Bridge differential settlement monitoring method based on differential laser projection

Abstract

The application provides a bridge differential settlement monitoring method based on differential laser projection, and relates to the technical field of civil engineering structure health monitoring, and comprises the following steps: deploying a laser projection unit and an image acquisition unit in a measured area of a bridge; establishing a mapping relationship between pixel coordinates and physical dimensions of the image acquisition unit through a static multi-point calibration method; continuously acquiring laser spot images, pre-processing the acquired images, extracting features, and constructing an initial feature library; real-time monitoring and data acquisition, projecting a laser pattern to a target unit, filtering out ambient light, and capturing laser spot images; noise preprocessing; extracting feature points and performing preliminary positioning; time dimension correlation verification; spatial dimension correlation verification; and determining final effective feature point displacement data. The application has the advantages of stable measurement reference, strong environmental adaptability, high cost-effectiveness, high automation degree, and reliable precision.

Description

Technical Field

[0001] This invention relates to the field of civil engineering structural health monitoring technology, and in particular to a method for monitoring differential settlement of bridges based on differential laser projection, which is mainly applicable to safety assessment of old bridges, post-disaster emergency detection, and long-term bridge monitoring. Background Technology

[0002] The long-term safe operation of large-scale infrastructure such as bridges and buildings relies heavily on accurate monitoring of structural deformation, particularly foundation settlement. Structural displacement monitoring technology aims to continuously and accurately acquire deformation data of key structural components, assess their health status, provide early warnings of potential risks, and offer a scientific basis for maintenance decisions. As engineering structures become larger and more complex, increasingly higher demands are placed on the accuracy, automation, long-term stability, and environmental adaptability of monitoring technologies. Traditional manual inspection methods are no longer sufficient to meet the needs of modern infrastructure's full life-cycle health management. Therefore, developing efficient, reliable, and intelligent automated monitoring technologies has become an important research direction in the field of civil engineering safety.

[0003] Currently, existing technologies applied to structural displacement monitoring mainly fall into the following categories: The first category is contact-based manual monitoring, such as using dial indicators or electronic dial gauges. While this method is intuitive and low-cost, it heavily relies on manual on-site operation and readings, resulting in low efficiency, susceptibility to subjective errors, and difficulty in achieving long-term, continuous, and automated monitoring. Furthermore, it poses safety risks in harsh environments or high-risk locations. The second category is monitoring systems based on total stations, which utilize phase-based laser ranging for high accuracy. However, the measurement process still requires manual intervention to precisely align the reflecting prism. Equipment installation is strictly limited by line-of-sight conditions, and single-point measurement modes struggle to achieve high-density, continuous spatial monitoring. System deployment and maintenance costs are also high. The third category is machine vision solutions, including those based on physical targets and targetless methods. There are two technical approaches. The former relies on specially designed marker plates and calibration, which is easily affected by changes in lighting, target occlusion, or contamination, resulting in insufficient stability. The latter utilizes natural features, but its algorithm is complex and highly dependent on image quality and feature texture. In long-term monitoring, it is prone to feature matching failure due to environmental changes, and its computational load is large, posing challenges to real-time performance. Its accuracy is significantly affected by camera parameter drift and atmospheric disturbances. The fourth type is fiber optic grating sensing technology, which has the advantages of distributed measurement and resistance to electromagnetic interference. However, this technology essentially measures strain rather than direct displacement, requiring indirect calculation through complex strain-displacement conversion models and finite element analysis. Model errors directly affect the reliability of displacement results, and the sensor deployment process is complex, with large initial investment. Local damage may lead to the failure of the entire line.

[0004] In summary, existing technologies have significant shortcomings in terms of automation and continuity, accuracy and reliability, cost and applicability. There is an urgent need for a new method for direct monitoring of structural displacement that can achieve high precision, high reliability, true automation, and is suitable for long-term field environments. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a bridge differential settlement monitoring method based on differential laser projection, which features stable measurement benchmarks, strong environmental adaptability, high cost-effectiveness, high automation, and reliable accuracy.

