Pavement crack analysis, which deals with crack detection and crack growth detection, is a crucial task for modern Pavement Management Systems (PMS). This paper proposed a novel approach that uses historical crack data as reference for automatic pavement crack analysis. At first, a multi-scale localization method, which including GPS based coarse localization, image-level localization, and metric localization has been presented to establish image correspondences between historical and query crack images. Then historical crack pixels can be mapped onto the query crack image, and these mapped crack pixels are seen as high-quality seed points for crack analysis. Finally, crack analysis is accomplished by applying Region Growing Method (RGM) to further detect newly grown cracks. The proposed method has been tested with the actual pavement images collected in different time. The F-measure for crack growth is 88.9%, which demonstrates the proposed method has an ability to greatly simplify and enhances crack analysis result.
Contour 2.1.2 Crack
Modern Pavement Management Systems (PMS) are playing more and more important roles in pavement survey, maintenance, and rehabilitation. An increasing number of transportation agencies are building and upgrading their PMS to enhance pavement management. In PMS, pavement crack analysis, which deals with crack detection and crack growth detection, is a crucial and core task. With the development of sensor and information technology, crack data can be quickly and automatically collected by vehicle-borne sensors. For example, many transportation agencies are routinely (e.g., quarterly, semi-annually, annually, etc) using sensor vehicles for pavement image collection. Thereby, how to accurately identify pavement cracks from the collected pavement images is becoming a key technology in automatic crack analysis in PMS.
Currently, many transportation agencies are collecting pavement image data in a routinely and periodic manner. For example, most of transportation agencies collect pavement data annually, while some of them collect pavement data every 6 months. As a result, historical crack images can be obtained by several crack data collections. The pavement cracks will not grow substantially in a short period of time. Hence, the historical crack data can be utilized to enhance current crack analysis. To achieve this goal, an important task is to establish the correspondence between historical and current crack images, which is also called localization. Therefore, this paper proposed a method that using historical data as reference for current crack analysis. Especially, the proposed method aim to address the localization problem by using a multi-scale strategy. Contributions of this paper are summarized as follow: 1) The authors proposed a method that using historical crack data as the reference for pavement crack analysis, which can greatly enhance the performance of pavement crack analysis; 2) The authors proposed a multi-scale localization strategy to match historical crack image with current crack image. The multi-scale localization method consists of GPS-based coarse localization, image-level localization and finally pixel-level localization; 3) By referring to historical crack data, the authors proposed a novel approach, called RGM, it can detect the condition change of pavement cracks easily. The condition change of pavement crack is especially important for pavement treatment strategies.
The rest of this paper is organized as follow: Section 2.1 introduces the multi-scale localization and crack data mapping. Section 2.2 introduces crack detection and analysis by RGM. Section 3 presents the experimental results. Section 4 draws the conclusion.
We got the authority of work from Wuhan university of technology. As illustrated in Fig 1, the proposed reference-based crack analysis method consists of three main modules: 1) multi-scale localization; 2) mapping historical crack image onto the query crack image; 3) crack post-processing and analysis. Each query crack data and historical crack data contains GPS information and a crack image. Each crack image is well associated with GPS information. In addition, the historical crack label, either extracted in manual or automatic way, are represented as a limited number of pixels belonging to crack (in pixel-level) in the pavement images. Hence, each historical crack data can be presented by using point sets as follows:(1)where n is the number of historical crack data. Gi is GPS information. Ii is the pavement crack image and Li represents all the crack pixels that are labelled in the historical image.
The module of multi-scale localization aims to establish the pixel-level image correspondence between the query crack data and the historical crack data. The proposed method adopts a coarse-to-fine strategy to achieve multi-scale localization. And it consists of three steps, which are GPS-based coarse localization, image-level localization and pixel-level localization. Based on the localization results, the labelled crack pixels in the historical image can be mapped onto the query crack image for crack analysis.
