200.
The following is a preferred embodiment of the present invention, and the method specifically comprises the following steps:
(1) data source
The embodiment of the technical scheme of the invention takes three-dimensional data of an asphalt pavement containing cracks as an example, and describes a method for extracting line scanning three-dimensional pavement cracks and attributes thereof facing an object.
(2) Data pre-processing
Due to the interference of a measuring environment (water stain and oil stain on a road surface or foreign matters in a measured area), part of abnormal noise (zero value points) may exist in the acquired data, and the abnormal noise points are replaced by non-abnormal sampling points close to the central area of the section; and correcting system errors caused by sensor installation, laser line radian and uneven light intensity distribution in the object section profile measured by the three-dimensional measuring sensor by using the calibration file, and converting image data into object data. And simultaneously splicing a series of pretreated sections along the driving direction to obtain the three-dimensional data of the asphalt pavement.
(3) Three-dimensional road surface data attitude fluctuation information removal
The method is characterized in that the low-frequency components in the three-dimensional pavement data are removed by analyzing the components of the three-dimensional pavement data provided in the prior art, and only the sum of sparse components and vibration components is reserved as the input of subsequent processing. The processing can reduce the influence of three-dimensional data attitude fluctuation acquired by the vehicle-mounted three-dimensional system on crack extraction, and is favorable for extracting subsequent crack depth information. For example, two typical three-dimensional road surface data containing cracks shown in fig. 3, and as can be seen from the labeled information, the posture deformation information in the data is disadvantageous for the visualization and extraction of the cracks.
Fig. 6 is a cross section point-by-point multi-window accumulated difference feature calculation example in the embodiment of the present invention, fig. 6(a) shows a cross section of a line-scanned three-dimensional road surface including cracks, fig. 6(b) shows a cross section subjected to attitude deformation removal, fig. 6(c) shows accumulated difference features of points of the cross section when the windows are 5, 15, 30, 50, and 80, and the cross section level effect is as shown in fig. 6(a) (b). Fig. 8(b) and 9(b) are three-dimensional road surface data obtained by cross-sectional splicing one by one and three-dimensional road surface high-frequency data obtained by cross-sectional splicing with attitude and undulation information removed.
(4) Section point-by-point multi-window accumulated difference feature acquisition and characterization
On the basis of the three-dimensional data with the attitude and the fluctuation information removed, for each point of the section, taking the current point as a starting point, respectively taking windows as 1, 2, 3, …, i, … and N (N is adjustable and is set as 200 by default) and calculating a point-by-point multi-window accumulated difference characteristic D according to the formulaN. FIG. 6(c) shows the cumulative difference d of each point of the cross section of FIG. 6(b) after the removal of the attitude at the windows 5, 15, 30, 50, 805、d15、d30、d50、d80. And when N is 200, the multi-window accumulated difference characteristic D of the section of fig. 6(b)NAs shown in fig. 7 (a). In particular, the figuresFIG. 7 shows an exemplary multi-window accumulated difference feature of a cross section, a crack point and a texture point, and FIG. 7(a) shows a multi-window accumulated difference feature D of the cross section shown in FIG. 6(a)N(N is 200), and fig. 7(b), 7(c), and 7(D) are D of crack points and texture points in the cross sectionNAnd (5) visualizing the characteristic mode.
Assuming that the section length is L, 200D is present at all points L-N in the front of the section shown in FIG. 6(b)NAnd (5) characterizing. Obtaining the multi-window accumulated difference characteristic D of each point for each section in the three-dimensional data according to the modeN。
D for selecting typical crack points and texture points of the section of FIG. 7(b)NThe features are visualized as shown in fig. 7(b), 7(c) and 7 (e). It can be seen that the window accumulated difference characteristics of the actual data obtained according to the method satisfy the theoretical model shown in fig. 5.
(5) Unsupervised three-dimensional pavement crack extraction based on point-by-point multi-window accumulated difference features and kmeans clustering
The embodiment of the invention combines Kmeans clustering to road surface data D for point-by-point multi-window accumulated difference characteristics acquired by three-dimensional road surface data after attitude deformation removalNThe method is characterized by carrying out unsupervised clustering, wherein the preset clustering number is K (K is more than 2, and the method defaults to 4). And respectively acquiring the average depth of each type of the acquired cluster type according to the point set with consistent clustering labels to the labels, taking the type with the minimum average depth corresponding to the K type labels as a suspected crack type Ks, setting other types as 0, and marking the point corresponding to the Ks type as 1. And forming an image object according to the connected domain by using the obtained point suspected crack point set binary image as a crack post-processing basic unit. And respectively obtaining an area average value A of the suspected crack object, a minimum external rectangular aspect ratio average value R and an area average value SA of the adjacent object. And (4) deleting the object with the object area smaller than A and the aspect ratio of the minimum external moment smaller than R by the crack post-treatment, and deleting the object with the object area smaller than A and the area of the adjacent object smaller than SA. According to the crack post-processing method, a relatively determined crack object is selected from the suspected crack objects as an extraction result.
