A computer vision-based long-term monitoring algorithm and system for bridge structure vibration

Abstract

The application provides a bridge structure vibration long-term monitoring algorithm and system based on computer vision, and relates to the technical field of beam monitoring.Aiming at the fact that there is no high-precision, strong-robustness and blockage self-recovery bridge monitoring algorithm based on vision in the prior art, the application adopts confidence degree calculation of matching points to remove matching points with low confidence degree score, calculates displacement of the bridge at multiple feature points to generate vibration time series data, extracts inherent frequency and corresponding modal shape and compares the inherent frequency and the corresponding modal shape with a benchmark frequency range and a benchmark modal shape to obtain a monitoring result, and the application performs blockage self-recovery in view of environmental interference, thereby improving the precision and reliability of bridge monitoring and improving the robustness of bridge monitoring activities in complex scenes.

Description

Technical Field

[0001] This invention relates to the field of bridge monitoring technology, and in particular to a long-term monitoring algorithm and system for bridge structural vibration based on computer vision. Background Technology

[0002] With the rapid development of computer vision and image processing technologies, vision-based structural health monitoring methods have demonstrated enormous application potential and development prospects in the field of bridge engineering monitoring due to their unique advantages of being non-contact, providing full-field measurement, flexible deployment, and low cost. Compared with traditional sensors, vision methods only require acquiring a sequence of structural images through a camera to reconstruct the full-field displacement response of the structure under load. This not only avoids the damage to the structure caused by installing sensors but also acquires spatially dense data far exceeding that of point measurements, providing a rich information foundation for comprehensively assessing the overall performance of the structure and identifying damage. This technical approach is expected to become a powerful supplement and even an important component of the existing monitoring system, and is a key technological direction for realizing intelligent and long-life operation and maintenance of infrastructure.

[0003] However, existing computer vision-based bridge monitoring methods still have significant shortcomings, especially for long-term, automated monitoring tasks. First, in complex and variable natural environments, existing methods lack robustness, leading to a significant reduction in the stability and reliability of their core component—image feature matching. This results in mismatches or matching failures, causing gross errors or interruptions in subsequent displacement data, ultimately lowering the overall accuracy and reliability of the monitoring system. Second, existing technologies are insufficient in distinguishing between environmental interference and actual structural damage. They also lack the self-recovery capability to address real-world problems such as dust accumulation on camera lenses during long-term operation, which degrades image quality and further reduces the overall accuracy and reliability of the monitoring system.

[0004] Therefore, developing a high-precision, robust, and self-recovering algorithm and system for long-term monitoring of bridge structure vibration based on computer vision is of great significance for long-term monitoring of bridge structure vibration. Summary of the Invention

[0005] To address this, the present invention provides a long-term monitoring algorithm and system for bridge structure vibration based on computer vision, which overcomes the problems of perception failure due to poor feature matching robustness and misdiagnosis due to weak environmental interference discrimination ability in the prior art.

[0006] To achieve the above objectives, this invention provides a computer vision-based long-term vibration monitoring algorithm for bridge structures, comprising:

[0007] S1, responding to the trigger instruction generated by the microcontroller's timer, controls the camera to capture timing images of the target bridge. The timing images include a reference image and multiple target images.

[0008] S2, extract multiple feature points from the reference image, and match each feature point in the reference image with each target image to obtain a preliminary matching point pair;

[0009] S3, generate the confidence score for each matching point pair, and remove matching point pairs with confidence scores lower than the first threshold to obtain the target matching point pair set;

[0010] S4. Based on the target matching point pair set, calculate the displacement of the target bridge at multiple feature points to generate vibration time series data of the target bridge.

[0011] S5. Based on vibration time series data, extract at least one natural frequency and corresponding mode shape of the target bridge. The mode shape is used to characterize the spatial deformation of the target bridge when it vibrates at the corresponding natural frequency.

[0012] S6 compares the natural frequency and mode shape with the preset reference frequency range and reference mode shape to obtain the monitoring results.

[0013] Furthermore, based on the target matching point pair set, the displacement of the target bridge at multiple feature points is calculated to generate vibration time-series data of the target bridge, including:

[0014] S41, Based on the target matching point pair set, calculate the pixel coordinate displacement vector between the feature points in the reference image and the corresponding matching points in the target image;

[0015] S42, based on utilizing the gray-level gradient changes of the image sequence in the spatiotemporal dimension, and / or by fitting the gray-level distribution of the feature point neighborhood through a surface, optimize the pixel coordinate displacement vector with sub-pixel precision to obtain the image plane displacement amount with sub-pixel precision.

[0016] S43, based on pre-calibrated camera parameters, transforms the image plane displacement into the physical displacement of the target bridge in three-dimensional space through coordinate transformation relationships. The camera parameters include internal camera parameters and external camera parameters relative to the target bridge.

[0017] S44, based on the correspondence between the timestamp corresponding to the trigger instruction of the microcontroller and each feature point, combines the physical displacement into a displacement-time sequence;

[0018] S45, aggregate the displacement-time series of all feature points to obtain vibration time series data to characterize the vibration response of the target bridge at multiple discrete locations.

[0019] Furthermore, the natural frequencies and mode shapes are compared with preset reference frequency ranges and reference mode shapes to obtain monitoring results, including:

[0020] Calculate the relative deviation between the current natural frequency and the reference frequency range. When the relative deviation exceeds the preset frequency deviation threshold, a frequency anomaly flag is generated. Calculate the modal guarantee criterion value between the current mode shape and the reference mode shape. When the modal guarantee criterion value is lower than the preset modal correlation threshold, a mode shape anomaly flag is generated.

[0021] When both frequency anomaly indicators and mode shape anomaly indicators exist simultaneously, the following steps are taken: Calculate the difference in mode shape curvature between the current mode shape and the reference mode shape at each characteristic point; identify continuous regions where the difference in mode shape curvature exceeds a preset curvature threshold; determine the location corresponding to the maximum value of the difference in mode shape curvature within the continuous region as the core damage region; calculate the relative deviation values ​​of several consecutive natural frequencies and assign weight coefficients based on the differences in the sensitivity of different frequencies to damage; calculate a comprehensive score based on the relative deviation values ​​of the natural frequencies and the corresponding weight coefficients; determine the damage severity level based on the correspondence between the comprehensive score and the damage severity level, where lower-order frequencies are assigned greater weight than higher-order frequencies; determine the structural risk level based on the location of the core damage region within the bridge structure; generate maintenance priorities using a predefined decision matrix based on the damage severity level and the structural risk level; and generate monitoring results characterizing the confirmation of structural damage based on the location information of the core damage region, the damage severity level, and the maintenance priorities.