[0006] The technical solution adopted by this invention to solve its technical problem is: a method for monitoring differential settlement of bridges based on differential laser projection, comprising the following steps: S1. System deployment and initialization, including S1.1 Deploy laser projection units and image acquisition units in the area of ​​the bridge to be measured; S1.2 Establish the mapping relationship between the pixel coordinates and physical size of the image acquisition unit through the static multi-point calibration method; S1.3 Under the condition that the system is installed and the environment is stable, control the image acquisition unit to continuously acquire laser spot images for a preset duration at a sampling frequency of 10Hz to obtain 18,000 frames of valid sample data; preprocess each frame of acquired image and extract the center coordinates, edge gradient features and shape factor features of the spot to construct an initial feature library; S2, Real-time Monitoring and Data Acquisition The laser projection unit and the image acquisition unit are activated. The laser projection unit projects a laser pattern onto the target unit, and the image acquisition unit filters out ambient light and captures the laser spot image through its equipped filter plate. During this process, the identifiability of the laser spot is ensured through an ambient light spectrum adaptive adjustment mechanism. S3. Feature point tracking and processing based on spatiotemporal correlation S3.1 Perform noise preprocessing on the real-time laser spot image acquired in step two; S3.2 Extract feature points from the preprocessed image and perform preliminary localization; S3.3. Perform correlation verification on the initially located feature points in the time dimension; S3.4. Perform correlation verification on the feature points that have passed the time verification in the spatial dimension to determine the final valid feature point displacement data.

[0007] In step S1.1, a laser projection unit is fixedly installed on pier A of the bridge to be tested, and an image acquisition unit is fixedly installed on pier B. The laser projection unit can project a cross laser pattern of a specific wavelength onto the target unit at the location of the image acquisition unit. The laser projection unit uses a cross and dot matrix composite laser source, which can emit dual-wavelength laser light of 650nm red light and 532nm green light. The image acquisition unit uses a binocular vision camera and is equipped with a multispectral filter group.

[0008] In step S1.2, N (N≥9) non-collinear static feature points (fixed marker points on the bridge surface, such as the edges of embedded parts or engravings) are uniformly selected in the bridge's measured area, such as the pier facade. A high-precision total station is used to measure the three-dimensional physical coordinates of each feature point, where the X-axis is the longitudinal direction of the bridge, the Y-axis is the transverse direction, and the Z-axis is the vertical direction. Simultaneously, the image acquisition unit is controlled to capture images of this area. A sub-pixel edge detection algorithm is used to extract the pixel coordinates of each feature point in the image, where u is the horizontal pixel index and v is the vertical pixel index. Based on the principle of perspective projection, a mapping model from pixel coordinates to physical coordinates is established, and the least squares method is used to solve for the mapping matrix M. For each feature point, the following conditions are met:

[0009] in, The perspective transformation matrix is ​​3×4. An overdetermined system of equations is constructed using the coordinate data of N feature points. The values ​​of each element of M are then solved, thereby realizing the transformation from arbitrary pixel coordinates to physical coordinates.

[0010]

[0011] Establish a structural constraint matrix. Based on bridge design specifications and structural mechanics analysis, assume the monitoring system deploys K measuring points, and define a K×K matrix. , of which elements This represents the maximum permissible relative settlement difference between the j-th and k-th measuring points, and its value is determined by the following formula:

[0012] In the formula, Let j be the horizontal distance between measuring points j and k. The maximum differential settlement between adjacent structures as specified in the bridge design documents, when j=k (No relative settlement at the same measuring point), matrix C is ensured through symmetry processing. This allows for the quantification of the relative displacement relationship between any two measuring points.

[0013] In step S1.3 (1) Coordinates of the center of the light spot The sub-pixel-level center coordinates are calculated using Gaussian surface fitting for an image with a gray-level distribution of... The light spot region is fitted with a two-dimensional Gaussian function using the least squares method:

[0014] in, The peak gray level of the light spot. , , respectively, are the standard deviations of grayscale distribution in the u and v directions. Subpixel-level center coordinates Given the background grayscale; solving for the extreme points of this function yields the coordinates of the light spot center in each frame of the image. The average of 18,000 frames of data is then used to obtain the initial center reference coordinates. ; (2) Edge gradient features Sobel edge detection is performed on the light spot region. The Sobel operator is shown in the formula below. The gradient magnitude and direction of the edge points are calculated, and a gradient direction histogram (HOG) is constructed: the light spot edge region is divided into 8×8 pixel cells, and the gradient magnitude distribution in 8 directions is counted in each cell, finally forming a 64-dimensional feature vector, which serves as the baseline template for the edge gradient.