Fig. 8 is a schematic diagram of an unsupervised three-dimensional pavement crack extraction example 1 based on point-by-point multi-window accumulated difference features and kmeans clustering in the embodiment of the present invention, fig. 8(a) is a depth-to-grayscale map of line-scanned three-dimensional pavement data, fig. 8(b) is three-dimensional pavement data after attitude and deformation information removal, fig. 8(c) is a suspected crack obtained by feature clustering, fig. 8(d) is a result of crack post-processing, fig. 9 is a schematic diagram of an unsupervised three-dimensional pavement crack extraction example 2 based on point-by-point multi-window accumulated difference features and kmeans clustering in the embodiment of the present invention, fig. 9(a) is a depth-to-grayscale map of line-scanned three-dimensional pavement data, fig. 9(b) is three-dimensional pavement data after attitude deformation removal, fig. 9(c) is a suspected crack obtained by feature clustering, fig. 9(d) is a result, fig. 8 and 9 illustrate the above-described procedure.
(6) Random forest classifier training based on point-by-point multi-window accumulated difference features
FIG. 10 is a schematic diagram of a crack/texture point feature template obtained by training a random forest classifier based on point-by-point multi-window accumulated difference features according to an embodiment of the present invention, as shown in FIG. 10, the embodiment of the present invention obtains point-by-point multi-window accumulated difference features D for three-dimensional road surface data after posture deformation removalNAnd on the basis, a random forest RF classifier and a small amount of labeled samples from the three-dimensional pavement A are combined to train the random forest classifier, wherein crack sample points and non-crack sample points are randomly selected from a labeled data set.
The sample size can be preset as required, in the embodiment, 20% of the crack marking data are randomly selected for training in the samples, but the proportion of the crack sample points and the non-crack sample points is required to be smaller than 1:1 and larger than 1:10 so as to meet the actual prior condition that the cracks in the pavement occupy a smaller amount and prevent the trained classifier from extracting too many or too few crack points. The crack marking required by training is not limited to crack types, pavement backgrounds and acquisition systems, and three-dimensional pavement data containing marked cracks of 1mm in transverse sampling interval can be used as a training set. The feature DNAnd a label set, wherein the obtained random forest classifier training model can be used for three-dimensional roads of a large number of different sourcesThe facet data provides a classification model.
(7) Different-source three-dimensional pavement crack extraction based on point-by-point multi-window accumulated difference characteristics and random forest
Using the random forest training model M obtained by training the data A in the step (6) as the three-dimensional road surface data D from the source BNAnd (4) carrying out supervision and classification on different source samples by using the characteristics. That is, for the test data B, only the feature D is extracted without any labeling and preset conditions of the dataNAnd then, marking the trained model M by using the data A, and finishing the crack extraction of the data B. Due to the difference of system and background characteristics, the classification result of the data B can be subjected to crack post-processing according to the step (5), and the crack extraction result of the three-dimensional data of the data source B is obtained on the basis of a small amount of labeling of the data source A. Based on the features DNThe method has the advantages that the data A training model is utilized to realize the classification of the data B of different sources, the sample sharing of the three-dimensional pavement crack marking data set is achieved, and the limitation that the traditional supervision classifier can only realize the supervision classification of the same data sample is overcome.
Fig. 11 is a schematic diagram of an example 1 of three-dimensional pavement crack extraction based on point-by-point multi-window accumulated difference features and different sources of a random forest according to the embodiment of the present invention, where fig. 11(a) and 11(B) are a three-dimensional graph of data a and data after pose deformation removal, fig. 11(c) is a three-dimensional graph of data B1 without labels, and fig. 11(d) is a crack extraction result of data B1 obtained by labeling a small amount of data a; fig. 11(e) is a three-dimensional diagram of data B2 not labeled, and fig. 11(f) is a crack extraction result of data B2 obtained with a small number of labels of data a.
Fig. 12 is a schematic diagram of an example 2 of three-dimensional pavement crack extraction based on point-by-point multi-window accumulated difference features and different sources of a random forest according to the embodiment of the present invention, where fig. 12(a) and 12(B) are a three-dimensional graph of data a and data after pose deformation removal, fig. 12(c) is a three-dimensional graph of data B1 without labels, and fig. 12(d) is a crack extraction result of data B1 obtained by using a small amount of labels of data a; fig. 12(e) is a three-dimensional view of data B2 without notation, and fig. 12(f) is a crack extraction result of data B2 obtained with a small number of notations of data a.
