INVENTORS: Xiangxiong Kong and Jian Li

CROSS-REFERENCE

[0001] This application claims the benefit of U.S. Provisional Application No.

62/647,136, filed March 23, 2018, the entire contents of which is hereby incorporated herein by reference.

BACKGROUND

[0002] Civic infrastructure, such as buildings, roads, bridges, towers, etc. are susceptible to structural damage and possible failure due to the significant loads that they sustain over long periods of time. In particular, bolted steel joints are prone to structural damage over long service periods due to self-loosening of bolts, mainly caused by repetitive loads and vibrations.

[0003] Self-loosening of bolts caused by repetitive loads and vibrations is one of the common defects that could weaken the structural integrity of bolted steel joints in civil structures. Many existing approaches for bolt loosening detection are based on physical sensors, hence, they require extensive sensor deployments, which may limit their abilities for cost-effective detection of loosened bolts in a large number of steel joints. Additionally, the extra work is required for the installation of sensors and cables, leading to complex and expensive monitoring systems.

SUMMARY

[0004] According to one embodiment, a method includes capturing a first image of a fastener at a first time, capturing a second image of the fastener at a second time, performing a feature-based image registration to create a third image, performing an intensity-based image registration to create a fourth image, determining registration errors based at least in part on a comparison of the first image and the fourth image, and performing a feature enhancement process on the registration errors to determine whether the fastener has loosened between the first time and the second time.

[0005] According to another embodiment, a system for fastener loosening detection can include a capture device and at least one computing device. The at least one computing device can be configured to obtain a first image of a fastener at a first time based at least in part on the capture device, obtain a second image of the fastener at a second time based at least in part on the capture device, perform a feature-based image registration to create a third image, perform an intensity-based image registration to create a fourth image, determine registration errors between the first image and the fourth image, and perform a feature enhancement process on the registration errors to determine whether the fastener has loosened between the first time and the second time.

[0006] According to another embodiment, a non-transitory computer-readable medium can embody a program executable by one or more computing devices. The program can cause the one or more computing devices to capture a first image of a fastener at a first time, capture a second image of the fastener at a second time, perform a feature-based image registration to create a third image, perform an intensity-based image registration to create a fourth image, determine registration errors between the first image and the fourth image, and determine whether the fastener has loosened between the first time and the second time based at least in part on a feature enhancement process algorithm and the registration errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Many aspects of the present disclosure can be better understood with reference to the following drawings. The components in the drawings are not necessarily drawn to scale, with emphasis instead being placed upon clearly illustrating the principles of the disclosure.

In the drawings, like reference numerals designate corresponding parts throughout the several views.

[0008] FIG. 1 illustrates a proposed vision-based fastener loosening detection approach according to various embodiments described herein.

[0009] FIG. 2 illustrates an example of the feature-based image registration algorithm according to various embodiments described herein.

[0010] FIG. 3 illustrates certain aspects of the feature-based image registration algorithm according to various embodiments described herein.

[0011] FIG. 4 illustrates an example of the principle of intensity -based image

registration according to various embodiments described herein.

[0012] FIG. 5 illustrates an example of the principle of feature enhancement according to various embodiments described herein.

[0013] FIG. 6 illustrates an example of the principle of result implementation according to various embodiments described herein.

[0014] FIG. 7 illustrates an example experimental validation test setup according to aspects of the embodiments described herein.

[0015] FIG. 8 illustrates results of an example experimental validation test setup according to aspects of the embodiments described herein.

[0016] FIG. 9 illustrates results of an example experimental validation test setup according to aspects of the embodiments described herein. [0017] FIG. 10 illustrates results of an example experimental validation test setup according to aspects of the embodiments described herein.

[0018] FIGS. 11 A and 11B illustrate a comparison of the fastener loosening detection results using two different image resolutions using the proposed vision-based approach according to aspects of the embodiments described herein.

[0019] FIG. 12 illustrates a comparison of the fastener loosening detection results at various rotation angles using the proposed vision-based approach according to aspects of the embodiments described herein.

[0020] FIGS. 13A-E illustrate a comparison of the performance of the Shi-Tomasi features for the feature-based image registration with other common types of feature-based image registration processes according to aspects of the embodiments described herein.

[0021] FIG. 14 illustrates a comparison of the results of the feature enhancement at various threshold levels using the proposed vision-based approach according to aspects of the embodiments described herein.

[0022] FIG. 15 illustrates a comparison of the fastener loosening detection results at lighting conditions using the proposed vision-based approach according to aspects of the embodiments described herein.

[0023] FIG. 16 illustrates results of an example experimental validation test setup according to aspects of the embodiments described herein.

[0024] FIG. 17 illustrates a test specimen according to aspects of the embodiments described herein.