[0015]

[0016] and These are the horizontal and vertical operators, respectively. (3) Shape factor characteristics The shape factor features include extracting the geometric parameters of the spot region, including roundness and aspect ratio, and taking the average value to obtain the initial shape factor reference value.

[0017] Circularity:

[0018] Among them, among them, The area of ​​the light spot. This is the perimeter of the light spot.

[0019] Aspect ratio:

[0020] Where a and b are the lengths of the major and minor axes of the smallest circumscribed ellipse of the light spot, respectively; The initial shape factor baseline value is obtained by averaging the S and R values ​​from 18,000 frames of data. and .

[0021] In step S2, the laser projection and image capture module ensures the identifiability of the laser spot under complex lighting conditions through an ambient light spectrum adaptive adjustment mechanism. The specific implementation process is as follows: First, a miniature fiber optic spectral sensor is integrated externally to the laser unit to collect ambient light spectral distribution data in real time and calculate the spectral irradiance of ambient light at the target wavelengths (532nm green light and 650nm red light). (W / (m2) nm); set the spectral irradiance threshold. Through laboratory calibration, it was determined that the strong light scene... =500W / m2 nm, foggy scene =100W / (m2 nm), establish wavelength switching decision logic: when > When the laser unit automatically switches to 650nm red light output; when At that time, maintain 650nm red light output, if < When the light output is at that time, it will switch to 532nm green light output.

[0022] In step S3.1, a multimodal noise dynamic classification and suppression algorithm is used to identify the noise type in the image in real time, such as vibration blur and raindrop occlusion, through a lightweight CNN, and to activate the filtering strategy accordingly. Among them, motion compensation filtering is used to process vibration blur, and neighborhood grayscale restoration is used to process occlusion. The CNN architecture uses an improved MobileNetV2 architecture to build a noise classification network, with the network depth compressed to 18 layers. The network structure is as follows: Input layer: Receives a 224×224×3 laser spot image; Feature extraction layer: It contains 6 depthwise separable convolutional modules. Each module consists of 3×3 depthwise convolution + 1×1 pointwise convolution. Noise features are extracted by increasing the dimensionality of pointwise convolution. The number of channels is 32, 64, 128 and 256 respectively. Each convolutional layer is followed by BatchNorm and ReLU6 activation functions. Pooling layer: Global average pooling is used to compress the feature map into a 256-dimensional vector; Output layer: Outputs the probability distributions of five noise types through a fully connected layer, including vibration blur, raindrop occlusion, dust adhesion, backlight overexposure, and no noise, using the Softmax activation function.

[0023] Where x is the input image, For the i-th type of noise, For the output value of the fully connected layer, Let be the probability that the image belongs to the i-th type of noise, and take the category with the highest probability as the recognition result.

[0024] In step S3.2, the Sobel operator is used to perform edge detection on the preprocessed image and sub-pixel localization is performed by Gaussian surface fitting to initially obtain the center coordinates of the laser spot.

[0025] In step S3.3, the temporal dimension correlation verification is achieved by constructing an LSTM-Attention prediction model to verify the temporal continuity of the light spot center coordinates, effectively identifying abnormal jumps caused by instantaneous interference.

[0026] The specific implementation process of step S3.3 is as follows: An LSTM-Attention prediction model is constructed, which adopts a hybrid architecture of Long Short-Term Memory (LSTM) network and attention mechanism. Taking the historical coordinate sequence as input, it predicts the theoretical coordinates of the center of the spot in the current frame. The structure is as follows: Input layer: Receives the sequence of light spot center coordinates from the previous 100 frames.

[0027]

[0028] Where t is the current frame time.

[0029] LSTM layer: Contains two hidden layers for extracting temporal features. The first LSTM layer outputs sequence H.

