Infrared target system based on linkage triggering of camera of image deflectometer and working method

Abstract

The invention discloses an infrared target system based on linkage triggering of an image deflection instrument camera and a working method, whether an abnormal event occurs on a bridge is judged through an image acquisition module, an infrared detection module is deployed, and bridge health monitoring is realized by adopting a target image deflection instrument; the detection control module is triggered based on an abnormal event, then target illumination instead of light emitting power is used as an evaluation index, when the target illumination exceeds a set illumination range, the system sends early warning information, and the wireless communication module sends the early warning information to the visual terminal. According to the invention, an attention mechanism is embedded on the basis of the traditional YoV8 network, the displacement information of the vehicle is fused into the channel attention, compared with the original YoV8 network, a feature map with direction perception and position sensitivity is formed, and a discrimination model is formed through the improved YoV8 network training and is embedded into an image processing unit to work in cooperation with a camera.

Description

Technical Field

[0001] The invention belongs to the technical field of bridge monitoring, operation and maintenance, and particularly relates to an infrared target target system based on the linkage triggering of an image deflection meter and a camera. Background Art

[0002] The static and dynamic deflections of a bridge reflect the overall vertical stiffness of the bridge structure, and are important indicators for evaluating the state and performance of the bridge structure. They are of great significance in bridge health monitoring. Therefore, accurately and conveniently monitoring the static and dynamic deflections during the operation stage of the bridge has become a practical need for evaluating the service performance of the bridge and providing timely structural safety warnings. At present, the main method for measuring the static and dynamic deflections of national and provincial trunk bridges with a large quantity and wide distribution is the image deflection meter. The image deflection meter is based on the digital image correlation (DIC) method, which is a measurement method for tracking (or matching) the same physical point (or pixel) between a reference image and a target image. Through correlation calculation and sub-pixel interpolation, the relative change in pixels of multiple target points on the image captured by the camera under the action of vehicle loads is accurately measured. Then, combined with the conversion between pixel coordinates and world coordinates, the actual deformation result is output. The DIC measurement system includes three parts: light spots, image acquisition, and image processing algorithms. Among them, the light spots have a great influence on the measurement accuracy of the static and dynamic deflections of the bridge. If the light spots are weak, the reference sub-region cannot be recognized in the camera field of view of the image deflection meter, resulting in large test errors. If the light spots are strong, they will interfere with other test light spots and consume a large amount of power. Considering that the infrared light has a longer wavelength and stronger ability to penetrate rain and fog than visible light, the current light spots are mainly infrared light, and the infrared light generally uses a target as a carrier.

[0003] For beam bridges, generally, the image deflection meter should be installed at the fixed point of the main span pier, the reference target should be installed at the fixed point of the opposite pier of the main span, and the target target should be installed at the quarter point and mid-span of the main beam bottom plate. However, when the infrared lamp beads of the target work for a long time in natural environments such as high temperature and high humidity, the electronic components will age. Coupled with poor heat dissipation of some targets, it will lead to light decay or even damage. At present, the infrared lamp beads of existing targets are mainly connected in series. As long as one lamp bead is damaged, the entire target will fail. At the same time, for cross-river bridges, there are more and more ships on inland waterways, and they are developing towards large-scale. Some old bridges have defects in initial design and construction, or the navigable clearance can no longer meet the navigational needs of modern ships. Ship collisions with bridges or targets occur frequently, resulting in damage to the infrared lamp beads of the targets and interruption of deflection monitoring, posing a hidden danger to the safe operation of the structure. In addition, since some targets are installed at the bottom of the bridge or in wading areas, inspection personnel need to rent bridge inspection vehicles or ships to maintain and replace the targets, which is costly and dangerous, bringing great inconvenience to the maintenance work. At present, in the field of existing bridge health monitoring, cameras are often equipped for the image deflection meter to monitor the bridge deck conditions to verify abnormal bridge deflections. Summary of the Invention

[0004] The object of the present invention is to provide an infrared target target system and a working method based on the linkage trigger of an image deflection meter camera, which combines the abnormal events of a bridge and the operation of an image deflection meter, can ensure the monitoring effect while reducing the operation and maintenance cost, and uses the target illumination rather than the luminous power as the evaluation index, which can effectively cope with the problem of lamp bead aging and improve the accuracy of control.