[0025] FIG. 18 illustrates results of an example experimental validation according to aspects of the embodiments described herein applied to the test specimen of FIG. 17. [0026] FIG. 19 illustrates results of an example experimental validation test setup according to aspects of the embodiments described herein applied to the test specimen of FIG. 17.

DETAILED DESCRIPTION

[0027] As noted above, bolted steel joints are prone to structural damage over long service periods due to self-loosening of bolts, mainly caused by repetitive loads and vibrations. Bolt loosening leads to a loss of clamping force acting on the joint, further causing stiffness degradation, and potentially structural failure. Therefore, monitoring health conditions of bolts in a timely fashion is essential to structural integrity as appropriate replacements or retrofits can be applied before the steel joints reach critical conditions.

[0028] Human visual inspection has been commonly applied to detect bolt loosening in civil structures. For example, the Federal Highway Administration (FHWA) requires routine inspections in two-year intervals for highway bridges in the United States. Trained bridge inspectors visually detect and record structural defects in bridges, including bolt loosening during inspections. However, human inspection is labor intensive and less effective as bolts may loosen between inspection intervals. More importantly, inspection results may contain errors due to inconsistencies in inspection skills and ability to interpret data among inspectors.

[0029] Advanced bolt loosening detection technologies have been developed using in the field of structural health monitoring (SHM) and nondestructive testing (NDT). Nevertheless, the successes of these methods may rely on extensive work of human operations and/or sensor deployments, which could be costly and less flexible for rapid inspections of bolted steel joints in civil structures. [0030] Computer vision-based technologies have received significant attentions in the SHM community due to the benefits of being low-cost, easy-to-deploy, and contactless. Several vision-based approaches have been reported for monitoring health conditions of civil structures at both global and local scales. In the field of vision-based fastener loosening detection, edge detection-based techniques have been used. However, these edge detection- based techniques compare the nut boundary (i.e. edges) before and after the nut rotation to determine loosening of the nut, which limits flexibility for automatically processing a large volume of images. Other approaches have combined vision-based technologies with machine learning algorithms to achieve a fastener loosening detection. However, prior knowledge about the damage state of the bolt (i.e. classifications of loosened bolt and tightened bolt) is required in order to train the machine learning algorithm, and the training procedure has to be repeated for new types of bolts with different dimensions or shapes. The above mentioned drawbacks limit the suitability of existing computer vision-based techniques.

[0031] In the context outlined above, a new computer vision-based bolt loosening detection method using image registration is described herein. The approach is based on mapping images at different inspection periods into the same coordinate system and uncovering differential features caused by the loosened bolt. This approach does not require extensive operations for finding the rotation of the nut’s boundaries and does not require prior knowledge about the monitored structure (such as bolt types) or damage state of the bolt. In these regards, the present approach is more flexible and cost-effective for engineering applications. The detection result of this approach is also presented for easy interpretation such that direct actionable decisions can be made to conduct condition-based maintenance procedures. The approach can include detecting loosened fasteners, such as bolts, nuts, etc., in captured images. Furthermore, when equipped with autonomous platforms, such as unmanned aerial vehicles (UAVs), vision-based SHM could bring higher flexibility and cost- effectiveness to structural inspection.

[0032] Turning to the drawings, FIG. 1 illustrates a proposed vision-based fastener loosening detection approach according to various embodiments described herein. As shown in FIG. 1, a bolted steel joint 103 (the monitored structure) comprises bolt 106 and is evaluated in two inspection periods, and bolt 106 is found to be loosened during the inspection interval. Two input images, denoted as Image 1 and Image 2, are collected by a camera during the two inspection periods at block 133. In various embodiments, the location of the camera relative to the bolted steel joint 103 may not be the same for the acquisition of the two images. Therefore, directly identifying the rotated bolt by overlapping the two input images is not possible. This can also be confirmed by the intensity comparison between Image 1 and Image 2, as shown in block 112, in which intensities of exactly matched pixels are illustrated as 0 (black), and intensities of unmatched pixels are in the region of 1 to 255 (grey to white), deepening on the level of their discrepancies.

[0033] In order to align two input images together, the approach applies a feature-based image registration method to transform Image 2 to Image 3 at block 136. The resulting Image 3 has the same coordinate system as Image 1. A region of interest 109 (ROI) may be assigned prior to this procedure as shown in Image 1. The purpose of defining the ROI 109 is to specify a region in Image 1 as the target region where Image 2 should match. In various embodiments, the ROI 109 covers a group of bolts and their adjacent structural surface and excludes unnecessary elements in the scene (e.g. the wall in the background in Image 1).

