Keywords

1 Introduction

Crack detection technology is becoming more and more mature, providing a lot of technical support for pavement maintenance. Scholars from all over the world have gradually developed a variety of pavement detection system based on intelligent detection technology [1], designed the double connectivity detection of pavement crack detection algorithm [2], image measurement method of asphalt pavement damage [3]. Asphalt pavement crack detection based on mathematical morphology [4] based on Prim pavement crack connection algorithm of minimum spanning tree [5]. The processing algorithm in HARRIS pavement damage detection system will be affected by environmental factors. Therefore, the elements of a sidewalk automatic measurement and 3D methods, based on 3D data of pavement crack detection algorithm [6, 7] is proposed.

The crack images are easily influenced by texture features and there's a lot of noise [8], so the next step is to filter the image, this method is often combined with neural network classifier, for distress image segmentation [9]. For early pavement management and repair, the circular Radon transform RGB camera image is used to propose a set of better road infrastructure evaluation indicators [10]. When using cameras and deep learning networks for crack detection, there is a problem of insufficient filtering of the road background. In order to improve the detection efficiency as a whole, a lightweight Tiny-Darknet is combined with YOLOv3, and open VINO is used for model optimization and reasoning to accelerate the detection of road cracks. But this method slows down the number of frames detected on the video image [11].

Traditional detection methods are time-consuming, while pavement detection vehicles are expensive [12]. In this paper, the improved UK scanner method divides video frame image into grid block, which can not only calculate the number of cracks, but also calculate the critical rate of cracks, and increase the accuracy of crack detection and statistics.

2 Crack Image Acquisition and Feature Extraction

Perform image extraction and a series of image processing on the road video, and then correct the image and remove duplicate parts. The process of crack image acquisition and feature extraction is shown in the Fig. 1.

Fig. 1.
figure 1

The process of crack image acquisition and feature extraction.

Full size image

2.1 Image Acquisition

A video data acquisition system is designed to collect images of pavement cracks, including a lightweight waterproof high-resolution GARMIN motion camera and a vehicle. Set a specific angle and height to the camera fixed on the vehicle to ensure complete collection of road images. Use GPS locator to record the longitude, latitude, altitude and time of the vehicle during the video capture process. The camera acquires 20 to 40 frames per second. According to the principle of dynamic photography, a four-point calibration method is used to calibrate the detection area of the road video frame image.

2.2 Feature Extraction

This paper uses the region growth algorithm to further process the image. First, enhancing non-scale filtering and adaptive threshold segmentation, the contour of the crack is obtained. However, this contour is not a complete crack but a part of the real crack. Second, calculate the complete crack size according to the extracting cracks image features. The crack growth algorithm based on regional growth is shown in Fig. 2.

Fig. 2.
figure 2

Crack growth algorithm implementation flow based on region growth.

Full size image

The traditional region growth algorithm can only grow adjacent pixels. If a crack is not a seed point and is not adjacent to other crack seeds, these cracks cannot be detected. The proposed crack growth algorithm optimizes the traditional region growth algorithm, which the search scope is extended to the n*n region. The condition is that the absolute difference between pixel values is less than or equal to the threshold K. Experiments are carried out based on the optimization region of n*n region convolution. Let n be 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31. The threshold K is set from 1 to 50. After comparative analysis, when n = 15, K = 15, the effect is best.

2.3 Frame Correction and the Overlapping Part Removal

When acquiring video frame image of road surface, the detection area in the image is irregular quadrilateral due to the difference in Angle of view and acquisition distance. It leads to the errors of different pixel points corresponding to the actual coordinate system values, which directly affects the calculation accuracy of geometric parameters of crack distress. Therefore, the image detection area shown in Fig. 3(a) needs to be corrected by reverse-engineering method of perspective transformation to restore to the real rectangular road surface image shown in Fig. 3(b). Then, by using the image crack feature extraction algorithm given in Sect. 2.2, the obtained binary image is shown in Fig. 3(c).

Fig. 3.
figure 3

Crack feature extraction.

Full size image

There is inevitable overlap between the two frames in the detection area. By analyzing the longitude, latitude, altitude and time information recorded every second, the method of removing overlap in the detection area is shown in Fig. 4.

Fig. 4.
figure 4

Remove overlap schematic diagram.

Full size image

Figure 4 shows only the first frame image can be fully used, and the actual length of pavement cracks in other frame images can be calculated according to Eq. (1). \(L_1\) is the total length of road surface, \(l\) is the corresponding length of road surface in the first video frame image, and \(N\) is the number of frames.

$$ len = \frac{L_1 - l}{{N - 1}} $$
(1)

3 Crack Image Distress Parameters Calculation and Assessment

The geometrical parameters of crack include crack area, crack length and crack width. As long as the grids are small enough, the grids can be approximately regarded as external rectangles of cracks. For each sub-block crack image, the image is scanned. If there is crack pixel, the sub-block is marked as a crack region sub-block. Otherwise, the sub-block will be not marked.

Set the initial pixel area of the crack \(A_x = 0\) and scan each pixel in the sub-block from left to right from top to bottom. If the pixel is a crack, increase the pixel area, \(A_x = A_x \,{ + }\,1\). If the pixel is the background, continue to scan the next pixel point until all the pixels in the sub-block are scanned, and the pixel area of the sub-block can be obtained. According to the actual length and width of each pixel, the actual area of each pixel can be calculated. The crack area in this sub-block \(A\) can be obtained by multiplying the number of pixels \(A_x\).