The first step of multi-scale localization is the coarse localization using GPS data, as illustrated in Fig 2. Let Gj be the GPS coordinate of jth query crack data and Gi be the GPS coordinate of ith historical crack data. The distance between Gj and Gi can be calculated as follow:(2)
With the GPS data matching, a set of historical crack image candidates can be obtained, whose associated GPS coordinates and the query GPS coordinates are within a threshold distance. Thus, this step is called GPS-based coarse localization. The GPS-based coarse localization allows us to derive a limited number of candidates from a huge amount of historical crack images collected. The mathematical of this task are described as follow:(3)where k is a threshold distance to select the candidate historical crack data. It decides that the ith historical crack data is close enough to the jth query crack data. dij is the distance between ith historical crack data and jth query crack data. In practice, according to the accuracy of GPS localization, the threshold distance is heuristically set to 10 meter. After GPS-based coarse localization, a limited number of candidate historical crack images, which satisfies Eq (2), are obtained.
In this paper, ORB [19] is utilized to extract local feature points both for query and candidate historical crack images. The ORB is an image match algorithm which combines oFAST (FAST with orientation) and rBRIEF (rotated BRIEF). Compared to the classic SIFT and SURF, ORB has much faster computation speed. More specific, it applies oFAST for feature point detection and rBRIEF for feature descriptor computation. The oFAST develops from FAST (From Accelerated Segment Test). FAST takes one parameter, the intensity threshold between the centre pixel and those in a circular window around the centre. Hence, FAST is very fast in implementation. Practically, FAST-9, which circular radius is 9, is usually used for good performance. The performance of oFAST is enhanced in two ways. First, the Harris corner measure is adopted to select the distinct FAST points. Second, the orientation information is added such that the extracted feature points are orientation invariant. The orientation is calculated from the moments of the image within a circular window:(4)where x, y are coordinate of an image and I(x,y) is moment of this image. As a result, the centroid of the image is computed from the moments as follows:(5)
With the computed homography matrix, the historical crack label can be mapped onto the query image as follow:(17)where n is the number of labelled crack pixels on historical image. [μi νi]T is the coordinates of pixel which is the mapped crack label on the historical image. [xi yi]T is the coordinate of mapped crack label on the query image.
Once the authors obtain the crack pixels on the query image by mapping the historical crack pixels, which were well labelled in the historical database, all the crack pixels can be detected afterwards. These mapped crack pixels on the query images are important clues for crack detection. The grown crack pixels of query images can be detected by RGM through mapped crack pixels.
Note that, the proposed method allows that each query crack image has unique pixel Gaussian model. Compared to the hard threshold methods in literature, the proposed method is thus more robust and adaptive for region growing. From the computed Gaussian model, the authors can quickly determine if a neighbour image I(pμ,pν) has the similar properties with the seed points. As crack region in the image usually has low image intensities, the authors can thus set a range of image intensities from the Gaussian model parameters, such as the mean and the standard deviation. Hence, a point (pμ,pν) is classified into crack if its intensity satisfied the following conditions:(21)where λ is determinate ratio that is empirically set in the practical applications. The detail for selection of determinate radio can be seen is Section 3.2.
As a result, the RGM can be applied to Eq (21) to detect all the neighbour homogenous points near the seed points. By utilizing the RGM, the authors can detect all the pixels belonging newly grown cracks in the query crack image. These newly grown crack pixels are important for us to analyse the crack growth and predict the crack severity.
In order to test the proposed method, two scenarios were selected for experiments. In order to test the proposed method, two scenarios were selected for experiments. The first scenario was Youyi Road near Yujiatou campus of Wuhan university of technology (WUT) (GPS coordinate: 114.363, 30.615). The second one was Linjiang Road, Wuhan city (GPS coordinate: 114.354, 30.621). The total distance for two scenarios is 6.5KM that amount of pavement crack data has been collected. Both scenarios have high traffic volume every day. The types of pavements for two scenarios are asphalt. Various kinds of pavement crack were recorded, such as longitudinal crack, transverse crack and fatigue crack, as illustrated in Fig 5. 2ff7e9595c
Comentários