[0022] Furthermore, when only frequency anomaly indicators exist:

[0023] Acquire the current ambient temperature data and compare and verify the current natural frequency with the ambient temperature-frequency correlation model established based on historical monitoring data;

[0024] When the frequency change conforms to the temperature change law and the mode shape characteristics remain stable, the camera is controlled to perform a preset swing action to reduce the impact of environmental interference and generate monitoring results characterizing the impact of environmental interference.

[0025] Furthermore, the camera is controlled to perform preset swaying movements to reduce the impact of environmental interference, including:

[0026] The swing amplitude of the preset swing action is determined based on the image quality of the time-series images;

[0027] The oscillation frequency of the preset oscillation motion is determined based on the rate of dust accumulation in the environment.

[0028] The timing of the pre-set swinging motion is determined based on the requirements of the target detection task;

[0029] Control commands for the preset swing motion are generated based on the swing amplitude, swing frequency, and swing timing of the preset swing motion.

[0030] Furthermore, the oscillation frequency of the preset oscillation motion is determined based on the rate of dust accumulation in the environment, including:

[0031] Obtain the concentration of dust in the environment, the ambient humidity, and the characteristic parameters of the camera lens;

[0032] The dust viscosity coefficient is determined based on the ambient humidity and the concentration of dust in the environment.

[0033] Based on the dust viscosity coefficient and the characteristic parameters of the camera lens, the type of dust adhesion is determined.

[0034] Based on the adhesion type of dust, determine the rate of dust accumulation in the environment;

[0035] The oscillation frequency of the preset oscillation motion is determined based on the rate of dust accumulation in the environment.

[0036] Furthermore, the characteristics of a camera lens include hydrophobicity, antistatic properties, and surface roughness.

[0037] Furthermore, if no frequency anomaly indicator or mode shape anomaly indicator is found within the preset number of tests:

[0038] Confirm that the target bridge is in a healthy state and generate monitoring results to characterize the health confirmation of the target bridge;

[0039] Based on the monitoring data within a preset number of tests, the reference frequency range is updated and the reference mode shape is optimized;

[0040] Based on the duration of the target bridge being in a healthy state, a timer trigger frequency adjustment instruction for the microcontroller is generated.

[0041] Furthermore, each feature point in the baseline image is matched with each target image to obtain preliminary matching point pairs, including:

[0042] S21, extract image features from the baseline image and the target image. Image features include feature point locations and corresponding high-dimensional feature descriptors.

[0043] S22: Based on the high-dimensional feature descriptors of feature points in the reference image and the target image, forward matching is performed from the reference image to the target image by calculating the similarity matrix between the feature descriptors, to find candidate matching points in the target image for each feature point in the reference image; backward matching is performed from the target image to the reference image, to find candidate matching points in the reference image for each feature point in the target image.

[0044] S23, perform a consistency check on the results of forward matching and backward matching, retain the corresponding matching point pairs in forward matching and backward matching, and obtain the preliminary matching point pairs.

[0045] This invention also provides a computer vision-based long-term vibration monitoring system for bridge structures, used to implement any computer vision-based long-term vibration monitoring algorithm for bridge structures, including:

[0046] The image generation unit is used to control the camera to capture a timing image of the target bridge in response to the trigger command generated by the timer of the microcontroller. The timing image includes a reference image and multiple target images.

[0047] The feature matching unit, connected to the image generation unit, is used to extract multiple feature points in the reference image and match each feature point in the reference image with each target image to obtain a preliminary matching point pair.

[0048] The matching filtering unit, connected to the feature matching unit, is used to generate a confidence score for each matching point pair and remove matching point pairs with confidence scores lower than a first threshold to obtain the target matching point pair set.

[0049] The displacement field reconstruction unit, connected to the matching filter unit, is used to calculate the displacement of the target bridge at multiple feature points based on the target matching point pair set, so as to generate the vibration time series data of the target bridge.

[0050] The modal parameter identification unit, connected to the displacement field reconstruction unit, is used to extract at least one natural frequency and corresponding mode shape of the target bridge based on vibration time series data. The mode shape is used to characterize the spatial deformation morphology of the target bridge when it vibrates at the corresponding natural frequency.

[0051] The structural health diagnostic unit, connected to the modal parameter identification unit, is used to compare the natural frequencies and mode shapes with preset reference frequency ranges and reference mode shapes to obtain monitoring results.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0053] First, this invention ensures the reliability of matching point pairs by filtering based on feature point confidence, reducing error input at the source. Then, it employs a sub-pixel optimization algorithm based on grayscale gradient changes and surface fitting to improve displacement measurement accuracy from the pixel level to 0.01 pixel level. Finally, through rigorous camera calibration and coordinate transformation models, it eliminates lens distortion and projection errors, accurately converting image displacement into three-dimensional physical displacement. This end-to-end precision control, from feature matching to displacement calculation, solves the technical bottleneck of existing visual monitoring methods that struggle to capture minute structural deformations due to insufficient precision, thus improving the accuracy and reliability of bridge monitoring.

[0054] Secondly, this invention constructs a progressive diagnostic process that includes frequency-mode dual verification, curvature damage localization, multi-order frequency weighted scoring, risk area matching, and maintenance priority determination. By collaboratively judging frequency and mode shape, it effectively distinguishes between actual damage and environmental interference; furthermore, it achieves precise location of the damaged area through mode shape curvature analysis; then, it achieves quantitative assessment of the damage degree through a weighted scoring system considering the sensitivity of different frequency orders; finally, it combines technical parameters with structural risk maps to output maintenance priorities that can directly guide maintenance decisions. This further improves the accuracy and reliability of bridge monitoring.

[0055] Third, this invention controls camera sway based on multi-source information such as image quality and dust accumulation rate, improving the self-recovery capability of lens contamination. By automatically updating the reference frequency and mode shape in a healthy state, the diagnostic criteria evolve naturally with the structural state. These self-adjusting and self-learning mechanisms enable the system to cope with long-term environmental changes and structural performance evolution, improving the robustness of bridge monitoring activities in complex scenarios. Attached Figure Description

[0056] Figure 1 This is a flowchart of a long-term monitoring algorithm for bridge structure vibration based on computer vision, according to an embodiment of the present invention.

[0057] Figure 2 This is a structural block diagram of a long-term bridge structure vibration monitoring system based on computer vision, according to an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0059] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0060] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate the direction or positional relationship, are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0061] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0062] Example 1

[0063] like Figure 1 As shown, this invention proposes a long-term vibration monitoring algorithm for bridge structures based on computer vision, which specifically includes the following steps:

[0064] Step S1: In response to the trigger command generated by the timer of the microcontroller, control the camera to capture a timing image of the target bridge. The timing image includes a reference image and multiple target images.

