A vibration displacement measurement method based on key point detection network

Abstract

This invention provides a vibration displacement measurement method based on a key point detection network, relating to the field of structural health monitoring technology. The method includes: acquiring vibration image data of a structure, wherein key points are marked in the vibration images; detecting key points in the vibration image data using a key point detection network to obtain key points; extracting displacement and physically mapping the key points to obtain vibration displacement data; wherein the backbone network of the key point detection network includes functional modules for fitting non-rigid deformations and motion blur, and suppressing illumination interference. Through this approach, the present invention improves the continuity and reliability of vibration displacement measurement in structural health monitoring.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring technology, and more specifically, to a vibration displacement measurement method based on a key point detection network. Background Technology

[0002] Structural vibration monitoring is a core technology for ensuring the safe operation of flexible engineering structures such as long-span bridges and lightweight buildings. Traditional contact measurements (such as accelerometers) can easily alter the dynamic characteristics of lightweight structures due to the added mass effect, and the wiring is complex. Non-contact measurements (such as laser vibrometers and microwave radar) can solve the problem of mass interference, but the equipment is expensive and it is difficult to achieve synchronous monitoring at multiple points across the entire area.

[0003] Computer vision technology, with its advantages of non-contact, long-distance, multi-point synchronous measurement, and low cost, has shown great potential in the field of structural vibration monitoring. Existing visual measurement methods are mainly divided into traditional image processing and deep learning-based methods. Among traditional methods, digital image correlation (DIC) calculates the displacement field by tracking high-contrast speckle patterns or template regions, but it has poor robustness to changes in illumination or line-of-sight occlusion. Optical flow methods, such as the KLT (Kanade-Lucas-Tomasi) algorithm, utilize corner features for sparse tracking. While it achieves sub-pixel accuracy for small displacements, it strictly relies on the assumptions of "constant brightness" and "small motion," making it difficult to measure large vibrations. Edge detection methods, such as the Canny and Sobel operators, achieve measurement by tracking the structural contour, but their integer-pixel-level coordinate extraction accuracy is insufficient for monitoring weak vibrations.

[0004] In recent years, the development of deep learning has provided new approaches to solving the challenges of vibration measurement in complex environments. The combination of RAFT optical flow and traditional KLT optical flow has effectively improved the accuracy of large displacement measurements. Efficient segmentation and measurement of multi-target rotating bodies have been achieved through semantic and instance segmentation networks. By nesting the backbone of the YOLO series network with MLP blocks using residuals, the speed and accuracy of vibration measurement have been balanced. However, target detection-based methods still face significant challenges when applied to high-precision measurements of flexible structures. For example, the elastic deformation of the target can cause unstructured jitter in the bounding box (BBox), introducing systematic errors that are difficult to eliminate. To address this issue, a method for measuring long-span bridges using an "anthropomorphic" approach is proposed. This method analogizes the bridge structure to a jointed torso and utilizes an improved YOLOv7 algorithm to achieve accurate tracking of key bridge points, demonstrating its superiority over traditional BBox tracking. Nevertheless, existing research largely focuses on ideal lighting conditions. In actual long-term monitoring, structures often face degraded imaging conditions such as sudden changes in lighting, strong exposure, line-of-sight obstruction, and motion blur caused by high-speed vibration. These factors can lead to severe degradation of image features, causing key point detection networks to fail, and consequently affecting the continuity and reliability of vibration displacement measurements in structural health. Summary of the Invention

[0005] In view of this, the present invention proposes a vibration displacement measurement method based on a key point detection network to solve the problems existing in the prior art.

[0006] To achieve the above objectives, this invention proposes a vibration displacement measurement method based on a key point detection network, comprising: Obtain vibration image data of the structure, wherein key points are marked in the vibration images; The vibration image data is subjected to key point detection by a key point detection network to obtain key points. The displacement of the key points is then extracted and physically mapped to obtain vibration displacement data. The backbone network of the key point detection network includes functional modules for fitting non-rigid deformations and motion blur, and suppressing illumination interference.