[0030] The calculation formula for LSTM layers is as follows:

[0031] in, , , These are the forget gate, input gate, and output gate, respectively. Candidate cell state, In cellular state, ( , , , ) is the weight matrix, ( , , , ) represents the bias term. It is the Sigmoid activation function. This is element-wise multiplication.

[0032] The attention layer assigns temporal attention weights to the hidden state H output by the LSTM, calculated as follows:

[0033]

[0034] in, Let be the attention weight for the i-th frame. Rate energy , For weight parameters, For bias terms The context vector is obtained by weighted fusion of hidden states:

[0035] Output layer: Outputs the predicted coordinates of the current frame through a fully connected layer. The activation function is a linear function:

[0036] in, , This is the weight matrix. , This is a bias term.

[0037] Let the measured coordinates of the current frame obtained by subpixel localization be... Calculate the Euclidean distance deviation between the predicted and actual coordinates:

[0038] If the value is ≤0.3mm, the measured coordinates are considered valid and output directly. ; When the value is greater than 0.3 mm, the coordinate is marked as a suspicious point, triggering the spatial dimension correlation verification process to further verify its rationality.

[0039] Step S3.4 calls the structural constraint matrix and, combined with the synchronous coordinate data of three adjacent measuring points, calculates the relative settlement difference between the suspected point and the three neighboring measuring points:

[0040] in, The vertical settlement of the suspected point. Let be the synchronous settlement of the j-th neighboring measuring point.

[0041] Determine whether the suspicious point conforms to the deformation law of the bridge structure. If it does, correct it and use the measured value; if it does not, it is determined to be an outlier and the value is removed.

[0042] The beneficial effects of this invention are as follows: By rigidly anchoring the laser projection and image acquisition unit directly to the bridge pier surface, the measurement reference can be synchronized with the deformation of the bridge itself, fundamentally avoiding measurement errors introduced by independent movement of the reference point, and ensuring the long-term stability of the reference and the structural following of the measurement system. Simultaneously, the use of a specific wavelength laser and a filtering device effectively suppresses interference from stray ambient light, and differential processing and other methods improve the signal-to-noise ratio in complex lighting environments, enhancing the system's environmental adaptability.

[0043] This method, by integrating an embedded processor and a pre-defined algorithm, automates the entire process from image acquisition, feature extraction, temporal and spatial correlation verification to settlement calculation, significantly reducing the frequency of manual inspections and interventions. While ensuring measurement accuracy, by optimizing the algorithm flow and hardware selection, a cost-effective monitoring system is constructed, providing a feasible technical solution for long-term, automated, and reliable monitoring of differential settlement in bridges. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the configuration of the present invention.

[0045] Figure 2 This is a diagram of the lightweight CNN structure of the present invention.

[0046] The attached diagram is labeled as follows: 1. Image acquisition unit; 2. Laser projection unit; 3. Pier A; 4. Pier B; 5. Filter plate; 6. Camera. Detailed Implementation

[0047] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0048] See Figure 1 As shown, this embodiment consists of a modular structure comprising a laser projection unit, an image acquisition unit 1, and a data processing core. The laser projection unit 2 is fixed to the pier A3 to be measured and projects a cross laser pattern of a specific wavelength onto the target unit of the pier B4. The image acquisition unit 1 includes an acrylic filter plate 5 and an imaging module. The acrylic filter plate 5 is mainly used to filter out ambient light. The embedded processor executes a preset algorithm to calculate the relative displacement. The entire system is directly anchored to the surface of the bridge body through rigid connectors to ensure that the measurement reference is synchronized with the structural deformation.

[0049] This embodiment includes a method for monitoring differential settlement of bridges based on differential laser projection, comprising the following steps: S1. System deployment and initialization, including S1.1 Deploy a laser projection unit 2 and an image acquisition unit 1 in the area of ​​the bridge to be measured. Fix the laser projection unit 2 on the pier A3 of the bridge to be measured and fix the image acquisition unit 1 on the pier B4. The laser projection unit 2 can project a cross laser pattern of a specific wavelength onto the target unit at the location of the image acquisition unit 1. The laser projection unit 2 adopts a cross and dot matrix composite laser source and can emit dual-wavelength laser light of 650nm red light and 532nm green light. The image acquisition unit 1 adopts a binocular vision camera 5 and is equipped with a multispectral filter group.

