Bridge displacement monitoring method and system based on machine vision and timing error compensation

Abstract

The application discloses a kind of bridge displacement monitoring method and system based on machine vision and timing error compensation, the method is first by image acquisition module obtains bridge LED target image, using PAC-RTDETR model extracts LED target region of interest;Subsequently, construct neighborhood gray scale variance weight in region of interest, and obtain subpixel center by weighted two-dimensional Gaussian surface fitting;Again, combined with Lucas-Kanade optical flow tracking and abnormal rollback relocation mechanism, get continuous pixel displacement sequence;Finally, original displacement, temperature and camera state feature input fusion physical priori PIM-Modern TCN model, output temperature drift and mechanical disturbance compensation after bridge displacement result.The application can improve the stability of LED target recognition under complex conditions, displacement extraction accuracy and long-term monitoring result reliability, suitable for long-distance, non-contact displacement monitoring of bridge structure.

Description

Technical Field

[0001] This invention belongs to the field of bridge health monitoring and machine vision measurement technology, and specifically relates to a bridge displacement monitoring method and system based on machine vision and time-series error compensation, which can be used for long-distance, non-contact displacement monitoring, dynamic response acquisition and long-term monitoring error correction of bridge structures. Background Technology

[0002] Bridge displacement is a crucial response parameter reflecting the stress state and service performance of a bridge structure. Traditional contact-based displacement monitoring methods typically require the deployment of displacement gauges, guy wire sensors, or fixed supports near the bridge's measuring points. This places high demands on on-site installation conditions, line-of-sight conditions, and the traffic operation environment, limiting its application in long-span bridges, operational railway bridges, and long-term online monitoring scenarios. Machine vision-based displacement monitoring methods offer advantages such as non-contact operation, long-distance monitoring, flexible deployment, and cost control, and have been increasingly applied in the field of bridge health monitoring in recent years.

[0003] Existing machine vision methods for bridge displacement monitoring still have several shortcomings. First, under long-distance observation conditions, LED targets typically appear as small-scale, low-pixel-ratio bright spots in images, making them susceptible to complex backgrounds, ambient light fluctuations, strong reflections, and image noise, leading to unstable target detection boxes. Second, traditional geometric center, gray-scale centroid, or ordinary Gaussian fitting methods mainly rely on the gray-scale distribution of a single frame image. When the target spot exhibits edge distortion, local reflections, occlusion, or blurring, the center estimation is prone to shift, making it difficult to balance single-frame positioning accuracy with long-term temporal continuity. Third, during long-term bridge monitoring, changes in ambient temperature, camera self-heating, wind loads, vehicle traffic, and micro-vibrations of the supports all affect the imaging system, causing slow changes or instantaneous disturbances in the camera's intrinsic and extrinsic parameters and the observation state, resulting in low-frequency baseline drift and high-frequency noise in the displacement sequence.

[0004] Therefore, it is necessary to propose a visual displacement monitoring method for the complex service environment of bridges, so that it can stably identify LED targets in complex backgrounds, achieve high-precision sub-pixel positioning in the target area, maintain tracking continuity in long-term monitoring, and further compensate for systematic errors caused by temperature drift and camera mechanical disturbance. Summary of the Invention

[0005] To overcome the defects and shortcomings of the prior art, the present invention provides a bridge displacement monitoring method and system based on machine vision and time-series error compensation.

[0006] Specifically, this invention provides a bridge displacement monitoring method based on machine vision and time-series error compensation, comprising the following steps: (1) deploying LED target modules at the bridge location to be measured to provide stable artificial visual features; (2) continuously acquiring bridge monitoring images containing LED targets through an image acquisition module; (3) synchronously acquiring ambient temperature, camera temperature, acquisition timestamp, and camera operating status information through a temperature and camera status acquisition module; (4) using a PAC-RTDETR target detection module to identify LED targets in the bridge monitoring images and outputting LED target detection boxes and regions of interest; (5) within the region of interest, obtaining the sub-pixel centers of LED targets by using a CF-OFG sub-pixel displacement extraction module with neighborhood gray-level variance weighted two-dimensional Gaussian surface fitting, and generating a continuous pixel displacement sequence by combining Lucas-Kanade optical flow tracking and anomaly backoff mechanism; (6) using a pixel-to-physical displacement conversion module to convert the pixel displacement sequence into the original physical displacement sequence of the bridge according to camera calibration parameters, actual size of LED targets, or target spacing constraints; (7) using a PIM-Modern... The TCN timing error compensation module compensates for the displacement error caused by temperature drift and camera mechanical disturbance based on the original physical displacement sequence of the bridge, temperature information and camera status information, and outputs the compensated bridge displacement result; (8) The compensated displacement result is uploaded through the data transmission module and displayed through the Web visualization module.