[0065] For example, a 32-bit high-performance microcontroller based on the ARM architecture is used, which has a high-precision timer peripheral.

[0066] An industrial-grade global shutter camera with an external trigger interface is used, which is connected to the main control device via USB.

[0067] Specifically, the sampling rate is 50Hz, the reference image update cycle is 24 hours, and the number of images captured in a single burst (i.e., the number of target images captured consecutively after each trigger) is 1024. The acquired images are first temporarily stored in the camera's internal buffer, and then transferred in batches via the data interface to the microcontroller's SD card or external RAM for temporary storage. During transmission, the microcontroller automatically timestamps each image, and this timestamp strictly corresponds to the interrupt point of the timer.

[0068] Upon initial run or when the baseline image update time arrives, the system designates the first high-quality, unblurred image captured at the current moment as the baseline image. This image serves as the reference benchmark for feature matching of all target images in this round. Each image captured after a timer trigger is marked as a target image. The system will continuously acquire 1024 target images, forming a data packet for analysis.

[0069] In addition to periodic updates, the system also incorporates event-triggered update logic. For example, when the average brightness of the image changes abruptly, or when image analysis detects a significant shift in the camera's field of view, the system can automatically trigger the reacquisition of the baseline image, thereby enhancing the system's adaptability to long-term environmental changes.

[0070] This invention enables reliable time-series image acquisition under precise time control and provides a high-quality and time-accurate data source for the entire monitoring system.

[0071] Step S2: Extract multiple feature points from the reference image, and match each feature point in the reference image with each target image to obtain preliminary matching point pairs, including:

[0072] S21, extract image features from the baseline image and the target image. Image features include feature point locations and corresponding high-dimensional feature descriptors.

[0073] S22: Based on the high-dimensional feature descriptors of feature points in the reference image and the target image, forward matching is performed from the reference image to the target image by calculating the similarity matrix between the feature descriptors, to find candidate matching points in the target image for each feature point in the reference image; backward matching is performed from the target image to the reference image, to find candidate matching points in the reference image for each feature point in the target image.

[0074] S23, perform a consistency check on the results of forward matching and backward matching, retain the corresponding matching point pairs in forward matching and backward matching, and obtain the preliminary matching point pairs.

[0075] For example, the AKAZE algorithm is used for feature extraction. Specifically, in the baseline image, the AKAZE algorithm constructs a nonlinear scale space, then detects extreme points in the scale space as stable feature points, extracting M feature points from the baseline image. For each detected feature point, its principal direction is calculated, and a 61-dimensional binary descriptor is extracted, which represents the gray-level gradient information of the feature point's neighborhood. The same operation is performed on the target image to obtain N feature points and their descriptors.

[0076] For each feature point descriptor in the baseline graph, calculate its Hamming distance to all feature point descriptors in the target graph. Find the two target graph feature points with the smallest Hamming distance and calculate the ratio of the nearest neighbor distance to the second nearest neighbor distance. If the ratio is lower than the first threshold, the feature point is considered a candidate matching point of the baseline graph feature point and a forward candidate pair is formed. Otherwise, the point is considered to have no match.

[0077] Perform the operation symmetrical to the forward matching, starting from the target image and ending at the reference image, repeat the ratio test of the nearest neighbor and the second nearest neighbor as the feature points in the target image, and search for candidate matching points in the reference image to form reverse candidate pairs.

[0078] Iterate through all forward candidate pairs and check if there is a symmetric pair of the forward candidate pairs in the set of backward matching candidate pairs. Only when they correspond to each other in forward and backward matching are the point pairs reserved as pre-matching point pairs and recorded in the set of matching point pairs.

[0079] This invention constructs a powerful mismatch filtering mechanism through a combination of bidirectional matching and consistency checks. This mechanism effectively identifies and eliminates ambiguous matching pairs generated in unidirectional matching due to image noise, sudden changes in illumination, repetitive textures, or local occlusion. This enhances the algorithm's adaptability to complex and variable engineering environments.

[0080] Step S3: Generate the confidence score for each matching point pair and remove matching point pairs with confidence scores lower than the first threshold to obtain the target matching point pair set;

[0081] For each initial matched point pair in the set of matched point pairs, a confidence score S is calculated. In this embodiment, the confidence score S consists of a matching distance score Sa and a local consistency score Sb. Specifically, based on the Hamming distance in forward matching, the smaller the distance, the higher the score. The Hamming distance is normalized to the [0,1] interval to obtain the matching distance score Sa. Taking this matched point pair as the center, the consistency of the displacement vectors of other matched point pairs in a small neighborhood around it, such as a 5x5 window, is examined through cosine similarity. The higher the consistency of the displacement direction and amplitude, the higher the local consistency score.

[0082] The overall confidence level is the sum of the products of the matching distance score Sa and the local consistency score Sb and their respective weight coefficients. In this embodiment, the weight coefficient of the matching distance score Sa is 0.7 and the weight coefficient of the local consistency score Sb is 0.3.

[0083] Optionally, the optional implementation range of the first threshold is [0.5, 0.7]. Preferably, the preferred embodiment of the first threshold is 0.6. Traverse all initial matching point pairs, remove point pairs with a comprehensive confidence level less than the preset threshold, and output the retained high-confidence point pairs.

[0084] This invention introduces a comprehensive evaluation index that integrates matching distance and local spatial consistency through confidence-based screening. This index not only eliminates obvious erroneous matches but also finely identifies potential error points that appear reasonable but actually deviate from the local deformation trend. This is equivalent to adding a layer of smoothness constraint based on prior structural mechanics on top of traditional geometric matching, thereby screening out a more reasonable and accurate set of preliminary matching point pairs in spatial distribution, providing a clean and reliable input for subsequent displacement calculations with sub-pixel accuracy.

[0085] Step S4: Based on the target matching point pair set, calculate the displacement of the target bridge at multiple feature points to generate vibration time-series data of the target bridge, including:

[0086] S41, Based on the target matching point pair set, calculate the pixel coordinate displacement vector between the feature points in the reference image and the corresponding matching points in the target image;

[0087] S42, based on utilizing the gray-level gradient changes of the image sequence in the spatiotemporal dimension, and / or by fitting the gray-level distribution of the feature point neighborhood through a surface, optimize the pixel coordinate displacement vector with sub-pixel precision to obtain the image plane displacement amount with sub-pixel precision.

[0088] S43, based on pre-calibrated camera parameters, transforms the image plane displacement into the physical displacement of the target bridge in three-dimensional space through coordinate transformation relationships. The camera parameters include internal camera parameters and external camera parameters relative to the target bridge.