After the feature-based image registration, the matching performance is improved as shown in the intensity comparison between Image 1 and 3, as shown in block 112. Nevertheless, misalignments (i.e. registration errors) still exist, especially around the areas of Bolt 1 and Bolt 3, even if they are intact during the inspection interval. [0034] To further reduce registration errors, the approach then utilizes an intensity-based image registration method at block 139, which transforms Image 3 into Image 4 to non- rigidly match Image 1. As can be found in the intensity comparison between Image 1 and Image 4 (FIG. ld), registration errors are significantly reduced around Bolt 1 and Bolt 3, while these errors still exist around the loosened Bolt 106 due to the bolt rotation. These errors are treated as the bolt loosening features which are introduced by multiple sources during the bolt rotation, such as: hexagon boundaries of the bolt head, the mark of A325, and other surface textures on the bolt head surface. Next, the approach enhances the above bolt loosening features by filtering out adjacent noisy contents in the registration errors at block 142. Next, the enhanced bolt loosening features obtained at block 142 are applied to the original input image (Image 1) so that the loosened bolt 106 can be directly visualized at block 145, and informed actionable decisions can be made to perform appropriate rehabilitations and/or retrofits to the monitored steel joint.

[0035] The feature-based image registration effectively aligns two input images into the same coordinate system based on a predefined ROI. However, small misalignments are usually associated with feature-based image registration. The intensity-based image registration, on the other hand, is able to adjust small misalignments but may have difficulty handling significant misalignments if the input images are taken from very different camera poses. By adopting these two image registration processes in a successive manner, the misalignments between two input images can be gradually reduced through each registration process. It should be noted that the algorithm for feature-based image registration is not tied to a particular intensity-based image registration method, and vice versa.

Image Acquisition

[0036] In various embodiments, a consumer-grade digital camera can be used for image acquisition. For example, a Nikon D7100 camera and a Sigma 17-50 mm lens with the auto shooting mode may be utilized. The distance between the camera and the monitored structure may depend on the resolution of the camera, and a typical distance is 20 to 50 cm. In various embodiments, the camera can be held by hand during image acquisition, and images can directly capture the detected bolt and its adjacent structural surface without any obstructions. Ambient lighting conditions are generally acceptable. The image plane can be either parallel or skew to the monitored structural surface. When collecting the images at different inspection periods, the lighting condition and camera pose should be similar between inspection periods in order to produce the optimal result. This approach does not require camera calibration.

Feature-Based Image Registration

[0037] The purpose of feature-based image registration is to align two images into the same coordinate system using matched features (i.e. correspondences). For this approach to be viable, features (also known as feature points, comer points, or key points) are first detected in both images. Then, a matching algorithm is applied to the images to find matched features between two images, based on which a geometric transformation matrix can be estimated to transform the second image to the coordinate system of the first image.

[0038] FIG. 2 illustrates an example of the feature-based image registration algorithm according to various embodiments described herein. FIG. 2 illustrates two input images of a concrete column taken by a digital camera with a resolution of 6000 pixels x 4000 pixels.

The feature-based image registration algorithm can be performed to match the front face of the column in two input images. First, Image 1 can be denoted as the first input image (FIG. 2a). A ROI (3500 pixels x 3500 pixels) can be selected in Image 1, which is shown in FIG. 2b, to cover the front face of the column. Next, a feature-based image registration algorithm, such as a Shi-Tomasi algorithm, can be applied to the ROI in Image 1 to extract features.

The extracted features are denoted as feature set 203. This feature extraction procedure is flexible and can be achieved by many other feature types as well. As can be seen in FIG. 2d, feature points 206, 209, and 212 are typical Shi-Tomasi features, which are based on the unique intensity change at a localized region in both horizontal and vertical directions.

[0039] FIG. 3 illustrates certain aspects of the feature-based image registration algorithm according to various embodiments described herein. FIG. 3a illustrates the second image of the concrete column captured at a different position relative to the image in FIG. 2, denoted as Image 2. Shi-Tomasi features are extracted for the entire region of Image 2, denoted as feature set 303 in FIG 3b. Next, the approach can use the Kanade-Lucas-Tomasi (KLT) tracker or other algorithm to match each point in feature set 203 (FIG. 2B) to any potential points in feature set 303. In this example, a detection application executed in a computing device matched 1,370 features that were found in FIG. 3c, where the circles are features in Image 1 and the crosses represent features in Image 2. Among all the matched features, some outliers can be found (FIG. 3d), indicating matching failures. These outliers can be further eliminated utilizing, for example, a Maximum Likelihood Estimation Sample Consensus (MLESAC) algorithm, and new matched results (i.e. inliers) are shown in FIG 3e and FIG.