[0007] This invention also provides a bridge displacement monitoring system based on machine vision and time-series error compensation, comprising:

[0008] LED target modules are deployed at the locations on the bridge to be measured to provide stable artificial visual features;

[0009] Image acquisition module, used to continuously acquire bridge monitoring images including LED targets;

[0010] The temperature and camera status acquisition module is used to synchronously acquire ambient temperature, camera temperature, acquisition timestamp, and camera operating status information;

[0011] The edge computing module deploys and runs the PAC-RTDETR target detection module, the CF-OFG sub-pixel displacement extraction module, and the PIM-Modern TCN timing error compensation module.

[0012] The PAC-RTDETR target detection module is used to identify LED targets in bridge monitoring images and output LED target detection boxes and regions of interest.

[0013] The CF-OFG subpixel displacement extraction module is used to extract the subpixel center of the LED target in the region of interest, and generate a continuous pixel displacement sequence by combining optical flow tracking and anomaly backoff mechanism.

[0014] The pixel-to-physical displacement conversion module is used to convert the pixel displacement sequence into the original physical displacement sequence of the bridge.

[0015] The PIM-Modern TCN timing error compensation module is used to compensate for displacement errors caused by temperature drift and camera mechanical disturbance based on the original physical displacement sequence of the bridge, temperature information and camera status information, and output the compensated bridge displacement result.

[0016] The data transmission module is used to upload the original displacement, compensated displacement, temperature information, equipment status and alarm information to the server or local monitoring terminal.

[0017] The web visualization module is used to display real-time displacement curves, historical monitoring data, alarm logs, and equipment operating status.

[0018] The advantages of this invention are:

[0019] 1. This invention uses the PAC-RTDETR model to extract the region of interest of LED targets. By using parallel dilated convolution and adaptive feature pyramid, the multi-scale feature representation of small-scale LED targets is enhanced, which helps to reduce false detections and false negatives under complex background, weak light, strong reflection and noise conditions.

[0020] 2. This invention employs a neighborhood gray-level variance weighted two-dimensional Gaussian surface fitting method, which incorporates the degree of local gray-level fluctuation into the sub-pixel center solution process, thereby reducing the impact of spot edge distortion, local reflection, and background disturbance on center estimation.

[0021] 3. This invention combines subpixel localization with Lucas-Kanade optical flow tracking and sets up an abnormal backtracking and relocalization mechanism, which can restart the detection and fine localization process when there is occlusion, blurring or local mistracking, thereby reducing the cumulative propagation of errors in the time series.

[0022] 4. This invention introduces physical prior branches and temporal feature branches through the PIM-Modern TCN model, and jointly models the camera parameter drift caused by temperature and the dynamic error caused by mechanical disturbance, which can simultaneously suppress low-frequency temperature drift error and high-frequency disturbance error.

[0023] 5. This invention can be deployed on edge computing devices, uploading only the compensated structured displacement results, reducing the bandwidth requirements for video data backhaul, and enabling real-time curve display, historical data query, and alarm prompts via a web interface, facilitating on-site engineering applications. Attached Figure Description

[0024] 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. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0025] Figure 1 This is an overall flowchart of the bridge displacement monitoring method and system described in this invention;

[0026] Figure 2 This is a structural diagram of the PAC-RTDETR target detection model described in this invention;

[0027] Figure 3 This is a structural diagram of the CSP-PAC parallel dilated convolution module described in this invention;

[0028] Figure 4 This is a diagram of the AFPN multi-scale feature fusion structure described in this invention;

[0029] Figure 5 This is a flowchart of the CF-OFG two-stage displacement extraction process described in this invention;

[0030] Figure 6 This is a flowchart illustrating the abnormal rollback mechanism described in this invention.