[0089] S44, based on the correspondence between the timestamp corresponding to the trigger instruction of the microcontroller and each feature point, combines the physical displacement into a displacement-time sequence;

[0090] S45, aggregate the displacement-time series of all feature points to obtain vibration time series data to characterize the vibration response of the target bridge at multiple discrete locations.

[0091] Specifically, the target matching point pair set is read. For each matching point in the set, the coordinates of the corresponding feature point in the reference image are subtracted from the coordinates of the point in the target image to obtain the pixel coordinate displacement vector. An optimization algorithm is then used for fine adjustment to obtain a displacement amount that exceeds the precision of integer pixels.

[0092] The pixel coordinate displacement vector is optimized and adjusted using optical flow based on gray-level gradient and surface fitting method;

[0093] The optical flow method based on gray-level gradients assumes that the gray-level distribution in a small region around a feature point remains stable in a continuous image. It calculates the gray-level gradient of the image sequence in the spatiotemporal dimension and iteratively searches for the sub-pixel displacement that minimizes the gray-level difference.

[0094] The surface fitting method locates the position with sub-pixel precision by fitting a gray-level distribution surface (such as a quadratic surface) within the neighborhood of the matching point and finding its extreme points.

[0095] Using pre-calibrated camera parameters, the displacement in the image is converted into the displacement of the bridge in the real 3D world. Specifically, parameters obtained through camera calibration are called, including intrinsic camera parameters such as focal length and principal point, and extrinsic camera parameters relative to the bridge, such as position and orientation. Based on the pinhole imaging model and the geometric constraints of the bridge's vibration plane, a mapping relationship is established from the image coordinate system to the bridge's world coordinate system. Through this mapping relationship, sub-pixel displacement is directly converted into 3D spatial displacement with physical units.

[0096] Based on the precise timestamps corresponding to the microcontroller's trigger commands, the displacement of each feature point is aligned on the time axis to form a chronologically ordered sequence, resulting in an independent displacement-time series for each feature point, characterizing its vibration history during the monitoring period. The time-series data of all spatially discrete feature points are aligned and integrated along the time dimension to construct a multi-dimensional dataset, generating the final vibration time-series data. This data simultaneously contains spatial information (the locations of multiple monitoring points on the bridge) and temporal information (the vibration time history of each point), comprehensively characterizing the dynamic vibration response of the target bridge.

[0097] This invention improves displacement measurement accuracy from the pixel level to the 0.01 pixel level through sub-pixel precision optimization, fundamentally enhancing the raw data accuracy. Subsequently, through rigorous coordinate transformation, it eliminates systematic errors caused by lens distortion and viewpoint projection using pre-calibrated camera parameters, improving the accuracy of the final physical displacement and enhancing the reliability of the monitoring algorithm. By constructing displacement-time series and aggregated vibration time series data, this invention systematically organizes discrete data points in space and time into a complete dataset containing information on spatial distribution and temporal evolution. This structured data format perfectly meets the input requirements of subsequent modal parameter identification algorithms, providing direct and high-quality data support for accurately extracting deep dynamic characteristics of bridges, such as natural frequencies, damping ratios, and mode shapes.

[0098] Step S5: Based on vibration time series data, extract at least one natural frequency and corresponding mode shape of the target bridge. The mode shape is used to characterize the spatial deformation of the target bridge when it vibrates at the corresponding natural frequency.

[0099] Specifically, the vibration time-series data from S45 includes a number of characteristic points and measurement points. Each measurement point contains a displacement-time series matrix of a number of sampled data points. The displacement sequence of each measurement point is detrended to eliminate slow-varying trend terms caused by temperature drift, etc. Subsequently, based on the expected frequency range of bridge vibration, a bandpass filter is used to retain the signal in that frequency band while filtering out high-frequency noise and extremely low-frequency interference.

[0100] Preferably, frequency domain decomposition is used as the core identification algorithm. For the time-series data of each measurement point in the preprocessed displacement data matrix, the Welch method is used to estimate its power spectral density. A power spectral density matrix is ​​assembled at all measurement points and all frequency points. Singular value decomposition is performed on the power spectral density matrix at each discrete frequency point. All singular values ​​corresponding to each frequency are connected to form a curve, creating a set of curves showing the change of singular values ​​with frequency. On the singular value spectrum, the frequency corresponding to the obvious peak of a certain order singular value curve is the first-order natural frequency. The first order corresponds to the highest peak, and subsequent peaks correspond to the second, third, and other higher-order frequencies. For the identified Nth-order natural frequency, the singular value decomposition result corresponding to that frequency point is found, and the left singular vector corresponding to the largest singular value is taken. This vector is normalized to obtain the mode shape corresponding to that natural frequency. This mode shape intuitively represents the spatial deformation morphology of each feature point when the bridge vibrates at a specific frequency.

[0101] This invention uses modal analysis algorithms (such as frequency domain decomposition) to extract the time-series displacement data of measurement points into natural frequencies and mode shapes that characterize the overall dynamic properties of the structure. These modal parameters are directly related to the mass and stiffness distribution of the bridge, and can reflect the health status of the structure more fundamentally and globally, providing a fundamental basis for diagnosis. By combining natural frequencies and mode shapes, a complementary diagnostic system of global localization and local qualitative analysis is formed, overcoming the misjudgment or omission that may be caused by a single indicator, and improving the accuracy and reliability of damage identification.

[0102] Step S6: Compare the natural frequencies and mode shapes with the preset reference frequency range and reference mode shapes to obtain the monitoring results, including:

[0103] Calculate the relative deviation between the current natural frequency and the reference frequency range. When the relative deviation exceeds the preset frequency deviation threshold, a frequency anomaly flag is generated. Calculate the modal guarantee criterion value between the current mode shape and the reference mode shape. When the modal guarantee criterion value is lower than the preset modal correlation threshold, a mode shape anomaly flag is generated.

[0104] When both frequency anomaly indicators and mode shape anomaly indicators exist simultaneously, the following steps are taken: Calculate the difference in mode shape curvature between the current mode shape and the reference mode shape at each characteristic point; identify continuous regions where the difference in mode shape curvature exceeds a preset curvature threshold; determine the location corresponding to the maximum value of the difference in mode shape curvature within the continuous region as the core damage region; calculate the relative deviation values ​​of several consecutive natural frequencies and assign weight coefficients based on the differences in the sensitivity of different frequencies to damage; calculate a comprehensive score based on the relative deviation values ​​of the natural frequencies and the corresponding weight coefficients; determine the damage severity level based on the correspondence between the comprehensive score and the damage severity level, where lower-order frequencies are assigned greater weight than higher-order frequencies; determine the structural risk level based on the location of the core damage region within the bridge structure; generate maintenance priorities using a predefined decision matrix based on the damage severity level and the structural risk level; and generate monitoring results characterizing the confirmation of structural damage based on the location information of the core damage region, the damage severity level, and the maintenance priorities.