3f. Applying the MLESAC algorithm to Image 2 yields, for example, a total of 1,175 matched features, based on which a projective geometric transformation matrix can be estimated so that Image 2 can be registered to the coordinate system of Image 1. The projective geometric transformation can remove the projective distortion between Image 1 and Image 2 taken under different camera poses. The feature points can be matched to generate an image after image registration is performed, an example of which is illustrated in FIG.3g and FIG. 3h where the circles (features of Image 1) match the crosses (features of Image 2).

Intensity-Based Image Registration [0040] In some embodiments, the purpose of intensity-based image registration is to further align two images based on the images intensity distributions. Instead of geometric transformation in feature-based image registration, intensity-based image registration is a non-rigid transformation process. With reference to FIG. 4, shown is an example of images illustrating the principle of intensity -based image registration according to various embodiments described herein. The images include a first image at FIG 4a and a second image at FIG. 4b illustrating the same hand under different poses. Due to the different locations of the image capture device in these two images, feature-based image registration may face difficulties in aligning the two images. The intensity-based image registration approach uses, for example, an algorithm to non-rigidly register Image 2 of FIG. 4b to Image 3 of FIG. 4c, such as, the algorithm proposed by Thirion ¹ . A typical three-level pyramid with 500, 400, and 200 iterations is adopted during this procedure. FIG. 4d and e further evaluate the registration errors through intensity comparisons. Instead of misalignment occurring in the unregistered images (Image 1 and Image 2), the two images are well aligned after the registration (FIG. 4e).

[0041] Despite the great performance of intensity -based image registration, registration errors may still occur if abrupt intensity changes occurred in Image 2. In this example, the location of the ring at the ring finger is changed between two image

acquisitions, where the ring in Image 2 is closer to the fingertip. Such an action induces abrupt intensity changes in a localized region, leading to registration errors as shown in FIG. 4f. However, from the perspective of detecting bolt loosening, such registration errors can be utilized for identifying discrepancies between two images, serving as good features for bolt loosening detection. Feature Enhancement

[0042] FIG. 5 illustrates an example of the principle of feature enhancement according to various embodiments described herein. Once the feature-based image registration process and the intensity-based image registrations process are completed successively, the loosened bolt can be identified through registration errors as shown in FIG. 5a. Nevertheless, directly identifying the loosened bolt would still require human intervention as the loosened bolt is surrounded by noise content (FIG. 5a). The next step in the approach is to remove the noise content so that bolt loosening features around Bolt 2 can be enhanced. A number of image processing techniques are adopted in this procedure. First, a rectangular window 503 is applied to the registration errors (FIG. 5a) so that unrelated results can be filtered out by assigning 0 intensity to the pixels outside the window. The dimensions of the window are predefined as the same sizes of the ROI 109 (shown in FIG. 1) prior to feature- based image registration.

[0043] Next, an image segmentation method, such as, for example a method proposed by Achanta et al. ² , is performed to segment registration errors (FIG. 5b) into a series of localized regions, termed as superpixels, as shown in FIG. 5c. For each superpixel i, the coefficient of variation of intensities at all pixels within this superpixel is computed and denoted as CVi. Then, by applying a cutoff threshold T, noise content can be eliminated from registration errors so that bolt loosening features can be preserved (FIG. 5f). To explain, suppose two typical superpixels are selected in FIG. 5c, where Superpixel 1 is from the loosened bolt, and Superpixel 2 represents the noise content. As shown in FIG. 5d and FIG. 5e, magnitudes of intensities change dramatically around the loosened bolt, such as Superpixel 1, while transit occurs smoothly in other regions, such as Superpixel 2. In this regard, extracting coefficients of variation CVs of superpixels can efficiently separate the loosened bolt from its background noise. Hence, a feature enhancement algorithm is proposed by assigning 0 intensities to superpixels whose CVs are less than a predefined threshold T. For superpixels with CVs that are larger than the predefined threshold T, no action is required. Utilizing this algorithm, the noise content can be removed, and the final result is shown in FIG. 5f.

Result Implementation

[0044] FIG. 6 illustrates an example of the principle of result implementation according to various embodiments described herein. The purpose of result implementation is to map the bolt loosening features (FIG. 6a) to the original input image so that the loosened bolt can be easily visualized. To achieve this goal, a two dimensional Gaussian filter is applied to FIG. 6a to blur the bolt loosening features (FIG. 6b). Then, the filtered bolt loosening features are further converted to RGB channels using the following rules: 1) black color in FIG. 6b is converted into white color, and 2) white color in FIG. 6b is converted into red color. Finally, by setting up the transparency levels and overlapping RGB channels to the original input image, the loosened bolt can be successfully identified (FIG. 6d).