[0031] Figure 7 This is a schematic diagram of the two-dimensional Gaussian sub-pixel positioning described in this invention; wherein (a) is a heat map of the grayscale distribution of LED spot within the ROI; and (b) is a comparison diagram of the contour lines fitted to the two-dimensional Gaussian surface and the original grayscale distribution.

[0032] Figure 8 This is a flowchart of the LK optical flow timing tracking and abnormal backoff relocation process described in this invention;

[0033] Figure 9 This is a structural diagram of the PIM-Modern TCN timing error compensation model described in this invention;

[0034] Figure 10 This is a diagram of the edge computing node data stream processing architecture described in this invention;

[0035] Figure 11 This is a flowchart illustrating the functional interaction of the web-based visualization platform described in this invention.

[0036] Figure 12 This is a schematic diagram of the on-site layout of the bridge described in this invention. Detailed Implementation

[0037] 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 so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0038] The technical solution of this invention is: a bridge displacement monitoring method and system based on machine vision and temporal error compensation, comprising an LED target, an image acquisition module, a temperature acquisition module, an edge computing module, a data transmission module, a server, and a visualization display module. The image acquisition module continuously acquires images of the bridge LED target; the edge computing module sequentially performs PAC-RTDETR target detection, neighborhood grayscale variance-weighted two-dimensional Gaussian subpixel localization, Lucas-Kanade optical flow tracking, anomaly backtracking relocation, and PIM-Modern TCN temporal error compensation; the data transmission module uploads the compensated displacement results to the server; and the visualization display module displays real-time displacement curves, historical data, and alarm information.

[0039] Preferably, the PAC-RTDETR target detection model is formed by introducing a parallel dilated convolution module (PAC) and an adaptive feature pyramid module (AFPN) on the basis of the RT-DETR detection framework. The multi-scale convolutional fusion of the PAC module can be expressed as:

[0040] (1)

[0041] In the formula, Indicates input features, Indicates different expansion rates.

[0042] In a further preferred embodiment, the AFPN module adjusts features at different scales through gating weights, and its fusion process can be represented as:

[0043] (2)

[0044] In the formula, Indicates the gating weight, and ⊙ indicates element-wise multiplication. and These represent the output features of the upsampling and downsampling paths, respectively.

[0045] More preferably, the neighborhood gray-level variance-weighted two-dimensional Gaussian subpixel localization first establishes a two-dimensional Gaussian model for the LED spot within the ROI:

[0046] (3)

[0047] In the formula, The two-dimensional spatial coordinates of the pixels within the region of interest (ROI). The amplitude of the Gaussian surface. For background items, and Let be the coordinates of the sub-pixel center to be solved. These represent the light spots at... shaft and Gaussian standard deviation along the axial direction.

[0048] Further preferred, pixels The mean gray level and the local gray level variance within the neighborhood are expressed as follows:

[0049] (4)

[0050] (5)

[0051] In the formula, Represented in pixels The local neighborhood centered on, The number of pixels in the neighborhood. This represents the gray level within the neighborhood.

[0052] A further preferred approach is to convert the local grayscale variance into pixel weights:

[0053] (6)

[0054] In the formula, To prevent the inclusion of tiny constants with zero denominators, a weighted least squares objective function is constructed:

[0055] (7)

[0056] The fitting parameters for the two-dimensional Gaussian surface are obtained by solving equation (7), and then... and Used as the center coordinates of the subpixel of the LED target.

[0057] More preferably, the Lucas-Kanade optical flow tracking is based on the assumptions of constant brightness and small displacement, with the constant brightness constraint being:

[0058] (8)

[0059] In the formula, This indicates that at time t, the image coordinates are at... The grayscale value at that location.

[0060] Performing a first-order Taylor expansion on equation (8) and neglecting higher-order terms, we obtain the optical flow constraint equation:

[0061] (9)

[0062] In the formula, This represents the spatial grayscale gradient of the image along the horizontal direction at the current pixel. This represents the spatial grayscale gradient of the image along the vertical direction at the current pixel. This represents the rate of change of the grayscale value of the current pixel over time. These represent the optical flow displacement components of the current pixel in the horizontal and vertical directions between adjacent frames, respectively.