[0050] S1.2. Establish the mapping relationship between pixel coordinates and physical dimensions of the image acquisition unit using a static multi-point calibration method. The specific process is as follows: N (N≥9) non-collinear static feature points are uniformly selected in the bridge's measured area, such as the bridge pier facade. Fixed marker points on the bridge surface, the edges of embedded parts, and grooves are used. A high-precision total station is used to measure the three-dimensional physical coordinates of each feature point, where the X-axis is the longitudinal direction of the bridge, the Y-axis is the transverse direction, and the Z-axis is the vertical direction. Simultaneously, the image acquisition unit is controlled to capture images of this area. The pixel coordinates of each feature point in the image are extracted using a sub-pixel edge detection algorithm, where u is the horizontal pixel index and v is the vertical pixel index. Based on the perspective projection principle, a mapping model from pixel coordinates to physical coordinates is established, and the least squares method is used to solve for the mapping matrix M: For each feature point, the following conditions are met:

[0051] in, The perspective transformation matrix is ​​3×4. An overdetermined system of equations is constructed using the coordinate data of N feature points. The values ​​of each element of M are then solved, thereby realizing the transformation from arbitrary pixel coordinates to physical coordinates.

[0052]

[0053] Establish a structural constraint matrix. Based on bridge design specifications and structural mechanics analysis, assume the monitoring system deploys K measuring points, and define a K×K matrix. , of which elements This represents the maximum permissible relative settlement difference between the j-th and k-th measuring points, and its value is determined by the following formula:

[0054] In the formula, Let j be the horizontal distance between measuring points j and k. The maximum differential settlement between adjacent structures as specified in the bridge design documents, when j=k (No relative settlement at the same measuring point), matrix C is ensured through symmetry processing. This allows for the quantification of the relative displacement relationship between any two measuring points.

[0055] S1.3 Under the condition that the system is installed and the environment is stable, control the image acquisition unit to continuously acquire laser spot images for 30 minutes at a sampling frequency of 10Hz to obtain 18,000 frames of valid sample data; preprocess each frame of the acquired image and extract the center coordinates of the spot, edge gradient features and shape factor features to construct an initial feature library; (1) Coordinates of the center of the light spot The sub-pixel-level center coordinates are calculated using Gaussian surface fitting for an image with a gray-level distribution of... The light spot region is fitted with a two-dimensional Gaussian function using the least squares method:

[0056] in, The peak gray level of the light spot. , , respectively, are the standard deviations of grayscale distribution in the u and v directions. Subpixel-level center coordinates Given the background grayscale; solving for the extreme points of this function yields the coordinates of the light spot center in each frame of the image. The average of 18,000 frames of data is then used to obtain the initial center reference coordinates. ; (2) Edge gradient features Sobel edge detection is performed on the light spot region. The Sobel operator is shown in the formula below. The gradient magnitude and direction of the edge points are calculated, and a gradient direction histogram (HOG) is constructed: the light spot edge region is divided into 8×8 pixel cells, and the gradient magnitude distribution in 8 directions is counted in each cell, finally forming a 64-dimensional feature vector, which serves as the baseline template for the edge gradient.

[0057]

[0058] and These are the horizontal and vertical operators, respectively. (3) Shape factor characteristics The shape factor features include extracting the geometric parameters of the spot region, including roundness and aspect ratio, and taking the average value to obtain the initial shape factor reference value.

[0059] Circularity:

[0060] Among them, among them, The area of ​​the light spot. This is the perimeter of the light spot.

[0061] Aspect ratio:

[0062] Where a and b are the lengths of the major and minor axes of the smallest circumscribed ellipse of the light spot, respectively; The initial shape factor baseline value is obtained by averaging the S and R values ​​from 18,000 frames of data. and .