Experimental Validation

[0045] To validate the approach in accordance with various embodiments of the present disclosure, three experimental tests were conducted in the laboratory. A digital camera was adopted for image acquisition. The resolution of collected input images was 6000 pixels ^c 4000 pixels. Ambient lighting conditions were applied to all the tests during image acquisition. The bolts in the tests are made by ASTM A325 steel with diameter of 19.05 mm (3/4 in.), which are a common type of high strength bolts applied in steel construction in the United States. Shi-Tomasi features and the KLT tracker are adopted for feature-based image registration. [0046] FIG. 7 illustrates an example experimental validation test setup according to aspects of the embodiments described herein. FIG. 7 shows the tested steel joints used in various experiments. The steel joint in Test 1 was from a gusset plate in a cross frame, the steel joint in Test 2 was a steel column flange, and the steel joint in Test 3 was a web region of a steel girder. Table 1 summaries the different testing parameters in three experiments, in which a total number of bolts, loosened bolts, surface textures, and camera orientations vary in order to validate the performance of the present approach. The Matlab Computer Vision Toolbox was adopted for applying all the algorithms mentioned described herein.

Table 1. Test Matrix

[0047] FIG. 8 illustrates results of an example experimental validation test setup according to aspects of the embodiments described herein. The three bolts (FIG. 1) are in the gusset plate denoted as Bolt 1, Bolt 2, and Bolt 3 in Image 1 (FIG. 8a). During the inspection interval, Bolt 2 was rotated, and then Image 2 was collected as shown in FIG. 8b. FIG. 8c shows the initial intensity comparison of two images where significant errors can be found due to the different camera poses. To improve the matching performance, the feature-based and intensity -based image registrations are applied successively, and their registration errors are shown in FIG. 8d and FIG. 8e. The feature-based image registration is based on the ROI defined near the group of bolts (see the red block in FIG. 8a). Then a number of image processing techniques are further applied in order to enhance the bolt loosening features and visualize the loosened bolt as discussed herein. These techniques include windowing (FIG. 8f), superpixel segmentation (FIG. 8g), feature enhancement (FIG. 8h), Gaussian filtering (FIG. 8i), and result overlapping (FIG. 8j).

[0048] FIG. 9 illustrates results of an example experimental validation test setup according to aspects of the embodiments described herein. Shown in FIG. 9 are the experimental results of Test 2 (Fig. 7). Eight bolts were adopted in Test 2, and two of them (i.e. Bolt 3 and Bolt 6) experienced rotations during the inspection interval, as shown in FIG. 9b. Nevertheless, this approach is still able to identify the loosened bolts as illustrated in FIG.

9j-

[0049] FIG. 10 illustrates results of an example experimental validation test setup according to aspects of the embodiments described herein. Shown in FIG. 10 are the experimental results of Test 3 (FIG. 7), in which multiple conditions were varied in order to validate the performance of the approach of the present disclosure. In particular, the orientation of the camera was skewed to the monitored surface instead of being parallel to the fastener. The surface treatment of the structural surface, on the other side, was a combination of painted and unpainted, as can be seen in FIG. 7c. Prior to processing the images, the ROI (red block in FIG. lOa) should be selected to only cover the detected bolts and their adjacent structural surface, while excluding any background that is far away from the monitored surface. The benefit of such a selection is twofold: 1) the ROI can facilitate feature-based image registration process by specifying a localized region for matching potential

correspondences, and 2) the ROI can also exclude unnecessary registration errors during the feature enhancement procedure (see FIG. lOf). As shown in FIG. lOj, the loosened bolt (i.e. Bolt 2) can be detected.

[0050] As a summary of these experimental results, the approach of the present disclosure can successfully detect and localize single or multiple loosened bolts from a group of bolts, regardless of the total number of bolts, structural surface textures, or camera orientations. The success of this approach, however, relies on tuning the cutoff threshold T, a parameter in the feature enhancement algorithm discussed above. As shown in Table 1, T is 50 in both Test 1 and Test 3, while T increases to 200 for Test 2 because more noise content occurred in the registration errors in Test 2, as demonstrated in FIG. 9f. A detailed discussion about the effect of T will be presented in Section 4.4.

Input Image Resolution

[0051] FIG. 11 illustrates a comparison of the fastener loosening detection results using two different image resolutions and the proposed vision-based approach according to aspects of the embodiments described herein. The resolution of input images is 6000 pixels x 4000 pixels for experimental validation. However, lower resolution images can achieve successful results. A parametric study is performed by downsizing the original input images and repeating the bolt loosening detection procedure. The two input images of Test 2 (FIG. 7) are used to understand the effect of resolution on successfully determining fastener loosening. Two image resolutions are selected including: 1) 6000 pixels x 4000 pixels, as shown in FIG. 1 la and 2) 750 pixels x 500 pixels, as shown in FIG. 1 lb. FIG. 11 summarizes the bolt loosening detection results for each scenario.