[0063] Within a local window, optical flow constraints are imposed on multiple pixels simultaneously, and the optical flow vector is solved using least squares:

[0064] (10)

[0065] In the formula, The optical flow displacement of the target point between adjacent frames. The image gradient matrix, The pixel weight matrix This represents the time gradient term. The target center is updated as follows:

[0066] (11)

[0067] Furthermore, the abnormal rollback mechanism adopts one or more of the following criteria:

[0068] (12)

[0069] In the formula, The threshold for the displacement between adjacent frames. The threshold value for the deviation between the optical flow prediction point and the Gaussian fitting point. To detect the confidence threshold, This represents the center coordinates of the current frame predicted by LK optical flow continuous tracking. The coordinates represent the verification coordinates obtained by independently performing neighborhood gray-level variance weighted two-dimensional Gaussian fitting within the same frame. When equation (12) holds, the system re-executes target detection and sub-pixel fine localization.

[0070] A further preferred embodiment of the conversion from pixel displacement to bridge physical displacement can be expressed as:

[0071] (13)

[0072] In the formula, The scaling factor is determined by camera calibration or the physical distance to the target. These are the initial pixel coordinates. These are the current pixel coordinates.

[0073] A further preferred embodiment of the PIM-Modern TCN temporal error compensation model includes a physical prior branch and a temporal feature extraction branch. The physical prior branch predicts the camera parameter drift vector based on the temperature sequence and camera state features:

[0074] (14)

[0075] In the formula, The neural network mapping function representing the physical prior branch, Represents a temperature sequence. Indicates camera state characteristics, Indicates the length of the input time window. This represents the predicted camera drift parameters or internal state vector.

[0076] In a further preferred embodiment, the physical mapping layer maps the camera drift parameters to the theoretical drift field:

[0077] (15)

[0078] In the formula, This represents the physical mapping function formed by the camera imaging relationship. This indicates the theoretical drift of the image point or displacement caused by temperature drift.

[0079] In a further optimized configuration, the adaptive fusion module fuses physical prior features and temporal perturbation features:

[0080] (16)

[0081] (17)

[0082] In the formula, These are physical prior characteristics. Temporal features extracted for Modern TCN This represents the trainable fully connected weight matrix and bias vector of an adaptive fusion weight calculation network. As environmental state characteristics, For adaptive fusion weights.

[0083] Further optimized, the final compensated bridge displacement is:

[0084] (18)

[0085] In the formula, This is the original visual displacement. This is the comprehensive displacement error compensation amount output by the model. The bridge displacement after compensation.

[0086] Furthermore, a composite loss function is used for model training:

[0087] (19)

[0088] In the formula, This is the displacement compensation error term. For physical parameter monitoring items, For time series smoothing constraints, and These are the weighting coefficients.

[0089] like Figure 1 As shown, the bridge displacement monitoring method and system based on machine vision and temporal error compensation provided by this invention includes an LED target, an image acquisition module, a temperature acquisition module, an edge computing module, a data transmission module, a server, and a visualization display module. Specifically, the implementation process of the method of this invention includes the following:

[0090] 1. LED target deployment and image acquisition

[0091] LED targets are deployed at the locations on the bridge structure to be measured. These LED targets can be actively emitting targets, and single-point or multi-point targets can be set according to the bridge structure dimensions and imaging distance. An industrial camera and fixed-focus lens are used in the image acquisition module to acquire a continuous sequence of images including the LED targets. During the acquisition process, ambient temperature, camera body temperature, sampling timestamps, and camera status information are recorded simultaneously.

[0092] The image acquisition module connects to the edge computing module via Gigabit Ethernet or other high-speed interfaces, and the raw image sequence is input in real time into the subsequent detection and displacement calculation process. To ensure monitoring continuity, the sampling frame rate can be set according to the bridge's dynamic response frequency, and a certain background area around the LED target is retained within the camera's field of view.

[0093] 2. PAC-RTDETR target detection

[0094] like Figure 2 As shown, the acquired bridge monitoring images are first input into the PAC-RTDETR target detection model. This model is based on the RT-DETR detection framework and embeds PAC and AFPN structures in the backbone network and hybrid encoder to enhance the perception capability of small LED targets in complex backgrounds.