[0105] Specifically, the currently identified natural frequencies are compared with the pre-established reference frequency range, and the relative deviation between the current frequency and the reference value is calculated. When the deviation of any frequency exceeds a preset threshold, a frequency anomaly flag is generated. The modal guarantee criterion is used to quantify the similarity between the current mode shape and the reference mode shape, and the correlation coefficient between the two modes is calculated. When the correlation coefficient is lower than a preset threshold, a mode shape anomaly flag is generated.

[0106] When both frequency anomaly markers and mode shape anomaly markers exist simultaneously, the mode shape curvature of the reference and current mode shapes at each characteristic point is calculated. The curvature difference between the two is calculated, and continuous regions where the curvature difference exceeds a threshold are identified. The location with the largest curvature difference within this continuous region is determined as the core damage area. The relative deviation values ​​of several consecutive natural frequencies are obtained, and weighting coefficients are assigned according to the differences in the sensitivity of different frequency orders to damage, with lower frequency orders given higher weights. A weighted comprehensive score is calculated, and the damage severity level is determined according to a preset level mapping relationship.

[0107] Specifically, the weight allocation ratio is: W1:W2:W3 = 0.5:0.3:0.2, where W1 corresponds to the first-order frequency weight, W2 corresponds to the second-order frequency weight, and W3 corresponds to the third-order frequency weight.

[0108] Specifically, the overall score is calculated using the following formula:

[0109] S=(W1·δ1+W2·δ2+W3·δ3) / (W1+W2+W3)

[0110] Wherein, δ1, δ2, and δ3 represent the relative deviation values ​​of the first, second, and third order frequencies, respectively.

[0111] Calculate the relative deviation δ of the first n natural frequencies. k (k=1,2,...,n), where:

[0112] δ k =|f {c,k} -f {b,k} | / f {b,k} ×100%

[0113] f {c,k} f is the current natural frequency of order k. {b,k} is the k-th order reference natural frequency.

[0114] Based on the location of the core damage area in the bridge structure, the structural risk level is determined. For example, the mid-span of the main beam, the top of the pier, and the cable anchorage area are predefined as high-risk areas.

[0115] The beams and bridge decks are predefined as medium-risk areas; the railings and sidewalks are predefined as low-risk areas; and the corresponding structural risk level is determined by matching the core damage area with the predefined risk area map.

[0116] Combining the damage severity level and structural risk level, maintenance priorities are generated through a predefined decision matrix. For example, when the damage severity is severe and the structural risk level is high, the highest maintenance priority is generated; when the damage severity is moderate and the structural risk level is high, or when the damage severity is severe and the structural risk level is medium, a high maintenance priority is generated; when the damage severity is minor and the structural risk level is high, or when the damage severity is moderate and the structural risk level is medium, a medium maintenance priority is generated; other combinations generate regular maintenance priorities.

[0117] Finally, monitoring results are generated, which include confirmed structural damage, detailed descriptions of core damage areas, damage severity levels, and maintenance priorities.

[0118] This invention effectively avoids misjudgments based on a single indicator caused by environmental interference or measurement noise by setting abnormality flags for both frequency and mode shape, requiring both to be triggered simultaneously before entering the damage confirmation process. This dual verification mechanism significantly improves the accuracy and robustness of damage diagnosis and reduces the system's false alarm rate. By identifying core damage areas through mode shape curvature analysis, the location of local stiffness degradation in bridge structures can be precisely pinpointed. Simultaneously, a weighted comprehensive scoring method based on multi-order frequency changes enables a quantitative assessment of damage severity. By predefining risk area maps and matching them with actual damage locations, an objective determination of structural risk levels is achieved. Furthermore, combined with damage severity, a decision matrix generates maintenance priorities, establishing a complete transformation chain from technical parameters to management decisions. This system allows limited maintenance resources to be prioritized for the most urgent and critical areas. The final monitoring results include key information such as damage confirmation, location description, severity level, and maintenance priority, forming a standardized and structured diagnostic report. This output format facilitates engineers' rapid understanding of the structural condition and enables them to take appropriate measures, greatly enhancing the practical value of the monitoring system.

[0119] When only the frequency anomaly flag exists:

[0120] Acquire the current ambient temperature data and compare and verify the current natural frequency with the ambient temperature-frequency correlation model established based on historical monitoring data;

[0121] When the frequency change conforms to the temperature change law and the mode shape characteristics remain stable, the camera is controlled to perform a preset swing action to reduce the impact of environmental interference and generate monitoring results characterizing the impact of environmental interference.

[0122] In this embodiment, when the system detects only a frequency anomaly flag, it first initiates the environmental interference judgment process. The system collects the current ambient temperature data in real time through an integrated temperature sensor, denoted as T, for example, T = 28.5℃. Simultaneously, the system calls an ambient temperature-frequency correlation model built based on historical monitoring data. This model is trained using long-term data, and its specific expression is:

[0123] F(T) = -0.0023T² + 0.15T + 10.24

[0124] Where F(T) is the predicted reference frequency at temperature T. The current measured natural frequency is compared with the model prediction value, and the relative deviation is calculated. If the frequency change does not conform to the temperature change law and the mode shape characteristics remain stable, the frequency anomaly is determined to be caused by the change in ambient temperature; if the frequency change conforms to the temperature change law and the mode shape characteristics remain stable, the frequency anomaly is determined to be caused by the influence of dust in the environment on the camera.

[0125] Specifically, dust in the environment itself does not change the bridge's true inherent frequency, but it can indirectly cause the calculated frequency value to be abnormal by contaminating the camera's imaging function. In addition, due to the high traffic volume on the bridge, the frequent passing vehicles will stir up more road dust particles, and the oily particulate matter emitted by vehicle exhaust is also very easy to adhere to the lens.

[0126] Given this environmental interference, the camera gimbal performs high-frequency, low-amplitude oscillations for dry road dust particles kicked up by vehicles. By applying acceleration, loosely attached particles overcome static friction and are detached from the lens surface. For oily particles in exhaust fumes, which have stronger adhesion, the system employs a low-frequency, high-amplitude oscillation mode to generate greater shearing force, peeling sticky contaminants off the lens.