[0007] Optionally, the key point detection network is provided with a backbone network, a neck network and a detection head in sequence; The deep semantic extraction stage of the backbone network includes a feature enhancement module; the feature enhancement module includes a fast spatial pyramid pooling module, a geometric adaptive focusing module, and an illumination adaptive lightweight module connected in sequence. The geometric adaptive focusing module is used to fit non-rigid deformation and motion blur, and the illumination adaptive lightweight module is used to suppress illumination interference. The neck network adopts a path aggregation network architecture, and the detection head adopts a decoupled head structure; the output features of the illumination adaptive lightweight module are processed by the C2PSA module and used as the deep feature output of the backbone network.

[0008] Optionally, in the geometric adaptive focusing module, for the input feature map of the geometric adaptive focusing module, the spatial offset is obtained by performing offset prediction on the input feature map through a convolutional layer, and the input feature map and the spatial offset are convolved by deformable convolution. After convolution processing, the output features of the geometric adaptive focusing module are obtained by normalization processing and activation function activation.

[0009] Optionally, in the illumination-adaptive lightweight module, for the input features, convolution is used to process them to obtain an intensity map. The intensity values ​​in the intensity map are divided into different statistical intervals. The original features of the statistical intervals are extracted, and the original features are aggregated, enhanced, and mapped to obtain prototype features. A query vector is generated based on the original features. A key vector and a value vector are generated based on the prototype features. The query vector, key vector, and value vector are processed by a multi-head self-attention module to obtain an attention output result. The attention output result is processed by convolutional branches of different scales to obtain different scale branch features. The different scale branch features are concatenated, and a gating network is used to generate corresponding branch weights based on the branch feature concatenation result. The different scale branch features are fused according to the branch weights, and the fused features are added to the attention output result by residual addition. The feature is then activated by an activation function to obtain the output feature of the illumination-adaptive lightweight module.

[0010] Optionally, the loss function of the key point detection network is direct coordinate regression loss.

[0011] Optionally, the process of displacement extraction and physical mapping of the key points includes: The initial static frame in the vibration image data is selected as the global reference benchmark, and the pixel displacement data of the key point at any time is extracted. Based on the known physical radius of the circular marker on the structure and the average pixel radius in the initial static frame, the mapping relationship between pixel coordinates and physical space coordinates is calculated. The pixel displacement data is then transformed into physical space according to the mapping relationship to obtain the vibration displacement data of the key point, i.e., the vibration displacement data of the structure.

[0012] Optionally, before performing keypoint detection on the vibration image data, the following steps are included: The keypoint detection network is optimized using a sample dataset. The sample dataset contains images of circular markers. The initial center and radius of the circular markers are obtained from the original frames in the sample dataset using Hough circle transform. A region of interest is constructed based on the initial center. The edge point set of the sub-pixel level edge contour is extracted from the region of interest. The edge point set is then subjected to regression analysis using a least-squares ellipse fitting algorithm to obtain the corresponding geometric center coordinates as the true label of the sample dataset.

[0013] Optionally, before performing keypoint detection on the vibration image data, the following steps are included: The keypoint detection network was validated using root mean square error, mean relative error, and Pearson correlation coefficient as indicators.