[0063] S2, Real-time Monitoring and Data Acquisition The laser projection unit and image acquisition unit are activated. The laser projection unit projects a laser pattern onto the target unit, and the image acquisition unit filters out ambient light and captures the laser spot image through its equipped filter plate. The laser projection and image acquisition modules ensure the recognizability of the laser spot under complex lighting conditions through an ambient light spectrum adaptive adjustment mechanism. The specific implementation process is as follows: First, a miniature fiber optic spectral sensor is integrated externally to the laser unit to collect ambient light spectral distribution data in real time and calculate the spectral irradiance of ambient light at the target wavelengths (532nm green light and 650nm red light). (W / (m2) nm); set the spectral irradiance threshold. Through laboratory calibration, it was determined that the strong light scene... =500W / m2 nm, foggy scene =100W / (m2 nm), establish wavelength switching decision logic: when > When the laser unit automatically switches to 650nm red light output; when At that time, maintain 650nm red light output, if < When the light output is at that time, it will switch to 532nm green light output. S3. Feature point tracking and processing based on spatiotemporal correlation S3.1. Perform noise preprocessing on the real-time laser spot image acquired in step two. Use a multimodal noise dynamic classification and suppression algorithm to identify the noise type in the image in real time, such as vibration blur and raindrop occlusion, through a lightweight CNN. Then, enable targeted filtering strategies, including motion compensation filtering to handle vibration blur and neighborhood grayscale restoration to handle occlusion. The CNN architecture uses an improved MobileNetV2 architecture to build a noise classification network, such as... Figure 2 As shown, the network depth is compressed to 18 layers, and the network structure is as follows: Input layer: Receives a 224×224×3 laser spot image; Feature extraction layer: It contains 6 depthwise separable convolutional modules. Each module consists of 3×3 depthwise convolution + 1×1 pointwise convolution. Noise features are extracted by increasing the dimensionality of pointwise convolution. The number of channels is 32, 64, 128 and 256 respectively. Each convolutional layer is followed by BatchNorm and ReLU6 activation functions. Pooling layer: Global average pooling is used to compress the feature map into a 256-dimensional vector; Output layer: Outputs the probability distributions of five noise types through a fully connected layer, including vibration blur, raindrop occlusion, dust adhesion, backlight overexposure, and no noise, using the Softmax activation function.

[0064] Where x is the input image, For the i-th type of noise, For the output value of the fully connected layer, Let be the probability that the image belongs to the i-th type of noise, and take the category with the highest probability as the recognition result.

[0065] S3.2 Extract feature points from the preprocessed image and perform preliminary localization. Use the Sobel operator to perform edge detection on the preprocessed image and use Gaussian surface fitting to perform sub-pixel localization to obtain the initial coordinates of the laser spot center.

[0066] S3.3. The initially located feature points are correlated and verified in the time dimension. This time-dimensional correlation verification is achieved by constructing an LSTM-Attention prediction model to verify the temporal continuity of the spot center coordinates, effectively identifying abnormal jumps caused by instantaneous interference. The specific implementation process is as follows: An LSTM-Attention prediction model is constructed, which adopts a hybrid architecture of Long Short-Term Memory (LSTM) network and attention mechanism. Taking the historical coordinate sequence as input, it predicts the theoretical coordinates of the center of the spot in the current frame. The structure is as follows: Input layer: Receives the sequence of light spot center coordinates from the previous 100 frames.

[0067]

[0068] Where t is the current frame time.

[0069] LSTM layer: Contains two hidden layers for extracting temporal features. The first LSTM layer outputs sequence H.

[0070] The calculation formula for LSTM layers is as follows:

[0071] in, , , These are the forget gate, input gate, and output gate, respectively. Candidate cell state, In cellular state, ( , , , ) is the weight matrix, ( , , , ) represents the bias term. It is the Sigmoid activation function. This is element-wise multiplication.

[0072] The attention layer assigns temporal attention weights to the hidden state H output by the LSTM, calculated as follows:

[0073]

[0074] in, Let be the attention weight for the i-th frame. Rate energy , For weight parameters, For bias terms The context vector is obtained by weighted fusion of hidden states:

[0075] Output layer: Outputs the predicted coordinates of the current frame through a fully connected layer. The activation function is a linear function:

[0076] in, , This is the weight matrix. , This is a bias term.