[0052] As shown in FIG. 11, similar registration errors can be found after two image registration processes. The superpixel segmentation also demonstrates robust performance, despite a slightly different segmentation layout in each scenario. Nevertheless, two loosened bolts can be consistently localized regardless of the image resolutions. This further verifies that the key components of the present approach are insensitive against input image resolutions. This finding allows reductions of data storage and computational cost through the utilization of lower resolution images.

Rotation Angles of Bolt Head [0053] FIG. 12 illustrates a comparison of the fastener loosening detection results at various rotation angles using the proposed vision-based approach according to aspects of the embodiments described herein. The ability of the approach of the present disclosure to successfully identify fastener loosening for different rotation angles of the bolt head is demonstrated in FIG. 12. Setup of Test 1 (FIG. 7) was adopted in this investigation, where the middle bolt in FIG. l2a was subjected to a series of counterclockwise rotations of 60,

120, 180, 240, and 300 degrees, respectively. Images were taken at the initial stage with the unloosened bolt (FIG. l2a) and stages hereafter (FIG. l2b to FIG. l2f). Images with loosened bolts were further paired with the initial images for the purpose of bolt loosening detection. All images were collected by a digital camera. Image planes are parallel to the monitored surface. To enhance image processing efficiency, the original input images are downsized to 1500 pixels ^c 1000 pixels. As can be seen in FIG. 12, the approach of the present disclosure can consistently localize the loosened bolt under different rotation angles.

Features for Tracking

[0054] The feature-based image registration adopted in this approach also shows great potential to be applied to other research fields in the SHM community, such as targetless displacement monitoring of civil structures

[0055] FIG. 13 illustrates a comparison of the performance of the Shi-Tomasi features for the feature-based image registration with other common types of feature-based image registration processes according to aspects of the embodiments described herein. FIG. 13 shows a comparison of the performance of the Shi-Tomasi features (FIG. 13 a) with an accelerated segment test (FAST) (FIG. l3b), Harris-Stephens (FIG. l3c), binary robust invariant scalable keypoints (BRISK) (FIG. l3d), and speeded up robust features (SURF) (FIG. l3de). Briefly, two input images of Test 3 in Section 3 (FIG. lOa and FIG. lOb) are adopted for the comparison. Five different types of features are extracted in the first input image within the ROI as shown in the first column of FIG. 13. Despite the total number and locations of these features (see the second column in FIG. 13), feature-based image registration can be successfully performed as shown in the third column of FIG. 13. As shown in the last columns of FIG. 13, registration errors can be significantly reduced after intensity-based image registration. In this regard, the loosened bolt can be consistently identified by this approach regardless of feature type.

Cutoff Threshold T of Image Segmentation

[0056] FIG. 14 illustrates a comparison of the results of the feature enhancement at various threshold levels using the proposed vision-based approach according to aspects of the embodiments described herein. As discussed above, a cutoff threshold T is used for eliminating noise content from the registration errors. Various experiments were performed to demonstrate the sensitivity of T in the process of feature enhancement. Two input images in Test 2 of Section 3 are adopted for this analysis, and results are shown in FIG. 14. As shown in FIG. 14, a larger cutoff threshold T can eliminate noise context in the initial registration errors; however, the bolt loosening features may also be deleted (see subfigure when T = 1000). On the other hand, a smaller cutoff threshold T can preserve bolt loosening features.

As a tradeoff, noise content may exist, as shown in the second subfigure when T = 50, leading to challenges in localizing the loosened bolts. For such reasons, T = 200 is adopted in the experiment in Section 2, while selecting a region of cutoff threshold T (from 200 to 600) may also be achievable for this particular dataset. A practical approach for determining the optimal cutoff threshold T would be a trial-and-error procedure. An initial T = 50 is suggested for the tests in this study and can be further adjusted based on the tuning result. Lishtins Condition

[0057] FIG. 15 illustrates a comparison of the fastener loosening detection results at lighting conditions using the proposed vision-based approach according to aspects of the embodiments described herein. Lighting condition is another important parameter in the approach of the present disclosure. Further tests were performed with varying lighting conditions. Setup of Test 1 was adopted in this analysis. As shown in FIG. 15b, the lighting condition varied by adding an additional light source with a floor lamp, leading to slight changes of shadows in the second input image. Bolt 1 (FIG. l5a) experienced a rotation under the inspection interval. Nevertheless, this approach can still detect the loosened bolt under such a condition.