[0014] On the other hand, the present invention provides a vibration displacement measurement system based on a key point detection network for performing the above-described method.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: To address the challenge of accurately locating key points under harsh conditions such as low light, high dynamic range, and motion blur, this invention employs YOLO-Vib by introducing GAFM and IALB to effectively enhance the model's feature extraction capabilities and sub-pixel-level key point localization accuracy under complex and degraded imaging conditions. Specifically, GAFM is first used to dynamically learn spatial sampling offsets, enabling the network's convolutional kernel receptive fields to adaptively fit the geometric deformation of the target and focus on deformed or blurred target marker regions. This effectively overcomes the problem of limited receptive fields in standard convolutions under motion blur and non-rigid deformation, and significantly suppresses edge blur interference caused by high-speed vibrations. To further improve feature stability under extreme lighting conditions, IALB is introduced. This module combines Dynamic Intensity-Aware Attention (DIA) and Illumination Gated Fusion (IGF) mechanisms. By modeling the global intensity distribution, global contextual dependencies are efficiently captured, effectively suppressing feature distribution degradation caused by local highlights, shadows, or low light, ensuring the model's feature discrimination and stability under complex lighting conditions. Furthermore, eliminating the "quantization step" error in keypoint localization is crucial for sub-pixel-level high-precision measurement of minute vibration displacements. To address this issue, a Direct Coordinate Regression Loss (DCRL) strategy is introduced, which optimizes keypoint location by minimizing the geometric Euclidean distance between the predicted keypoint coordinates and the physical ground truth. This effectively overcomes the shortcomings of traditional object detection networks (based on IoU or OKS) that are sensitive to small target scales and have weak gradient responses in the sub-pixel range, significantly improving the model's numerical accuracy and convergence stability in continuous displacement tracking. These contributions enable YOLO-Vib to achieve high detection accuracy even under harsh conditions such as low lighting, high dynamic range, and motion blur, further laying the foundation for the development and application of YOLO-Vib in visual vibration measurement. Attached Figure Description

[0016] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is an overall experimental flowchart in an embodiment of the present invention; Figure 2 This is a schematic diagram of the vibration data acquisition system in an embodiment of the present invention; Figure 3 This is a schematic diagram of the YOLO-Vib model in an embodiment of the present invention; Figure 4 This is a structural diagram of the geometry-adaptive focusing module and the illumination-adaptive lightweight module in an embodiment of the present invention; Figure 5 This is a comparison chart of the results of different methods on dataset A in this embodiment of the invention; Figure 6 These are schematic diagrams of scene images of dataset B after different degrees of degradation processing in embodiments of the present invention; Figure 7 This is a comparison chart of displacement curve tracking of different methods in this embodiment of the invention on dataset B; Figure 8 This is a comparison chart of displacement curves of different methods in the embodiments of the present invention on the outdoor bridge dataset C. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] Non-contact measurement based on computer vision has become an important development direction in the field of structural health monitoring (SHM) due to its absence of added mass effects. However, degraded imaging conditions commonly found in real engineering scenarios (such as high-speed motion blur, low illumination, and shadow interference) severely limit the robustness of the methods, and existing methods struggle to maintain sub-pixel-level measurement accuracy under dynamic large deformations. To address this, this invention proposes a highly robust visual vibration measurement framework, YOLO-Vib, based on key point detection.

[0019] To address the scarcity of high-quality training data, this scheme constructs a physical truth generation framework based on geometric constraints, resolving the pain point of data-driven methods lacking high-precision truth values ​​in sub-pixel-level vibration measurements. To overcome the feature extraction bottleneck under complex working conditions, the scheme designs a dual enhancement mechanism: a geometrically adaptive focusing module (GAFM) is introduced, which effectively suppresses non-rigid deformation and edge blurring caused by high-speed vibration by dynamically learning convolutional sampling offsets; an illumination-adaptive lightweight module (IALB) is introduced, which uses global intensity distribution modeling to ensure the discriminative stability of features under extreme illumination changes. Furthermore, the direct coordinate regression loss function (DCRL) is used to optimize geometric positioning errors, significantly improving the model's physical measurement accuracy.

[0020] This invention relates to a visual vibration measurement technology for structural health monitoring. By integrating the feature self-learning capabilities of deep neural networks with visual measurement techniques, this invention achieves high-precision vibration measurement under complex operating conditions.

[0021] To achieve the above technical objectives, such as Figure 1 As shown, the present invention provides the following technical solution: S1: Build a high-speed vibration data acquisition system to obtain a continuous vibration image sequence of the structural object.

[0022] S2: A semi-automatic annotation strategy based on geometric constraints is adopted to construct a sub-pixel-level high-precision dataset.