[0077] Let the measured coordinates of the current frame obtained by subpixel localization be... Calculate the Euclidean distance deviation between the predicted and actual coordinates:

[0078] If the value is ≤0.3mm, the measured coordinates are considered valid and output directly. ; When the value is greater than 0.3 mm, the coordinate is marked as a suspicious point, triggering the spatial dimension correlation verification process to further verify its rationality.

[0079] S3.4. In the spatial dimension, perform correlation verification on the feature points that have passed the time verification to determine the final valid feature point displacement data. Call the structural constraint matrix and combine it with the synchronous coordinate data of the three adjacent measuring points to calculate the relative settlement difference between the suspected point and the three neighboring measuring points:

[0080] in, The vertical settlement of the suspected point. Let be the synchronous settlement of the j-th neighboring measuring point.

[0081] Determine whether the suspicious point conforms to the deformation law of the bridge structure. If it does, correct it and use the measured value; if it does not, it is determined to be an outlier and the value is removed.

[0082] The technical features of this invention not described can be implemented by or using existing technology, and will not be repeated here. Of course, the above description is not a limitation of this invention, and this invention is not limited to the examples above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention should also be within the protection scope of this invention.

Claims

1. A method for monitoring differential settlement of bridges based on differential laser projection, characterized in that, Includes the following steps: S1. System deployment and initialization, including S1.1 Deploy a laser projection unit (2) and an image acquisition unit (1) in the area of ​​the bridge to be measured. S1.2 Establish the mapping relationship between the pixel coordinates and physical size of the image acquisition unit (1) by static multi-point calibration method; S1.3, Control Image Acquisition Unit (1) continuously acquires laser spot images, preprocesses each acquired frame of image, and extracts the center coordinates, edge gradient features and shape factor features of the spot to construct an initial feature library; S2, Real-time Monitoring and Data Acquisition The laser projection unit (2) and the image acquisition unit (1) are activated. The laser projection unit (2) projects a laser pattern onto the target unit, and the image acquisition unit (1) filters out ambient light and captures the laser spot image through its equipped filter plate (5). S3. Feature point tracking and processing based on spatiotemporal correlation S3.1 Perform noise preprocessing on the real-time laser spot image acquired in step two; S3.2 Extract feature points from the preprocessed image and perform preliminary localization; S3.

3. Perform correlation verification on the initially located feature points in the time dimension; S3.

4. Perform correlation verification on the feature points that have passed the time verification in the spatial dimension to determine the final valid feature point displacement data.

2. The method for monitoring differential settlement of bridges based on differential laser projection according to claim 1, characterized in that, In step S1.1, a laser projection unit (2) is fixedly installed on pier A (3) of the bridge to be tested, and an image acquisition unit (1) is fixedly installed on pier B (4). The laser projection unit (2) can project a cross laser pattern of a specific wavelength onto the target unit at the location of the image acquisition unit (1). The laser projection unit uses a cross and dot matrix composite laser source, which can emit dual-wavelength laser light of 650nm red light and 532nm green light. The image acquisition unit (1) uses a binocular vision camera (6) and is equipped with a multispectral filter group.

3. The method for monitoring differential settlement of bridges based on differential laser projection according to claim 2, characterized in that, In step S1.2, N non-collinear static feature points are uniformly selected in the bridge's measured area, such as the bridge pier facade, and the three-dimensional physical coordinates of each feature point are measured using a high-precision total station. At the same time, the image acquisition unit is controlled to take pictures of the area, and the pixel coordinates of each feature point in the image are extracted by a sub-pixel edge detection algorithm. Based on the principle of perspective projection, a mapping model from pixel coordinates to physical coordinates is established, and the mapping matrix M is solved by the least squares method.

4. The method for monitoring differential settlement of bridges based on differential laser projection according to claim 3, characterized in that, In step S1.3, the subpixel-level center coordinates are calculated using the Gaussian surface fitting method, and the gradient magnitude and direction of the edge points are calculated by Sobel edge detection on the spot area to construct a gradient direction histogram (HOG). The shape factor features include extracting the geometric parameters of the spot area, including roundness and aspect ratio, and taking the average value to obtain the initial shape factor reference value.