[0058] Despite the success of the present approach in this particular investigation, a significant change of lighting condition around the bolts could affect performance as significant changes in lighting conditions would provoke extensive intensity change, inducing excessive registration errors. For instance, the new shadow of the angle caused by the lighting change denoted in the second input image (FIG. 15b) cannot be eliminated by the two image registration processes, hence the registration error appears in FIG. l5e. If such a change in lighting condition occurs around the bolts, the robustness of this approach would be affected. Nut Loosening

[0059] FIG. 16 illustrates results of an example experimental validation test setup according to aspects of the embodiments described herein. Nut loosening is another common phenomenon caused by self-loosening of the bolt. Utilizing the proposed methodology, nut loosening can also be detected. FIG. 16 illustrates an example through the setup of Test 1 (FIG. 7). Instead of bolt heads, nuts were installed at the facial side of the gusset plate, as shown in Image 1 (FIG. l6a). The third nut from the left experienced a counterclockwise rotation (about 15 degree) during inspection interval, and then Image 2 was collected (FIG. l6b). The result indicates that the approach is able to identify the loosened nut (FIG. l6j).

Bolt Type [0060] FIG. 17 illustrates a test specimen according to aspects of the embodiments described herein. FIG. 18 illustrates results of an example experimental validation according to aspects of the embodiments described herein applied to the test specimen of FIG. 17.

[0061] Application of the present approach for a different bolt type is demonstrated in FIG. l8a. A double angle steel joint with two bolts was adopted in this experiment. The dimensions of the double angles were 2L76.2mm ^c 50.8 mm x 4.8 mm (2L3in. x 2in. x 3/l6in.). The diameter of each bolt was 7.9 mm (5/16 in.), which is much smaller than the bolt (19.05 mm) applied in Section 3. The second nut from the left was rotated about 30 degrees in the counterclockwise direction, as shown in Image 2 (FIG. 18b). Such a rotation leads to registration errors around the loosened nut, which can be detected by the present approach, as shown in FIG. 18j .

Gay Caused by Nut Loosening

[0062] Instead of finding the rotation of bolts’ heads and nuts, an alternative strategy for bolt loosening detection is to identify the change of the gap between the nut and the bolted surface. This strategy would be particularly useful for practical implementation if the front views of the bolt heads and/or nuts are difficult to obtain in field conditions (e.g. the space in front of the monitored structure is occupied by other objects). FIG. 19 illustrates results of an example experimental validation test setup according to aspects of the embodiments described herein. As demonstrated in FIG. l9b, the loosened nut results in a gap at the second bolt. The changes of intensities associated with this outward movement of the nut become features of bolt loosening detection (FIG. l9j).

[0063] In various embodiments of the present disclosure, the fastener can include a nut, a bolt, a screw, or any other type of fastener that can accurately be identified using the approach presented herein. [0064] In various embodiments, images are captured using a consumer-grade digital camera. In various other embodiments, images may be acquired from a video feed using suitable methods, as will be apparent to one skilled in the art.

[0065] The embodiments described herein can be embodied in hardware, software, or a combination of hardware and software. If embodied in software (in part), each procedural step or element can be embodied as a module or group of code that includes program instructions to implement the specified logical function(s). The program instructions can be embodied in the form of, for example, source code that includes human-readable statements written in a programming language or machine code that includes machine instructions recognizable by a suitable execution system, such as a processor in a computer system or other system.

[0066] If embodied in hardware (in part), each element can represent a circuit or a number of interconnected circuits that implement the specified logical function(s). The hardware can be implemented as a circuit or state machine that employs any suitable hardware technology. The hardware technology can include, for example, one or more microprocessors, discrete logic circuits having logic gates for implementing various logic functions upon an application of one or more data signals, application specific integrated circuits (ASICs) having appropriate logic gates, programmable logic devices ( e.g ., field- programmable gate array (FPGAs), and complex programmable logic devices (CPLDs)).

[0067] The hardware can also include at least one processing circuit. Such a processing circuit can include, for example, one or more processors and one or more storage or memory devices coupled to a local interface. The local interface can include, for example, a data bus with an accompanying address/control bus or any other suitable bus structure. The storage or memory devices can store data or components that are executable by the processors of the processing circuit. [0068] Also, one or more or more of the components described herein that include software or program instructions can be embodied in a non-transitory computer-readable medium for use by or in connection with an instruction execution system, such as a processor or processing circuit. The computer-readable medium can contain, store, and/or maintain the software or program instructions for use by or in connection with the instruction execution system.

[0069] A computer-readable medium can include a physical media, such as, magnetic, optical, semiconductor, and/or other suitable media. Examples of a suitable computer- readable media include, but are not limited to, solid-state drives, magnetic drives, or flash memory. Further, any logic or component described herein can be implemented and structured in a variety of ways. For example, one or more components described can be implemented as modules or components of a single application. Further, one or more components described herein can be executed in one computing device or by using multiple computing devices.