[0127] This invention effectively removes dust, oily particles, and other contaminants from the lens surface through periodic or triggered oscillations, directly restoring image clarity and contrast. It ensures stable and accurate detection and tracking of feature points, improving the accuracy of subsequent displacement calculations and vibration analysis from the data source. This mechanism endows the system with crucial self-diagnostic and self-recovery capabilities. When only a frequency anomaly is detected while the mode shape is normal, oscillation cleaning is triggered, thus distinguishing between false anomalies caused by sensor contamination and real structural damage, reducing the system's false alarm rate and improving robustness in complex environments.

[0128] Traditional cleaning methods require frequent on-site visits by technicians, which is costly, risky, and inefficient for cameras installed on tall, remote structures such as bridges. Automatic swing cleaning enables remote, automated daily maintenance, significantly reducing the frequency of manual climbing and saving substantial manpower, time, and financial costs.

[0129] Controlling the camera to perform preset swaying movements to reduce the impact of environmental interference includes:

[0130] The swing amplitude of the preset swing action is determined based on the image quality of the time-series images;

[0131] The oscillation frequency of the preset oscillation motion is determined based on the rate of dust accumulation in the environment.

[0132] The timing of the pre-set swinging motion is determined based on the requirements of the target detection task;

[0133] Control commands for the preset swing motion are generated based on the swing amplitude, swing frequency, and swing timing of the preset swing motion.

[0134] The oscillation frequency of the preset oscillation motion is determined based on the rate of dust accumulation in the environment, including:

[0135] Obtain the concentration of dust in the environment, the ambient humidity, and the characteristic parameters of the camera lens;

[0136] The dust viscosity coefficient is determined based on the ambient humidity and the concentration of dust in the environment.

[0137] Based on the dust viscosity coefficient and the characteristic parameters of the camera lens, the type of dust adhesion is determined.

[0138] Based on the adhesion type of dust, determine the rate of dust accumulation in the environment;

[0139] The oscillation frequency of the preset oscillation motion is determined based on the rate of dust accumulation in the environment.

[0140] The characteristics of a camera lens include hydrophobicity, antistatic properties, and surface roughness.

[0141] For example, this invention determines the swing amplitude by analyzing multiple quality metrics of time-series images. First, the image sharpness index is evaluated, which is obtained by calculating the gradient variance of the image; for example, a normal baseline value of 280 and a current measurement value of 185 represent a decrease of 34%. The feature point matching success rate is detected, for example, decreasing from a normal 98% to 82%. The signal-to-noise ratio is detected, for example, decreasing from a baseline of 38 dB to 29 dB.

[0142] Based on the degree of degradation of these image quality parameters, if the sharpness index drops by more than 30% and the feature point matching success rate is below 85%, a strong cleaning mode is required, with the swing amplitude set to ±25°. If the sharpness degradation is between 15% and 30%, a standard cleaning mode is used, with an amplitude of ±15°. For minor quality degradation, the system uses a maintenance cleaning mode, with an amplitude of ±8°.

[0143] For example, the oscillation frequency is determined based on an assessment of the rate of dust accumulation in the environment. Dust concentration is monitored in real time using a particulate matter sensor; for instance, the current dust concentration is measured to be 0.25 mg / m³, while a humidity sensor displays an ambient humidity of 68%. Lens characteristics include: a hydrophobic contact angle of 105°, meeting strong hydrophobic standards; and a surface resistivity of 10 Ω·cm for antistatic properties. 8 Ω, which is a medium level of antistatic properties; the surface roughness is 0.02μm, which meets the standard of mirror-level smoothness.

[0144] In this embodiment, the baseline viscosity value of 1.2, measured under laboratory conditions using a standard dust sample, is used as the reference for calculation.

[0145] For example, the humidity effect coefficient is used to characterize the effect of ambient humidity on dust viscosity, using a humidity correction formula:

[0146] Humidity influence coefficient = 1 + (current humidity - reference humidity) × humidity sensitivity;

[0147] The reference humidity is the air humidity under standard operating conditions. The humidity sensitivity is related to the type of dust. In this implementation, the reference humidity is taken as 50% and the humidity sensitivity is 0.003, and the humidity influence coefficient is calculated to be 1.054.

[0148] The dust viscosity coefficient is the product of the base viscosity and the humidity effect coefficient.

[0149] Lens characteristic parameters are converted into standardized values ​​through parameter normalization. For example, hydrophobicity rating: a contact angle of 105° corresponds to a rating of 1.3; antistatic rating: surface resistivity of 10... 8 Ω corresponds to a score of 1.1; Surface smoothness score: Roughness of 0.02μm corresponds to a score of 1.2.

[0150] Adhesion type identification is performed based on viscosity coefficient and lens characteristic scores:

[0151] When the viscosity coefficient is greater than 2.0 and the hydrophobicity score is less than 1.0, it is judged as wet adhesion; when the viscosity coefficient is less than or equal to 2.0 and the electrostatic mimicry score is less than 1.2, it is judged as electrostatic adsorption; when the viscosity coefficient is greater than 2.0 and the electrostatic mimicry score is less than 1.2, it is judged as mixed adhesion.

[0152] For example, a baseline accumulation rate measured under standard conditions is used as a benchmark. Then, an environmental factor is calculated, obtained by querying the humidity coefficient corresponding to the current humidity level, and multiplied by an accumulation factor specific to the mixed adhesion type. Simultaneously, a lens protection factor is calculated, which is the product of the lens's hydrophobicity and antistatic properties, used to quantify the lens's surface resistance to dust accumulation. The final predicted accumulation rate is the product of the baseline accumulation rate, the environmental factor, and the lens protection factor.

[0153] In this embodiment, the accumulation coefficient of wet adhesion is 1.5, the accumulation coefficient of electrostatic adsorption is 1.2, and the accumulation coefficient of mixed adhesion is 2.0.

[0154] Based on the calculated accumulation velocity, the system determines the fundamental oscillation frequency using a preset velocity-frequency mapping table:

[0155] Low accumulation range (<0.05 mg / cm²·day): corresponding frequency 0.5 Hz

[0156] Medium accumulation range (0.05-0.15 mg / cm²·day): corresponding frequency 1.0 Hz

[0157] High accumulation range (0.15-0.30 mg / cm²·day): corresponding frequency 2.0 Hz

[0158] Severe accumulation zone (>0.30 mg / cm²·day): corresponding frequency 3.0 Hz

[0159] This invention determines the timing of the swing based on the requirements of the monitoring task. The monitoring task is divided into three priorities: high priority includes the structural anomaly verification period and the period when heavy vehicles pass through; medium priority includes the regular monitoring cycle and the data transmission period; low priority includes the system maintenance window and the low traffic period at night.