The results of fig. 11 and 12 illustrate the above process, and from the point of view of two sets of different source random forest classification results, the classification results using different sources are superior to the unsupervised clustering results in step (5). Two groups of experiments show the effectiveness of point-by-point multi-window accumulated difference characteristics, and for data which are not marked, clustering can be carried out on the data according to characteristic aggregations according to the step (5) to obtain crack information; and (4) according to the step (7), the labeling information of different source data can be utilized to realize different source supervision and classification results, and higher crack extraction precision is realized.
The technical problems solved by the invention are as follows:
(1) source and destination of data. The invention designs a point-by-point multi-window accumulated difference characteristic which can be suitable for various pavement backgrounds and various crack types by utilizing the pavement data acquired by a line scanning three-dimensional measuring sensor, and performs unsupervised/supervised extraction of different source three-dimensional pavement data cracks on the basis.
(2) And acquiring and representing the point-by-point multi-window accumulated difference characteristics of the section. In the cross section of the line scanning three-dimensional pavement data, cracks are interfered by various factors, edge clearness and V-shaped structures are not always presented, and more, due to the interference of a pavement texture background, a plurality of V-shaped structures in the data are not real crack positions. Due to the influence of factors such as driving posture, deformation diseases, pavement material abrasion, crack types and pavement background difference, cracks in the cross section data of the line scanning three-dimensional pavement data are easy to present asymmetric V-shaped or non-V-shaped structures with unclear edges.
Therefore, for the crack extraction of line scanning three-dimensional data, the traditional template matching and edge detection are difficult to obtain a relatively ideal crack detection effect. On the basis of the three-dimensional road surface data after attitude and deformation information is removed, the invention adopts the section point-by-point multi-window accumulated difference characteristic to carry out modeling representation on the fluctuation characteristics of the cracks and the textures of the cross section of the three-dimensional road surface data.
(3) Extracting unsupervised three-dimensional pavement cracks based on point-by-point multi-window accumulated difference characteristics and kmeans clustering. The traditional unsupervised machine learning method is difficult to directly obtain a good effect in online scanning three-dimensional pavement data crack extraction, the online scanning three-dimensional data has influence of factors such as driving posture deformation diseases and the like, and the traditional template matching and edge detection method has limited applicability to different pavement cracks with different data. On the basis of the line scanning three-dimensional data without posture deformation, the method firstly utilizes the point-by-point multi-window accumulated difference characteristic to overcome the difference of different crack types of different road surface backgrounds, can directly utilize the widely used Kmeans clustering method to obtain crack information from the data, and can accurately and quickly obtain the line scanning three-dimensional road surface crack information without marking the information.
(4) And training a random forest classifier based on point-by-point multi-window accumulated difference characteristics. The traditional supervised learning method or the deep learning method has extremely high requirements on the homogeneity of training samples and test data, can hardly achieve the purpose of sample sharing between different data, and can obtain a better supervised classification effect only by using labeled samples from the same data.
Therefore, the function of the machine learning model in the practical application of online scanning three-dimensional pavement crack detection is greatly limited. The method comprises the steps of combining a random forest RF classifier and a small amount of labeled samples from the three-dimensional road surface A to train the random forest classifier for point-by-point multi-window accumulated difference characteristics acquired by three-dimensional road surface data after posture deformation removal, wherein crack sample points and non-crack sample points are randomly selected from a labeled data set.
The crack marking required by training is not limited to crack types, pavement backgrounds and acquisition systems, and three-dimensional pavement data containing marked cracks of 1mm in transverse sampling interval can be used as a training set. The obtained random forest classifier training model can provide classification models for a large number of different sources of three-dimensional pavement data.
(5) And extracting the three-dimensional pavement cracks based on different sources of the point-by-point multi-window accumulated difference characteristics and the random forest. The random forest training model M obtained by training the data A is used for carrying out supervision and classification on different source samples on the three-dimensional pavement data DN characteristics of the source B.
For the test data B, marking and presetting conditions of the data are not needed, and after the features are extracted, the trained model M is marked by the data A, so that the crack extraction of the data B can be completed.
Based on the proposed characteristics, the data A training model is utilized to realize the classification of the data B of different sources, the sample sharing of the three-dimensional pavement crack marking data set is achieved, and the limitation that the traditional supervision classifier can only realize the supervision classification of the same data for the same data sample can be overcome.
Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and certainly can also be implemented by hardware. With this understanding in mind, the above-described technical solutions may be embodied in the form of a software product, which can be stored in a computer-readable storage medium such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the methods described in the embodiments or some parts of the embodiments.
Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.