[0023] S3: In the detection network, a geometric adaptive focusing module is designed for adaptive fitting of non-rigid deformation and motion blur.

[0024] S4: In the detection network, an illumination-adaptive lightweight module is designed to suppress lighting interference in complex environments.

[0025] S5: For the detection network, a direct coordinate regression loss strategy is designed to achieve sub-pixel-level high-precision positioning optimization.

[0026] S6: The detection network is validated on vibration image data of large indoor and outdoor bridge structures, and vibration displacement is measured on the acquired vibration image sequence through the detection network.

[0027] like Figure 2 As shown, the specific steps for setting up a high-speed vibration data acquisition system to obtain a vibration image sequence in step S1 are as follows: Circular markers with known physical radii are affixed to the measurement nodes of the flexible structure under test. A high-speed industrial camera is set up at a certain distance from the target structure, ensuring that the camera's optical axis is strictly perpendicular to the target vibration plane to reduce perspective distortion. The acquisition system is activated, continuously capturing vibration videos of the structure under excited conditions at a set frame rate and resolution, generating an original vibration image sequence containing consecutive frames.

[0028] The S2 section describes the construction of a semi-automated dataset based on geometric constraints. Addressing the low accuracy of traditional manual annotation, a coarse-to-fine geometric constraint annotation method is proposed: First, the Hough circle transform is used to obtain the initial center and radius of the circular markers in the original frame. Then, a region of interest is constructed using the initial center, and the Canny operator and a dual-threshold strategy are used to extract sub-pixel-level edge contours. Subsequently, a least-squares ellipse fitting algorithm is used to perform regression analysis on the edge point set, solving for the coordinates of the ellipse's geometric center, unaffected by quantization errors, as the true label, and calculating the equivalent pixel radius. Finally, the high-precision coordinates are converted into a standard dataset format for network training.

[0029] The S3 design for the geometrically adaptive focusing module is as follows: To overcome the limitations of the fixed receptive field of standard convolutional layers in dealing with high-speed vibrating targets, a GAFM module is designed and embedded in the deep semantic features of the backbone network. This module first predicts the spatial offset field of the input feature map through a lightweight convolutional layer. For each standard convolutional sampling point on the feature map, the network assigns it a learnable two-dimensional spatial offset. This mechanism allows the receptive field of the convolutional kernel to dynamically adjust according to the image content, adaptively fitting the non-rigid distortion and edge blurring caused by high-speed vibration of the target, effectively focusing on effective pixels containing high-frequency edge information.

[0030] The design concept and specific content of the S4 lightweight illumination adaptive module are as follows: To address the interference of complex lighting conditions (such as uneven lighting and metal reflection) on feature extraction in industrial environments, a lightweight illumination adaptive module incorporating a dynamic intensity-aware attention mechanism and an illumination-gated fusion network is designed. First, the dynamic intensity-aware attention mechanism effectively suppresses local highlight or shadow interference by calculating the attention between local pixels and global illumination statistical prototype features. Then, the gated fusion network uses bi-branch convolutions to capture spatial details at different scales and generates dynamic weights through the gated network, achieving dynamic weighted fusion of multi-scale local details and global illumination features, thereby enhancing the discriminative power and stability of the features.

[0031] The idea and specific content of the S5 design for the direct coordinate regression loss strategy are as follows: In vibration measurement tasks for flexible structures, the traditional OKS loss function based on target bounding box area normalization is insensitive to small displacements and easily leads to gradient nonlinearity. Therefore, this invention employs direct coordinate regression loss, directly penalizing the geometric Euclidean distance between the predicted keypoint center coordinates and the true sub-pixel coordinates generated in step S2. This strategy eliminates the "quantization step" error based on detection box optimization, providing the network with a more direct geometric optimization objective and significantly improving the model's numerical accuracy and convergence stability in continuous displacement tracking.