[0160] When deciding when to perform a swipe, first check if the system is currently in a high-priority task period; if so, delay the swipe operation. If there is a data transmission gap and the system idle period exceeds 5 minutes, execute the swipe immediately. Otherwise, schedule it for the next maintenance window. The system has detected a data transmission gap and no high-priority tasks are scheduled for the next 15 minutes, therefore, a swipe cleaning is triggered immediately.

[0161] This invention determines the type and rate of dust adhesion by comprehensively analyzing multi-dimensional parameters such as ambient humidity, dust concentration, and lens surface characteristics, and then precisely sets the oscillation frequency accordingly. This predictive cleaning strategy based on environmental conditions can initiate cleaning before dust accumulates to the point of affecting image quality. Compared to fixed-frequency cleaning methods, it avoids equipment wear caused by excessive oscillation, ensures continuous lens cleaning, and significantly enhances the system's adaptability and robustness in harsh environments such as dusty and humid conditions.

[0162] This invention determines the optimal timing for oscillation by considering the requirements of the target detection task, thus avoiding critical monitoring periods and performing cleaning actions only during data processing gaps or when the system is idle. This ensures that cleaning operations do not affect normal monitoring functions, and the decentralized operation avoids the instantaneous occupation of system resources by centralized cleaning, reducing the impact of cleaning operations on monitoring tasks and improving the robustness of the system.

[0163] If no frequency anomaly or mode shape anomaly is detected within the preset number of tests:

[0164] Confirm that the target bridge is in a healthy state and generate monitoring results to characterize the health confirmation of the target bridge;

[0165] Based on the monitoring data within a preset number of tests, the reference frequency range is updated and the reference mode shape is optimized;

[0166] Based on the duration of the target bridge being in a healthy state, a timer trigger frequency adjustment instruction for the microcontroller is generated.

[0167] Specifically, if no frequency anomaly indicators or mode anomaly indicators are found within multiple consecutive monitoring cycles (e.g., 10 consecutive tests), the target bridge is confirmed to be in a structurally healthy state. A standardized health confirmation monitoring report is generated, which mainly includes:

[0168] Monitoring period and duration;

[0169] Monitoring findings, for example, indicate that the structure is in good condition and no obvious damage was observed.

[0170] Stability assessment of key parameters.

[0171] Extract all valid monitoring data within a preset number of detections (e.g., the most recent 10 times), including natural frequency values ​​and mode shape data of each order. Based on the new frequency dataset, recalculate the reference range of each frequency order using statistical methods.

[0172] The new benchmark range = [mean - 3 × standard deviation, mean + 3 × standard deviation].

[0173] Historical health data is retained, and the baseline database is updated using a sliding window approach, dynamically adjusting the monitoring frequency based on the duration of the health status. For example:

[0174] When the duration of health is less than 1 month, maintain the standard triggering frequency, which is 3 times per hour;

[0175] When the duration of health is within 1-3 months, reduce to a moderate frequency, i.e., once every 12 hours;

[0176] When the duration of good health is greater than 3 months, the monitoring frequency should be reduced to once a day.

[0177] If any abnormal flag is detected, immediately revert to the standard trigger frequency.

[0178] This invention confirms a healthy state only when no abnormalities are detected within a preset series of consecutive detection cycles. This effectively eliminates the impact of fluctuations caused by accidental environmental interference or measurement noise. This statistical process control-based approach significantly improves the reliability and accuracy of state determination.

[0179] This invention dynamically updates the reference frequency range and optimizes the reference mode shape using new monitoring data under healthy conditions, enabling the system's diagnostic reference to evolve synchronously with the actual state of the structure. This enhances the system's applicability and diagnostic accuracy during long-term service and effectively prevents false alarms caused by outdated references.

[0180] This invention significantly reduces unnecessary monitoring frequency by intelligently adjusting the timer trigger frequency of a microcontroller based on the duration of a health status. This reduces the consumption of data storage and computing resources, providing technical support for long-term field monitoring scenarios that rely on limited energy sources such as solar power, and improving the economic efficiency and long-term sustainability of the monitoring system.

[0181] Example 2

[0182] like Figure 2 As shown, this invention also proposes a computer vision-based long-term bridge structure vibration monitoring system to implement any of the computer vision-based long-term bridge structure vibration monitoring algorithms described in Example 1, including:

[0183] The image generation unit is used to control the camera to capture a timing image of the target bridge in response to the trigger command generated by the timer of the microcontroller. The timing image includes a reference image and multiple target images.

[0184] The feature matching unit is used to extract multiple feature points from the reference image and match each feature point in the reference image with each target image to obtain a preliminary matching point pair.

[0185] The matching filtering unit is used to generate a confidence score for each matching point pair and remove matching point pairs with confidence scores lower than a first threshold to obtain the target matching point pair set.

[0186] The displacement field reconstruction unit is used to calculate the displacement of the target bridge at multiple feature points based on the target matching point pair set, so as to generate the vibration time series data of the target bridge.

[0187] The modal parameter identification unit is used to extract at least one natural frequency and corresponding mode shape of the target bridge based on vibration time series data. The mode shape is used to characterize the spatial deformation of the target bridge when it vibrates at the corresponding natural frequency.

[0188] The structural health diagnostic unit is used to compare the natural frequencies and mode shapes with preset reference frequency ranges and reference mode shapes to obtain monitoring results.

[0189] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A long-term vibration monitoring algorithm for bridge structures based on computer vision, characterized in that, include: S1, in response to the trigger command generated by the timer of the microcontroller, controls the camera to capture a timing image of the target bridge, the timing image including a reference image and multiple target images; S2, extract multiple feature points from the reference image, and match each feature point in the reference image with each of the target images to obtain a preliminary matching point pair; S3, generate the confidence score for each matching point pair, and remove matching point pairs with confidence scores lower than the first threshold to obtain the target matching point pair set; S4. Based on the target matching point pair set, calculate the displacement of the target bridge at multiple feature points to generate vibration time series data of the target bridge. S5. Based on the vibration time series data, extract at least one natural frequency and corresponding mode shape of the target bridge. The mode shape is used to characterize the spatial deformation of the target bridge when it vibrates at the corresponding natural frequency. S6. The natural frequency and the mode shape are compared with the preset reference frequency range and reference mode shape to obtain the monitoring results; The step of comparing the natural frequency and the mode shape with a preset reference frequency range and reference mode shape to obtain monitoring results includes: Calculate the relative deviation between the current natural frequency and the reference frequency range. When the relative deviation exceeds a preset frequency deviation threshold, generate a frequency anomaly flag. Calculate the modal guarantee criterion value between the current mode shape and the reference mode shape. When the modal guarantee criterion value is lower than a preset modal correlation threshold, generate a mode shape anomaly flag. When both the frequency anomaly indicator and the mode shape anomaly indicator are present, the mode shape curvature difference between the current mode shape and the reference mode shape at each feature point is calculated; continuous regions where the mode shape curvature difference exceeds a preset curvature threshold are identified; the location corresponding to the maximum value of the mode shape curvature difference in the continuous region is determined as the core damage region; the relative deviation values ​​of several consecutive natural frequencies are calculated, and weight coefficients are assigned according to the differences in the sensitivity of different frequencies to damage; a comprehensive score is calculated based on the relative deviation values ​​of the natural frequencies and the corresponding weight coefficients; the damage severity level is determined based on the correspondence between the comprehensive score and the damage severity level; wherein, the weight assigned to lower-order frequencies is greater than the weight assigned to higher-order frequencies; the structural risk level is determined based on the location of the core damage region in the bridge structure; maintenance priorities are generated through a predefined decision matrix based on the damage severity level and the structural risk level; and monitoring results characterizing the confirmation of structural damage are generated based on the location information of the core damage region, the damage severity level, and the maintenance priority. When only the frequency anomaly flag exists: Acquire the current ambient temperature data and compare and verify the current natural frequency with the ambient temperature-frequency correlation model established based on historical monitoring data; When the frequency change conforms to the temperature change law and the mode shape characteristics remain stable, the camera is controlled to perform a preset swing action to reduce the impact of environmental interference and generate monitoring results characterizing the impact of environmental interference.