[0005] Technical solution: An infrared target target system based on the linkage trigger of an image deflection meter camera, which is used for the health monitoring of a bridge. The system includes:

[0006] An image acquisition module, including monitoring devices and an image processing unit deployed on the bridge, is used to collect bridge surface vehicle communication data and judge whether an abnormal event occurs on the bridge. The abnormal events include vehicle overload or overload caused by the aggregation of a number of vehicles in a unit area;

[0007] An infrared detection module, deployed on the bridge, includes a target infrared lamp bead, a constant current source and an image deflection meter. The target infrared lamp bead is used for the selection of the reference area and the target area of the image deflection meter, and the constant current source is used to provide a stable current output;

[0008] A detection control module, including an illuminance sensor and an illuminance control module. When the image acquisition module detects an abnormal event, the illuminance sensor monitors the target illuminance in real time. When the target illuminance exceeds the defined range, the illuminance control module drives the constant current source to adjust the output current size, so as to adjust the illuminance of the target; when the target illuminance exceeds the set illuminance range, the system issues a warning message, and the warning message is sent to the visualization terminal by the wireless communication module;

[0009] A power supply module is used to supply power to the electrical equipment and electrical components in the system.

[0010] Further, the target infrared lamp beads are arranged in series or parallel, and the combination of series and parallel can avoid the problem of the paralysis of the entire system caused by the failure of the target infrared lamp beads.

[0011] The present invention also provides a working method of an infrared target target system based on the linkage trigger of an image deflection meter camera, including the following steps:

[0012] S1. Set the target target and the reference target, adjust the focal length of the image deflection meter camera until the target target and the reference target can be clearly seen in the camera field of view. Record the illuminance at this moment by the illuminance sensor, and use this illuminance as the upper limit E_max of illuminance control; adjust the constant current source until the target target can just be recognized by the image deflection meter camera field of view. Record the illuminance at this moment by the illuminance sensor, and use this illuminance as the lower limit E_min of illuminance control;

[0013] S2. Monitor the vehicle conditions on the bridge deck through the image processing unit in the image acquisition module, determine whether there is an abnormal event, trigger the image disturbance meter to start working through the abnormal event, and the illuminance sensor monitors the illuminance E of the target infrared lamp beads;

[0014] S3. If the current target illuminance is within the set range (E_min, E_max), the target is in a normal working state;

[0015] S4. If the current target illuminance E is higher than E_max, the controller controls the constant current source to reduce the output current, thereby reducing the illuminance E of the infrared lamp beads until the illuminance E is within the set illuminance range;

[0016] S5. If the current target illuminance E is lower than E_min, the controller turns on the standby infrared lamp beads, the standby infrared lamp beads start to work, and the constant current source of the standby infrared lamp beads is adjusted until the illuminance E is within the set illuminance range;

[0017] S6. When the target illuminance E exceeds the set illuminance range, a warning message is sent and transmitted to the background through the wireless communication module, providing decision support for the maintenance and replacement of the target by the management and maintenance unit, and reducing the structural safety operation risk.