[0095] like Figure 3As shown, the PAC module employs a split-then-merge cross-stage partial connection strategy, dividing the input features into two paths along the channel dimension: one path enters a multi-dilation rate parallel convolution branch to extract LED spot features under different receptive fields; the other path serves as a side branch to preserve original local detail information. The two feature paths are concatenated along the channel dimension and then fused through convolution for output.

[0096] like Figure 4 As shown, the AFPN module includes a parallel upsampling branch, a parallel downsampling branch, and a gated fusion mechanism. During upsampling, the network simultaneously utilizes deconvolution and interpolation upsampling to recover feature resolution; during downsampling, the network uses strided convolution and pooling in parallel modeling; the gated branch adaptively adjusts the contribution of features at different scales based on the feature response. Through this process, the system obtains the LED target detection box and the region of interest (ROI).

[0097] 3. Neighborhood gray-level variance weighted two-dimensional Gaussian subpixel localization

[0098] like Figure 5 and Figure 7 As shown, after obtaining the LED target ROI, the system enters the fine positioning stage. First, the grayscale image within the ROI is extracted, and the main energy region of the LED spot is used as the fitting object. Since the spot may be affected by reflection, defocus, noise, and edge distortion in actual bridge monitoring, directly using the grayscale centroid or ordinary Gaussian fitting can easily cause abnormal pixels at the edges to have a significant impact on the center estimation.

[0099] This invention calculates the gray-level variance of the pixel neighborhood within the Region of Interest (ROI), uses this variance to describe the degree of local gray-level fluctuation, and determines the pixel weight accordingly. Pixels with drastic gray-level fluctuations, which may be affected by noise or edge distortion, are assigned smaller weights; pixels with more stable gray-level distributions are assigned larger weights. Subsequently, a weighted two-dimensional Gaussian surface fitting is used to solve for the sub-pixel center coordinates of the LED target.

[0100] 4. LK optical flow tracking and abnormal backtracking relocation

[0101] like Figure 8 As shown, in order to maintain the continuity of the displacement sequence, the present invention uses the sub-pixel center coordinates of the previous frame as the initial point of LK optical flow tracking, solves the local optical flow displacement in the ROI of the adjacent frame, and updates the center coordinates of the LED target according to Equation (11).

[0102] In actual monitoring, short-term occlusion, blurred light spots, bright background interference, or tracking point drift may cause the optical flow results to fail. This invention incorporates an abnormal rollback mechanism, such as... Figure 6As shown, when the detection confidence, the deviation between the optical flow prediction point and the Gaussian fitting point, the displacement increment of adjacent frames, or the fitting residual exceeds the threshold, the system pauses the current optical flow tracking results, calls the PAC-RTDETR model again to re-detect the LED target, and re-executes the sub-pixel center localization.

[0103] 5. Coordinate Transformation and Bridge Displacement Calculation

[0104] After obtaining the continuous sub-pixel center sequence, the system calculates the pixel displacement for each frame based on the center coordinates of the initial frame or reference frame. For near-vertical imaging scenes, the pixel displacement can be converted into the physical displacement of the bridge structure using the known LED target spacing or the scaling factor obtained from camera calibration. For scenes with observation angle deviations, coordinate correction is performed using camera intrinsic and extrinsic parameters and the imaging model.

[0105] The above processing yields the original visual displacement sequence of the bridge structure. This sequence may still contain errors caused by temperature drift and camera mechanical disturbances, thus requiring a timing error compensation stage.

[0106] 6. PIM-Modern TCN Timing Error Compensation

[0107] like Figure 9 As shown, this invention employs the PIM-Modern TCN model to compensate for errors in the original visual displacement sequence. This model includes a physical prior branch, a Modern TCN temporal feature extraction branch, and an adaptive fusion module. The physical prior branch takes the temperature sequence and camera state features as input, predicts camera parameter drift or internal state vectors, and generates a theoretical drift field through a physical mapping layer.

[0108] The Modern TCN temporal feature extraction branch employs large-kernel deep convolution and residual connection structures to extract long-term drift trends and short-term dynamic perturbation features from displacement sequences. The adaptive fusion module generates fusion weights based on environmental conditions and feature responses, weights-fusing physical prior features and temporal features to output a comprehensive displacement error compensation.