2. The computer vision-based long-term bridge structure vibration monitoring algorithm according to claim 1, characterized in that, The step of calculating the displacement of the target bridge at multiple feature points based on the target matching point pair set to generate vibration time-series data of the target bridge includes: S41, Based on the target matching point pair set, calculate the pixel coordinate displacement vector between the feature points in the reference image and the corresponding matching points in the target image; S42, based on the gray-level gradient changes of the image sequence in the spatiotemporal dimension, and / or by fitting the gray-level distribution of the feature point neighborhood with a curved surface, the pixel coordinate displacement vector is optimized with sub-pixel precision to obtain the image plane displacement amount with sub-pixel precision. S43, based on pre-calibrated camera parameters, the image plane displacement is transformed into the physical displacement of the target bridge in three-dimensional space through coordinate transformation relationship. The camera parameters include camera internal parameters and camera external parameters relative to the target bridge. S44, based on the correspondence between the timestamp corresponding to the trigger command of the microcontroller and each feature point, the physical displacement is combined into a displacement-time sequence; S45, aggregate the displacement-time series of all feature points to obtain vibration time series data to characterize the vibration response of the target bridge at multiple discrete locations.

3. The computer vision-based long-term monitoring algorithm for bridge structure vibration according to claim 1, characterized in that, The control of the camera to perform a preset swaying motion to reduce the impact of environmental interference includes: The swing amplitude of the preset swinging action is determined based on the image quality of the time-series images; The oscillation frequency of the preset oscillation action is determined based on the rate of dust accumulation in the environment. The timing of the preset swinging motion is determined based on the requirements of the target detection task; Control commands for the preset swinging motion are generated based on the swing amplitude, swing frequency, and swing timing of the preset swinging motion.

4. The computer vision-based long-term bridge structure vibration monitoring algorithm according to claim 3, characterized in that, The determination of the oscillation frequency of the preset oscillation motion based on the accumulation rate of dust in the environment includes: The concentration of dust and humidity in the environment, as well as the characteristic parameters of the camera lens, are obtained. The dust viscosity coefficient is determined based on the ambient humidity and the concentration of dust in the environment. Based on the dust viscosity coefficient and the characteristic parameters of the camera lens, the dust adhesion type is determined; Based on the adhesion type of the dust, the accumulation rate of dust in the environment is determined; The oscillation frequency of the preset oscillation action is determined based on the rate of dust accumulation in the environment.

5. The computer vision-based long-term bridge structure vibration monitoring algorithm according to claim 4, characterized in that, The lens characteristics of the camera include hydrophobicity, antistatic properties, and surface roughness.

6. The computer vision-based long-term vibration monitoring algorithm for bridge structures according to claim 1, characterized in that, If no frequency anomaly or mode shape anomaly is detected within the preset number of tests: Confirm that the target bridge is in a healthy state and generate monitoring results to characterize the health confirmation of the target bridge; Based on the monitoring data within the preset number of detections, the reference frequency range is updated, and the reference mode shape is optimized; Based on the duration of the target bridge being in a healthy state, a timer trigger frequency adjustment instruction for the microcontroller is generated.

7. The computer vision-based long-term monitoring algorithm for bridge structure vibration according to claim 1, characterized in that, The step of matching each feature point in the reference image with each of the target images to obtain a preliminary matching point pair includes: S21, extract image features from the reference image and the target image, wherein the image features include feature point positions and corresponding high-dimensional feature descriptors; S22, based on the high-dimensional feature descriptors of feature points in the reference image and the target image, by calculating the similarity matrix between the feature descriptors, forward matching is performed from the reference image to the target image to find candidate matching points in the target image for each feature point in the reference image; backward matching is performed from the target image to the reference image to find candidate matching points in the reference image for each feature point in the target image. S23, perform a consistency check on the results of forward matching and backward matching, retain the corresponding matching point pairs in forward matching and backward matching, and obtain the preliminary matching point pairs.

8. A computer vision-based long-term vibration monitoring system for bridge structures, used to implement the computer vision-based long-term vibration monitoring algorithm for bridge structures as described in any one of claims 1 to 7, characterized in that, include: An image generation unit is used to control a camera to capture a time-series image of the target bridge in response to a trigger command generated by a timer of a microcontroller. The time-series image includes a reference image and multiple target images. A feature matching unit, connected to the image generation unit, is used to extract multiple feature points in the reference image and match each feature point in the reference image with each of the target images to obtain a preliminary matching point pair; A matching filtering unit, connected to the feature matching unit, is used to generate a confidence score for each matching point pair and remove matching point pairs with confidence scores lower than a first threshold to obtain a target matching point pair set. The displacement field reconstruction unit, connected to the matching filtering unit, is used to calculate the displacement of the target bridge at multiple feature points based on the target matching point pair set, so as to generate the vibration time series data of the target bridge. The modal parameter identification unit, connected to the displacement field reconstruction unit, is used to extract at least one natural frequency and corresponding mode shape of the target bridge based on the vibration time series data. The mode shape is used to characterize the spatial deformation morphology of the target bridge when it vibrates at the corresponding natural frequency. The structural health diagnostic unit, connected to the modal parameter identification unit, is used to compare the natural frequency and the mode shape with a preset reference frequency range and reference mode shape to obtain monitoring results.

Images