Initially set the length and width of crack projection \(T_c = 0,T_k = 0\). Scan each column pixel in the sub-block from left to right. If a column pixel contains crack pixels, the projection width will increase, that is \(T_k = T_k + 1\), otherwise no operation will be carried out and continue to scan the next column pixel until all the columns are scanned, then the projection width of crack \(T_k\) can be obtained. Similarly, scan each row pixel in the sub-block from top to bottom to calculate the projection length Tc.

The camera correction information for each pixel is calculated based on the actual length and width. Projection length Tc and projection width Tk become an actual length Lc and an actual width Lk. Since the sub-blocks of the whole pavement are small, the crack length can be approximated according to the Pythagorean theorem, as shown in Eq. (2).

$$ L = \sqrt {L_c^2 + L_k^2 } $$
(2)

The crack width can be approximately divided by the actual crack area by the actual length.

The UK Scanner pavement crack damage assessment method is shown in Fig. 5. The left picture is the original pavement image, and the right picture is the crack map measured after grid division.

Fig. 5.
figure 5

UK Scanner pavement crack damage assessment schematic.

Full size image

The UK scanner method divides the pavement crack grade into three categories: light, medium and heavy. Different grades of cracks are different in the evaluation of pavement condition. The measurement method of DMI (distress measurement index) for relevant research and evaluation is shown in Eq. (3).

$$ DMI = \frac{1*N_1 + 2*N_2 + 3*N_3 }{{\left( {1 + 2 + 3} \right)*N}} $$
(3)

In Eq. (3), \(N_1\) is the number of light crack distresses, \(N_2\) is the number of medium crack distresses, \(N_3\) is the number of heavy crack distresses, N is the total number of cracks, and 1, 2, 3 are the specified coefficients of cracks of different grades.

The crack width standard is \(k_1\) mm, \(k_2\) mm, where \(k_1 \le k_2\), the crack width \(D \le k_1\) is the light crack, the crack width \(k_1 < D \le k_2\) is the medium crack, the crack width \(D > k_2\) is heavy cracks. Calculation of the pavement condition index (DMI): The different grades of cracks have different degrees of distress to the pavement. Set the light crack coefficient as, the medium crack coefficient as \(k_M\), and the heavy crack coefficient as \(k_L\). Calculate the test pavement index DMI as shown in Eq. (4), and calculate the pavement distress index. The crack data is processed, classified, calculated and stored in the database.

$$ DMI = \frac{k_S *N_S + k_M *N_M + k_L *N_L }{{N_S + N_M + N_L }} $$
(4)

4 Test Results Analysis

In order to test the accuracy of crack feature extraction algorithm, 20 image samples were extracted from different video samples collected, and two methods were used for processing. 1) The grid was divided with 15*6 (pixel) as the side length, and introduce ‘N1’ as the pseudo-ground truth. 2) The image is detected by crack feature extraction algorithm. The grid is divided into 15*6 (pixel) edges, and the number N2 of the grid containing cracks is counted in the background of the system. The relative error calculation formula obtained by comparing the detection results of the two methods is shown in Eq. (5). The statistical results are shown in Table 1.

$$ \begin{array}{*{20}c} {\delta = \frac{{\left| {N_1 - N_2 } \right|}}{N_1 } \times 100\% } \\ \end{array} $$
(5)
Table 1. Statistical table of cracks relative errors.
Full size table

The average relative error of the 20 samples is 31.57%, and the error between the manual detection value and the system extraction value is analyzed as follows: N1 > N2 means that the grid containing cracks has not been detected, and N1 < N2 means that many interferences have been mistakenly identified as cracks, such as water and shadows. As the upper half of the perspective image has less image information than the lower half, the incomplete and missing part of the corrected image leads to blurring, which affects the detection results.

The length of the segment taken from the collection video is 4′25″52 s, and the detection distance is 6213 m. All cracks were divided into three degrees according to the principle that the crack width less than 3 mm was considered as mild crack, the crack width between 3 and 6 mm was considered as moderate crack, and the crack width greater than 6 mm was considered as severe crack.

The number of crack grids, the total number of crack grids, the total number of grids and the percentage of crack grids in each interval are respectively counted. The test data are counted every 1000 m, and the number of crack grids with different damage degrees in each interval is analyzed and compared, as shown in Fig. 6. The blue part is the number of mild crack grids, the orange part is the number of moderate crack grids, and the gray part is the number of severe crack grids.

Fig. 6.
figure 6

Crack level grid number statistics.

Full size image

The actual damage degree of crack distress in test sections is shown in Fig. 7. It can be seen that the distress degree of the measured section is the most serious in the range of 3000–4000 m.

Fig. 7.
figure 7

Pavement distress assessment distribution.

Full size image

5 Conclusions

This paper proposes methods to solve the problem of crack feature extraction and parameter quantification in video. The main conclusions can be summarized in the following two aspects:

  1. 1)

    The image is segmented using automatic thresholds, and non-crack interferences are filtered according to the area and crack morphological characteristics. The traditional area growth algorithm is improved to solve the problem of the algorithm's inability to grow across areas and identify the characteristics of cracks.

  2. 2)

    A crack quantification method based on the combination of UK Scanner grid method and projection method is proposed. Count the crack information of all grid sub-blocks and use the improved UK Scanner pavement crack damage evaluation method to evaluate the detected road.