[0032] The calculation principle of the measurement evaluation index involved in S6 is as follows: To comprehensively and quantitatively evaluate the accuracy and stability of the algorithm proposed in this invention in the displacement measurement of flexible structures, three indicators are used for performance measurement: root mean square error (RMSE), mean relative error (MRE), and Pearson correlation coefficient (PCC). RMSE is used to characterize the overall magnitude of the absolute deviation between the network-predicted displacement value and the sensor's actual displacement value, which can intuitively reflect the algorithm's positioning accuracy in physical space. MRE is used to characterize the average proportion of the prediction error to the actual value, mainly used to evaluate the measurement stability of the algorithm when dealing with different vibration amplitudes (different scales). PCC is used to characterize the strength and direction of the linear correlation between the predicted displacement time series and the actual displacement time series, used to measure the consistency of the trend of the algorithm during continuous dynamic tracking.

[0033]

[0034]

[0035]

[0036] in, N The total number of frames in the test sequence. For the first The structural physical displacement values ​​predicted by the model of this invention at frame time; For the first The frame contains the physical displacement value measured by a high-precision displacement sensor; This represents the average value of the network-predicted displacement sequence over the test period. This represents the average value of the sensor's actual displacement sequence within the same time period.

[0037] Regarding the overall structure of the detection network model described above: 1. The YOLO-Vib network uses CSP-Darknet as its backbone, employs PANet for multi-scale feature pyramid fusion, and uses a decoupled head for keypoint prediction. Its core innovation lies in the concatenation of a geometrically adaptive focusing module (GAFM) and an illumination-adaptive lightweight module (IALB) at the deep semantic feature level of the backbone network, and the reconstruction of the prediction head loss function into a direct coordinate regression loss (DCRL). The overall architecture of the model can be found in [link to model architecture]. Figure 3 .

[0038] 2. YOLO-Vib Keypoint Detection Infrastructure The keypoint detection network architecture mainly consists of three core parts: the backbone network, the neck network, and the keypoint prediction head.

[0039] The backbone network uses the CSP-Darknet architecture as the basic feature extractor, which consists of several cascaded convolutional layers and C3k2 residual blocks. It aims to transform the input raw vibration image into a multi-scale feature map containing different semantic levels.

[0040] The feature enhancement module is deployed in the deep semantic extraction stage of the backbone network, specifically including the fast spatial pyramid pooling (SPPF) module connected in sequence, as well as the geometric adaptive focusing module (GAFM) and the illumination adaptive lightweight module (IALB) that are the core of this invention.

[0041] The neck network adopts the Path Aggregation Network (PANet) architecture, which includes an upsampling layer, feature splicing nodes, and a C3k2 module for efficient feature aggregation, responsible for achieving cross-scale fusion of deep semantic features and low-level geometric features.

[0042] The keypoint prediction head adopts a decoupled head structure, which includes a coordinate regression branch and a visibility classification branch, and has been structurally optimized specifically for sub-pixel level point regression tasks.

[0043] The original vibration image is processed by the backbone network through successive convolutional layers and C3k2 residual blocks to extract features. At the deep feature level, the GAFM module learns the spatial offset, adaptively capturing the geometric distortion features caused by high-speed vibration. The intensity-sensing mechanism of the IALB module eliminates interference from lighting fluctuations in complex environments, thus obtaining highly robust feature representations. The neck network aggregates the enhanced deep features with shallow features that retain more positional information through a bidirectional path, and utilizes the C3k2 module to improve the model's spatial perception ability of small marker points. Finally, the keypoint prediction head calculates based on the fused feature map, directly predicting the sub-pixel-level center coordinates of circular markers through a decoupled regression branch, and using a classification branch to determine their visibility status, achieving high-precision localization of key points in the flexible structure.