[0018] The present invention also provides a method for identifying abnormal events of bridges based on YOLOV8. The method is used for the image processing module unit in the system. The method includes constructing and improving the YOLOV8 target detection network model, using spatial and channel reconstruction convolutions to replace the convolutional layers in the original feature extraction network module, and adopting a pyramid skip connection method in the feature extraction network module to meet the recognition requirements of vehicle images of different scales caused by the distance difference;

[0019] The implementation process of the spatial reconstruction convolution includes: first, using the group normalization scaling factor to evaluate the information content of different feature maps, separating the feature maps with large information content and the feature maps with relatively small information content, and corresponding to the spatial content; then adding the features with more information and the features with less information to generate features with more information and save space;

[0020] The implementation process of the channel reconstruction convolution includes: splitting the features into two parts, performing different 1x1 convolution processes through different ratio paths, then performing transformation operations, further transforming the two parts of the features through global convolution and point convolution, and finally fusion. The two transformed features are weighted and fused through pooling and non-linear activation functions to form the final channel refined features.

[0021] Further, in the implementation process of the spatial reconstruction convolution, the mathematical representation of the group normalization process is:

[0022]

[0023] W = Gate(Sigmoid(W γ (GN(X)))))

[0024] where μ and σ are the mean and standard deviation, γ and β are trainable variables, ε is a constant for ensuring stability, and W γ is the normalized weight; then the weight values of the feature map reweighted by Wγ are mapped to the range (0, 1) through the sigmoid function and gated by a threshold; the weights higher than the threshold are set to 1 to obtain the informative weight W1, while the weights higher than the threshold are set to 0 to obtain the less informative weight W2; further, the input feature X is multiplied by W1 and W2 respectively to obtain two weighted features, which are the informative feature and the less informative feature

[0025] Spatial reconstruction refers to cross-reconstruction, specifically combining the two different informative features after weighting to obtain X ω1 and X ω2 , and after connecting them, the spatially refined feature map Xω is obtained. The mathematical calculation is expressed as:

[0026]

[0027] Unnecessary information is reduced through spatial and channel reconstruction volumes, effectively improving the feature representation efficiency, while reducing the number of model parameters and computational costs, meeting the purpose of real-time monitoring.

[0028] Furthermore, the YOLOV8 object detection network model includes a backbone feature extraction network, a feature fusion network, and a prediction network. The backbone feature extraction network is used to extract feature information from the photos of vehicles on the bridge deck, including the number of yellow card vehicles and the traffic volume of vehicles per unit time. The feature fusion network fuses the features extracted by the feature extraction network, fusing features of different scales. The prediction network is used to train the prediction results.

[0029] Even further, the sample data set for training the YOLOV8 object detection network model includes the following processing: First, a data set is obtained by collecting image information of vehicles on the bridge deck, and this data set is preprocessed. Then, the data set is labeled to construct a pre-training data set and a training data. The preprocessing includes image rotation, contrast enhancement, brightness enhancement, and data set augmentation to obtain the sample data set.

[0030] Beneficial effects: The system provided by the present invention realizes real-time detection of the traffic conditions of vehicles on the bridge deck through the image acquisition module, triggers the health monitoring of the bridge for abnormal events, and improves the accuracy and precision of monitoring through the image perturbation instrument. The present invention replaces the existing continuous monitoring with the continuity of the image acquisition module, improves the service life of the monitoring system (image perturbation instrument), etc. For non-abnormal events, the existing conventional detection can be used, which can also reduce the operation and maintenance costs and meet the low-power scenario requirements of the existing lightweight bridge monitoring system. Brief Description of the Drawings

[0031] Figure 1 is a schematic structural diagram of the system of the present invention;

[0032] Figure 2 is a schematic flowchart of the working method of the present invention;

[0033] Figure 3 is an application scenario illustrated in the embodiment;

[0034] Figure 4 is a structural diagram of a spatial and channel reconstruction convolutional network;

[0035] Figure 5 is a structural diagram of a spatial reconstruction unit network;

[0036] Figure 6 is a structural diagram of a channel reconstruction unit network;

[0037] Figure 7 is a structural diagram of a signal feature pyramid network. Detailed Embodiments

[0038] In order to enable those skilled in the art to clearly and accurately understand the technical solutions provided by the present invention, the following further introduction is made in conjunction with the accompanying drawings of the specification.