[0109] Finally, the system corrects the original visual displacement according to equation (17) to obtain the compensated bridge displacement result. During the model training stage, the composite loss function shown in equation (18) can be used to ensure that the compensation result satisfies the requirement of minimizing displacement error, while maintaining the physical consistency of camera parameter drift prediction and the smoothness of the compensation sequence.

[0110] 7. Edge computing deployment and web visualization output

[0111] like Figure 10As shown, this invention can deploy PAC-RTDETR target detection, neighborhood gray-level variance-weighted two-dimensional Gaussian localization, LK optical flow tracking, and PIM-Modern TCN error compensation on an edge computing module. The edge directly receives camera image data and completes displacement calculation, uploading only structured data such as timestamps, original displacement, compensated displacement, temperature information, and alarm status to the server.

[0112] like Figure 11 As shown, the server and visualization module provide functions such as user login, real-time monitoring, historical data query, and alarm logs. The real-time monitoring interface displays the bridge displacement time history curve and system operating status; the historical data module allows querying and exporting monitoring records by time range; and the alarm module can output alarm information based on displacement thresholds, communication status, or compensation deviations.

[0113] like Figure 12 As shown, in engineering field applications, visual observation stations can be deployed on the bridge side bank or roadbed platform, and LED targets can be fixed at key sections of the bridge or the locations of components to be measured. After the camera's optical axis is aligned with the LED target area, long-distance, non-contact displacement monitoring can be performed.

[0114] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A bridge displacement monitoring method based on machine vision and temporal error compensation, characterized in that, The steps include: (1) Deploying LED target modules at the locations on the bridge to be measured to provide artificial visual features; (2) Continuously acquire bridge monitoring images containing LED targets through the image acquisition module; (3) Synchronously acquire ambient temperature, camera temperature, acquisition timestamp and camera operating status information through the temperature and camera status acquisition module; (4) Use the PAC-RTDETR target detection module to identify the LED targets in the bridge monitoring images and output the LED target detection box and region of interest; (5) In the region of interest, use the CF-OFG subpixel displacement extraction module to obtain the subpixel center of the LED target by fitting a two-dimensional Gaussian surface with neighborhood gray-level variance weighting, and combine it with Lucas-Kanade optical flow tracking and anomaly backoff mechanism to generate a continuous pixel displacement sequence; (6) Use the pixel-physical displacement conversion module to convert the pixel displacement sequence into the original physical displacement sequence of the bridge according to the camera calibration parameters, the actual size of the LED target or the target spacing constraint; (7) Use the PIM-Modern The TCN timing error compensation module compensates for displacement errors caused by temperature drift and camera mechanical disturbance based on the original physical displacement sequence of the bridge, ambient temperature, camera temperature, and camera operating status information, and outputs the compensated bridge displacement result. (8) The compensated displacement results are uploaded via the data transmission module and displayed via the Web visualization module.

2. The bridge displacement monitoring method based on machine vision and time-series error compensation according to claim 1, characterized in that: The PAC-RTDETR target detection module includes a parallel dilated convolution module (PAC) and an adaptive feature pyramid module (AFPN). The parallel dilated convolution module uses multiple convolution branches with different dilation rates to extract local bright spot features and contextual features of the LED target. The adaptive feature pyramid module achieves multi-scale feature adaptive fusion through parallel upsampling, parallel downsampling, and gated branches.

3. The bridge displacement monitoring method based on machine vision and time-series error compensation according to claim 2, characterized in that: The PAC module adopts a cross-stage partial connection structure, which divides the input features into a feature modeling branch and a bypass preservation branch. The feature modeling branch is used for multi-dilation rate parallel convolution, and the bypass preservation branch is used to preserve the original local detail information. After channel concatenation and convolution fusion, the two are used to output multi-scale LED target features.

4. The bridge displacement monitoring method based on machine vision and time-series error compensation according to claim 1, characterized in that: The weighted two-dimensional Gaussian surface fitting uses the pixel grayscale value within the region of interest as the observation value, the two-dimensional Gaussian function as the fitting function, and the weight obtained by the neighborhood grayscale variance as the pixel contribution coefficient. The fitting parameters are solved by weighted least squares, and the sub-pixel center coordinates of the LED target are obtained by inverse solution based on the fitting parameters.