[0044] The deep features output by the C2PSA module from the IALB module are input into the upsampling layer of the neck network. The neck network structure consists of the following layers in sequence: upsampling layer, first concatenation layer, first C3k2 module, upsampling layer, second concatenation layer, second C3k2 module, convolutional layer, third concatenation layer, third C3k2 module, convolutional layer, fourth concatenation layer, and fourth C3k2 module. The specific feature fusion connection relationship is as follows: the first splicing layer splices the output of the first upsampling layer with the features output by the third C3k2 module in the backbone network; the second splicing layer splices the output of the second upsampling layer with the features output by the second C3k2 module in the backbone network; the third splicing layer splices the output features of the first C3k2 module with the output of its preceding convolutional layer; the fourth splicing layer splices the output of the C2PSA module at the end of the backbone network with the output of its preceding convolutional layer; the output ends of the second, third, and fourth C3k2 modules are also connected to key point detection heads.

[0045] 3. Geometric Adaptive Focusing Module Structure Standard convolution operations rely on a fixed grid sampling pattern (such as 3x3), which makes it difficult to effectively capture the dynamic deformation features of the target structure. Therefore, a geometrically adaptive focusing module is introduced. By learning the spatial offset of the sampling points, the receptive field of the convolution kernel can adaptively fit the geometric deformation of the target. For example... Figure 4 As shown in (a), this module first predicts the offset field through a lightweight convolutional layer. This module first predicts the offset field through a lightweight convolutional layer (OffsetConv). For each location on the input feature map... Sampling points of standard convolution It is given a learnable 2D offset. The offset field, along with the original feature map, is input to the deformable convolutional layer DeformConv, whose output feature map... for:

[0046] in, Indicates the number of sampling points. For n sampling points A fixed offset relative to the center position, Indicates the corresponding weight. Represents the corresponding feature map; The final output is then processed by normalization and the SiLU activation function. This mechanism gives the network the ability to "dynamically sample," enabling it to focus on effective pixels that contain high-frequency edge information.

[0047] 4. Light-adaptive lightweight modular structure To address the interference of complex lighting conditions (such as uneven lighting and shadow occlusion) on feature extraction in industrial environments, this invention, inspired by Histoformer, designs a lightweight lighting adaptive module. This module consists of two sequential sub-modules: Dynamic Intensity-Aware Attention (DIA) and Illumination Gated Fusion (IGF), aiming to efficiently capture global contextual dependencies.

[0048] (1) Dynamic intensity perception attention DIA avoids the high cost of calculating pairwise correlations between pixels across the entire image, instead calculating the attention between pixels and a fixed set of prototype features, reducing computational complexity to a much lower level. For example... Figure 4 As shown in (b) above, firstly using Convolution from input features Generate intensity map The intensity values ​​are then divided into L statistical intervals, with L set to 16 in this embodiment. Corresponding features are extracted via convolution to obtain the original features. These original features are then aggregated and enhanced using an MLP (Multi-Level Processing) layer, and projected through a linear layer to generate global prototype features. Subsequently, a query vector Q is generated using the original features, and a key vector K and a value vector V are generated using the prototype features. These are then input into the Multi-Head Self-Attention module.

[0049] This mechanism allows each local pixel to query global illumination statistics through an attention mechanism, thereby effectively suppressing interference from local highlights or shadows.

[0050] (2) Illumination Gated Fusion The output of DIA is integrated at multiple scales via the IGF module. For example... Figure 4 As shown in (c), this module utilizes two parallel branches (with convolutional kernel sizes of [sizes not specified]). and Captures spatial details at different scales. Each branch includes batch normalization and SiLU activation. Two feature streams. and After concatenation, dynamic weights λ∈[0,1] are generated through a gating network. The final feature fusion formula is:

[0051] in Activated for Sigmoid. This involves element-wise multiplication. Finally, a residual connection is introduced to add the input features to the fused features, and the output is activated by ReLU. This part achieves dynamic weighting of local details and global illumination features, enhancing the stability of the features.

[0052] 5. Loss Function In vibration measurement tasks for flexible structures, the accuracy of key point detection directly affects the accuracy of displacement calculation. Traditional target detection uses the OKS loss function, which relies on normalization of the detection box area. This indirect optimization method is prone to gradient nonlinearity and insensitivity to small displacements. Therefore, this study employs Direct Coordinate Regression Loss to directly penalize the geometric deviation between predicted and true coordinates.