[0070] The above-described examples of the present disclosure are merely possible examples of implementations set forth for a clear understanding of the principles of the disclosure. Many variations and modifications can be made without departing substantially from the spirit and principles of the disclosure. All such modifications and variations are intended to be included herein within the scope of this disclosure and protected by the following claims.

[0071] Clause 1. A method, comprising capturing a first image of a fastener at a first time; capturing a second image of the fastener at a second time; performing a feature-based image registration to create a third image; performing an intensity-based image registration to create a fourth image; determining registration errors based at least in part on a comparison of the first image and the fourth image; and performing a feature enhancement process on the registration errors to determine whether the fastener has loosened between the first time and the second time.

[0072] Clause 2. The method of clause 1, wherein the performing the feature-based image registration comprises using a feature-based image registration algorithm.

[0073] Clause 3. The method of clause 1 or 2, further comprising defining a region of interest in the first image.

[0074] Clause 4. The method of any of clauses 1-3, wherein performing the feature enhancement process on the registration errors comprises creating a fifth image that identifies whether the fastener has loosened at the second time.

[0075] Clause 5. The method of any of clauses 1-4, further comprising aligning the second image based at least in part on a coordinate system of the first image.

[0076] Clause 6. The method of any of clauses 1-5, further comprising reducing differences between the first image and the third image by registering the third image to the fourth image.

[0077] Clause 7. The method of any of clauses 1-6, wherein performing the feature enhancement process comprises: segmenting the registration errors into a plurality of regions; analyzing each of the plurality of regions to determine a coefficient of variation for each of the plurality of regions, the coefficient of variation for each of the plurality of regions being based at least in part on a variation of intensity; and removing any region among the plurality of regions that has a coefficient of variation below a threshold value.

[0078] Clause 8. The method of any of clauses 1-7, further comprising visually identifying that the fastener has loosened based at least in part on any region among the plurality of regions that has a coefficient of variation above the threshold value.

[0079] Clause 9. The method of any of clauses 1-8, wherein performing the feature- based image registration comprises comparing the first image and the second image, and performing the intensity-based image registration comprises comparing the first image and one of the second image or the third image.

[0080] Clause 10. A system for fastener loosening detection, comprising: a capture device; at least one computing device configured to at least: obtain a first image of a fastener at a first time based at least in part on the capture device; obtain a second image of the fastener at a second time based at least in part on the capture device; perform a feature-based image registration to create a third image; perform an intensity-based image registration to create a fourth image; determine registration errors between the first image and the fourth image; and perform a feature enhancement process on the registration errors to determine whether the fastener has loosened between the first time and the second time.

[0081] Clause 11. The system of clause 10, wherein the at least one computing device is further configured to at least define a region of interest in the first image.

[0082] Clause 12. The system of clause 10 or 11, wherein the at least one computing device is further configured to at least perform the feature enhancement process by creating a fifth image that identifies whether the fastener has loosened at the second time.

[0083] Clause 13. The system of any of clauses 10-12, wherein the at least one computing device is further configured to at least align the second image based at least in part on a coordinate system of the first image.

[0084] Clause 14. The system of any of clauses 10-13, wherein the at least one computing device is further configured to at least reduce differences between the first image and one of the third image by registering the third image to the fourth image.

[0085] Clause 15. The system of any of clauses 10-14, wherein the at least one computing device is configured to perform the feature enhancement process by: segmenting the registration errors into a plurality of regions; analyzing each of the plurality of regions to determine a coefficient of variation for each of the plurality of regions, the coefficient of variation for each of the plurality of regions being based on a variation of intensity; and removing any region among the plurality of regions that has a coefficient of variation below a threshold value.

[0086] Clause 16. The system of clause 15, wherein the at least one computing device is further configured to at least identify that the fastener has loosened based on any region among the plurality of regions that has a coefficient of variation above the threshold value.

[0087] Clause 17. A non-transitory computer-readable medium embodying a program that, when executed by at least one computing device, causes the at least one computing device to at least: capture a first image of a fastener at a first time; capture a second image of the fastener at a second time; perform a feature-based image registration to create a third image; perform an intensity-based image registration to create a fourth image; determine registration errors between the first image and the fourth image; and determine whether the fastener has loosened between the first time and the second time based at least in part on a feature enhancement process algorithm and the registration errors.

[0088] Clause 18. The non-transitory computer-readable medium of clause 17, wherein the program further causes the at least one computing device to at least define a region of interest in the first image.

[0089] Clause 19. The non-transitory computer-readable medium of clause 17 or 18, wherein the program further causes the at least one computing device to at least align the second image based at least in part on a coordinate system of the first image.

[0090] Clause 20. The non-transitory computer-readable medium of any of clauses 17- 19, wherein the program further causes the at least one computing device to at least reduce differences between the first image and one of the third image by registering the third image to the fourth image.