5. The bridge displacement monitoring method based on machine vision and time-series error compensation according to claim 1, characterized in that: The Lucas-Kanade optical flow tracking constructs a set of gray-level gradient equations within a local window centered on the sub-pixel center, and uses the least squares method to solve for the optical flow displacement between adjacent frames, so that the center position of the LED target is continuously updated in the time series.

6. The bridge displacement monitoring method based on machine vision and time-series error compensation according to claim 1, characterized in that: The abnormal rollback mechanism adopts one or more of the following criteria: the displacement increment of adjacent frames exceeds the displacement threshold, the deviation between the optical flow prediction point and the Gaussian fitting point exceeds the deviation threshold, the target detection confidence is lower than the confidence threshold, or the fitting residual exceeds the residual threshold; when any criterion is met, target detection and sub-pixel localization are re-executed.

7. The bridge displacement monitoring method based on machine vision and time-series error compensation according to claim 1, characterized in that: The pixel-to-physical displacement conversion includes the conversion relationship based on camera intrinsic parameters, extrinsic parameters, and calibration scaling factor; when the camera is approximately perpendicular to the plane of the bridge measurement point, the pixel displacement is converted into physical displacement through the scaling factor; when there is an observation angle deviation, coordinate correction is performed in conjunction with the camera imaging model.

8. The bridge displacement monitoring method based on machine vision and time-series error compensation according to claim 1, characterized in that: The PIM-Modern TCN temporal error compensation module includes a physical prior branch, a temporal feature extraction branch, and an adaptive fusion module. The physical prior branch is used to establish the mapping relationship between temperature changes and camera parameter drift. The temporal feature extraction branch is used to extract long-term drift features and dynamic disturbance features from the original displacement sequence. The adaptive fusion module is used to fuse physical prior features and temporal features to obtain the displacement error compensation amount.

9. The bridge displacement monitoring method based on machine vision and time-series error compensation according to claim 8, characterized in that: The camera parameter drift includes one or more of the following: lateral pixel scale change, vertical pixel scale change, principal point lateral coordinate change, principal point vertical coordinate change, and camera attitude angle change; the camera parameter drift is converted into a theoretical drift field through a physical mapping layer, and then the displacement error compensation amount is generated from the theoretical drift field. The PIM-Modern TCN timing error compensation module is trained using a composite loss function that includes displacement error terms, physical parameter supervision terms, and timing smoothing terms. This makes the compensated displacement sequence close to the reference displacement or stable benchmark, while constraining the camera parameter drift output by the physical prior branch to have physical consistency.

10. A bridge displacement monitoring system based on machine vision and time-series error compensation, characterized in that, include: LED target modules are deployed at the locations on the bridge to be measured to provide artificial visual features; Image acquisition module, used to continuously acquire bridge monitoring images including LED targets; The temperature and camera status acquisition module is used to synchronously acquire ambient temperature, camera temperature, acquisition timestamp, and camera operating status information; The edge computing module deploys and runs the PAC-RTDETR target detection module, the CF-OFG sub-pixel displacement extraction module, and the PIM-Modern TCN timing error compensation module. The PAC-RTDETR target detection module is used to identify LED targets in bridge monitoring images and output LED target detection boxes and regions of interest. The CF-OFG subpixel displacement extraction module is used to extract the subpixel center of the LED target in the region of interest, and generate a continuous pixel displacement sequence by combining optical flow tracking and anomaly backoff mechanism. The pixel-to-physical displacement conversion module is used to convert the pixel displacement sequence into the original physical displacement sequence of the bridge. The PIM-Modern TCN timing error compensation module is used to compensate for displacement errors caused by temperature drift and camera mechanical disturbance based on the original physical displacement sequence of the bridge, temperature information and camera status information, and output the compensated bridge displacement result. The data transmission module is used to upload the original displacement, compensated displacement, ambient temperature, camera temperature, equipment status, and alarm information to the server or local monitoring terminal. The web visualization module is used to display real-time displacement curves, historical monitoring data, alarm logs, and equipment operating status.

Images