[0053]

[0054] 6. Displacement Extraction and Physical Quantity Calculation Module To obtain the physical vibration response (unit: mm) with engineering significance, it is necessary to establish an image pixel coordinate system. With physical space coordinate system The precise mapping relationship between them. First, based on the global average pixel radius. (Unit: pixel) and the known physical radius of the circular label. (Unit: mm), Define pixel-physical calibration coefficient k (Unit: mm / pixel) is: .

[0055] Based on this, this scheme adopts an absolute displacement measurement strategy based on a fixed reference frame to eliminate the cumulative drift of traditional optical flow methods in long-term tracking. A static frame is set before the vibration excitation is applied (…). () serves as the global reference baseline. For any given time... t Key point pixel coordinates Its relative to the initial position pixel displacement , The calculation is as follows: ,

[0056] The physical displacement response at this moment, in conjunction with the calibration coefficient k. Represented as: ,

[0057] This method avoids the cumulative error introduced by inter-frame recursive calculation by always anchoring to the initial frame, thus ensuring the stability and accuracy of long-term vibration measurements.

[0058] The above technical solution will be explained in conjunction with the accompanying drawings and data: Dataset A contains a high-resolution continuous vibration image sequence with a resolution of 960×1000 and a frame rate of 100FPS. On this data, this invention conducted comparative experiments with traditional optical flow methods (such as Lucas-Kanade and Farneback), feature matching algorithms (SIFT), and various deep learning-based keypoint detection networks. The results are shown in Table 1. The method of this invention surpasses traditional visual vibration measurement methods and existing deep learning baseline models in terms of the measurement accuracy of minute vibration displacements. Figure 5 As shown, the displacement curve extracted by the method of this invention closely matches the actual physical reference height of the high-precision laser sensor, especially with minimal dynamic tracking error at the peaks and troughs. The numerical results in Table 1 also fully confirm this; Table 1 shows the measurement results of YOLO-Vib and other vibration measurement methods on dataset A.

[0059] Table 1 ; To further evaluate the stability of the method of this invention when faced with poor image visual quality due to environmental interference or physical limitations of the equipment, this invention performed image processing on the original acquired benchmark data to generate a vibration image sequence (denoted as dataset B) containing various typical degradation scenarios, such as... Figure 6 As shown in the figure, a linear motion blur model is used to degrade the image to simulate severe motion blur caused by high-frequency vibration or insufficient camera exposure time; a global linear attenuation and contrast scaling method is used to simulate luminance disturbance, and a low-light and high-contrast (local overexposure) video environment is generated by adjusting parameters; in addition, Gaussian noise and random polygon shadow occlusion are introduced to simulate real outdoor occlusion and sensor noise interference. Figure 7 The displacement curve tracking images of different methods under degradation conditions are shown. The results of each method under synthetic degradation conditions are shown in Table 2-3, which presents the measurement results of YOLO-Vib and other vibration measurement methods on five degradation datasets. As can be seen from the table, the method of this invention can overcome the trajectory drift caused by the failure of the "constant brightness" assumption or feature loss in traditional methods when facing extreme illumination changes and high-speed motion blur, maintaining sub-millimeter-level measurement accuracy. Table 2 ; Table 3 ; To verify the applicability of this invention in a real engineering environment, this embodiment further collected vibration video data (resolution 1024×896, 50FPS) of a large outdoor bridge (Collaborative Innovation Bridge) under real interference such as natural light fluctuations and background clutter as the outdoor test object (denoted as dataset C). Table 4 shows the comparison results of different methods in outdoor bridge measurement, and the displacement curves of each method are as follows. Figure 8 As shown in the results, existing deep learning models (such as YOLOv12-KP) easily lose fine-grained features under weak outdoor vibrations, resulting in their inability to detect minute vibrations. The displacement curve of YOLO-Vib closely matches the real physical benchmark of a high-precision laser sensor, demonstrating that the method of this invention exhibits greater robustness to environmental changes and can adapt to various environmental variations.