5. The method for monitoring differential settlement of bridges based on differential laser projection according to claim 1, characterized in that, In step S2, the laser projection and image capture module ensures the identifiability of the laser spot under complex lighting conditions through an ambient light spectrum adaptive adjustment mechanism. The specific implementation process is as follows: First, a miniature fiber optic spectral sensor is integrated externally to the laser unit to collect ambient light spectral distribution data in real time and calculate the spectral irradiance of ambient light at the target wavelengths (532nm green light and 650nm red light). (W / (m2) nm); set the spectral irradiance threshold. Through laboratory calibration, it was determined that the strong light scene... =500W / m2 nm, foggy scene =100W / (m2 nm), establish wavelength switching decision logic: when > When the laser unit automatically switches to 650nm red light output; when At that time, maintain 650nm red light output, if < When the light output is at that time, it will switch to 532nm green light output.

6. The method for monitoring differential settlement of bridges based on differential laser projection according to claim 1, characterized in that, In step S3.1, a multimodal noise dynamic classification and suppression algorithm is used to identify the noise type in the image in real time, such as vibration blur and raindrop occlusion, through a lightweight CNN, and to activate the filtering strategy accordingly. Among them, motion compensation filtering is used to process vibration blur, and neighborhood grayscale restoration is used to process occlusion. The CNN architecture uses an improved MobileNetV2 architecture to build a noise classification network, with the network depth compressed to 18 layers.

7. The method for monitoring differential settlement of bridges based on differential laser projection according to claim 6, characterized in that, In step S3.2, the Sobel operator is used to perform edge detection on the preprocessed image and sub-pixel localization is performed by Gaussian surface fitting to initially obtain the center coordinates of the laser spot.

8. The method for monitoring differential settlement of bridges based on differential laser projection according to claim 7, characterized in that, In step S3.3, the temporal dimension correlation verification is achieved by constructing an LSTM-Attention prediction model to verify the temporal continuity of the light spot center coordinates, effectively identifying abnormal jumps caused by instantaneous interference.

9. The method for monitoring differential settlement of bridges based on differential laser projection according to claim 8, characterized in that... The specific implementation process of step S3.3 is as follows: An LSTM-Attention prediction model is constructed, which adopts a hybrid architecture of Long Short-Term Memory (LSTM) network and attention mechanism. Taking the historical coordinate sequence as input, it predicts the theoretical coordinates of the center of the spot in the current frame. The structure is as follows: Input layer: Receives the sequence of light spot center coordinates from the previous 100 frames. Where t is the current frame time. LSTM layer: Contains two hidden layers for extracting temporal features. The first LSTM layer outputs sequence H. The calculation formula for LSTM layers is as follows: in, , , These are the forget gate, input gate, and output gate, respectively. Candidate cell state, In cellular state, ( , , , ) is the weight matrix, ( , , , ) represents the bias term. It is the Sigmoid activation function. This is element-wise multiplication. The attention layer assigns temporal attention weights to the hidden state H output by the LSTM, calculated as follows: in, Let be the attention weight for the i-th frame. Rate energy , For weight parameters, For bias terms The context vector is obtained by weighted fusion of hidden states: Output layer: Outputs the predicted coordinates of the current frame through a fully connected layer. The activation function is a linear function: in, , This is the weight matrix. , This is a bias term. Let the measured coordinates of the current frame obtained by subpixel localization be... Calculate the Euclidean distance deviation between the predicted and actual coordinates: If the value is ≤0.3mm, the measured coordinates are considered valid and output directly. ; When the value is greater than 0.3 mm, the coordinate is marked as a suspicious point, triggering the spatial dimension correlation verification process to further verify its rationality.

10. The method for monitoring differential settlement of bridges based on differential laser projection according to claim 1, characterized in that, Step S3.4 calls the structural constraint matrix and, combined with the synchronous coordinate data of three adjacent measuring points, calculates the relative settlement difference between the suspected point and the three neighboring measuring points: in, The vertical settlement of the suspected point. Let be the synchronous settlement of the j-th neighboring measuring point. Determine whether the suspicious point conforms to the deformation law of the bridge structure. If it does, correct it and use the measured value; if it does not, it is determined to be an outlier and the value is removed.

Images