[0060] Table 4 ; Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A vibration displacement measurement method based on a key point detection network, characterized in that, include: Obtain vibration image data of the structure, wherein key points are marked in the vibration images; The vibration image data is subjected to key point detection by a key point detection network to obtain key points. The displacement of the key points is then extracted and physically mapped to obtain vibration displacement data. The backbone network of the key point detection network includes functional modules for fitting non-rigid deformations and motion blur, and suppressing illumination interference.

2. The method according to claim 1, characterized in that, The key point detection network consists of a backbone network, a neck network, and a detection head arranged sequentially. The deep semantic extraction stage of the backbone network includes a feature enhancement module; The feature enhancement module includes a fast spatial pyramid pooling module, a geometric adaptive focusing module, and an illumination adaptive lightweight module connected in sequence. The adaptive focusing module is used to fit non-rigid deformation and motion blur, and the illumination adaptive lightweight module is used to suppress illumination interference. The neck network adopts a path aggregation network architecture, and the detection head adopts a decoupled head structure; the output features of the illumination adaptive lightweight module are processed by the C2PSA module and then used as the deep feature output of the backbone network.

3. The method according to claim 2, characterized in that, In the geometric adaptive focusing module, for the input feature map of the geometric adaptive focusing module, the spatial offset is obtained by performing offset prediction on the input feature map through a convolutional layer, and then the input feature map and the spatial offset are convolved through deformable convolution. Finally, the output feature of the adaptive focusing module is obtained by normalization and activation function.

4. The method according to claim 2, characterized in that, In the aforementioned light-adaptive lightweight module, the input features are processed through convolution to obtain an intensity map; the intensity values ​​in the intensity map are divided into different statistical intervals; the original features of the statistical intervals are extracted, and the original features are aggregated, enhanced, and mapped to obtain prototype features; a query vector is generated based on the original features, and a key vector and a value vector are generated based on the prototype features; the query vector, key vector, and value vector are processed through a multi-head self-attention module to obtain an attention output result; the attention output result is processed through convolutional branches of different scales to obtain different scale branch features; The different scale branch features are concatenated, and the corresponding branch weights are generated based on the concatenation results through a gating network. The different scale branch features are fused according to the branch weights, and the residuals of the fused features and the attention output results are added together. The features are then activated by an activation function to obtain the output features of the illumination adaptive lightweight module.

5. The method according to claim 1, characterized in that, The loss function of the keypoint detection network is direct coordinate regression loss.

6. The method according to claim 1, characterized in that, The process of displacement extraction and physical mapping of the key points includes: The initial static frame in the vibration image data is selected as the global reference benchmark, and the pixel displacement data of the key point at any time is extracted. Based on the known physical radius of the circular marker on the structure and the average pixel radius in the initial static frame, the mapping relationship between pixel coordinates and physical space coordinates is calculated. The pixel displacement data is then transformed into physical space according to the mapping relationship to obtain the vibration displacement data of the key point, i.e., the vibration displacement data of the structure.

7. The method according to claim 1, characterized in that, Prior to key point detection of the vibration image data, the following steps are included: The keypoint detection network is optimized using a sample dataset. The sample dataset contains images of circular markers. The initial center and radius of the circular markers are obtained from the original frames in the sample dataset using Hough circle transform. A region of interest is constructed based on the initial center. The edge point set of the sub-pixel level edge contour is extracted from the region of interest. The edge point set is then subjected to regression analysis using a least-squares ellipse fitting algorithm to obtain the corresponding geometric center coordinates as the true label of the sample dataset.

8. The method according to claim 1, characterized in that, Prior to key point detection of the vibration image data, the following steps are included: The keypoint detection network was validated using root mean square error, mean relative error, and Pearson correlation coefficient as indicators.

9. A vibration displacement measurement system based on a key point detection network, characterized in that, Used to perform the method described in any one of claims 1-8.

Images