[0039] Combined with Figure 1 As shown, an infrared target target system based on the linkage trigger of an image deflection instrument camera of the present invention is used for the health monitoring of a bridge. The system includes:

[0040] A power supply module, which includes a 220v AC power supply, a transformer and a surge protector. The transformer is used to convert 220V AC power into DC power, and the surge protector prevents the large current caused by direct lightning or indirect lightning from instantaneously breaking through the infrared lamp beads;

[0041] The image acquisition module includes a high-definition bridge deck camera and an image processing system, which is used to determine whether there is heavy traffic or suspicious overloaded vehicles, that is, to judge abnormal time. Common abnormal events include heavy vehicles crossing the bridge, such as vehicles over a hundred tons, and also include the accumulation of vehicles on the bridge deck caused by traffic congestion. The present invention judges whether the bridge has experienced a heavy traffic event through the bridge deck camera, thereby avoiding the target being in a constantly lit state, prolonging the life of the target, and greatly reducing the maintenance cost.

[0042] The infrared detection module includes infrared lamp beads and a constant current source. The infrared lamp beads provide a light source for the selection of the reference area and the target area of the image deflection meter, and the constant current source can provide a stable current output.

[0043] The detection control module includes an illuminance sensor and an illuminance control module. The illuminance sensor monitors the illuminance of the target in real time. When the illuminance of the target exceeds the defined range, the illuminance control module drives the constant current source to adjust the magnitude of the output current, thereby adjusting the illuminance of the target.

[0044] When the illuminance of the target exceeds the set illuminance range, the system issues a warning message, and the wireless communication module sends the warning message to the visualization terminal.

[0045] Combined with Figure 2 , for the above system, the present invention provides the following working modes.

[0046] 1. Install the target and the reference target according to the construction drawing. According to the specific bridge type analysis of the bridge, the reference target and the deflection meter are installed at the fixed points of the abutment or the transverse diaphragm in the box chamber, and the target is installed at the main linear change position of the bridge and the quarter point position of the main beam. All targets should be avoided from being installed on the bridge deck to prevent the monitoring results from being inaccurate due to the influence of vehicle lights.

[0047] 2. Turn on the power of the target and the reference target, adjust the camera focus of the image deflection meter until the target and the reference target can be clearly seen in the camera's field of view. The illuminance sensor records the illuminance at this moment, and this illuminance is used as the upper limit E_max of illuminance control; adjust the constant current source until the target can just be recognized by the camera's field of view of the image deflection meter, and the illuminance sensor records the illuminance at this moment, and this illuminance is used as the lower limit E_min of illuminance control.

[0048] 3. Install the bridge deck camera to align with the bridge deck (note that the bridge deck camera needs to maintain a certain distance (greater than or equal to 50m) from the mid-span of the bridge). The camera is embedded with an image processing module. When it is judged that there are suspicious overloaded vehicles or heavy traffic vehicle congestion, a signal is triggered to the target to ensure that the deflection meter starts to work.

[0049] Combined with Figure 3 As shown, the abnormal events given in this embodiment include the following two types:

[0050] (1) There are ≥ 3 yellow-plate vehicles in the camera shooting area. According to the survey of existing monitoring system data, conventional blue-plate cars are relatively light in weight, usually 2-3 tons, and have no impact on the health of the bridge.

[0051] (2) Suspected overloaded vehicles, such as overloaded vehicles and vehicle congestion, are identified by training on existing overloaded vehicle image datasets.

[0052] 4. When an abnormal event occurs, the image deflectometer starts working, and the illumination sensor monitors the illumination E of the target infrared lamp beads;

[0053] 5. If the current target illumination is within the set range (E_min, E_max), the target is in normal working condition;

[0054] 6. If the current target illumination E is higher than E_max, the controller controls the constant current source to reduce the output current, thereby reducing the illumination E of the infrared lamp beads until the illumination E is within the set illumination range;

[0055] 7. If the current target illumination E is lower than E_min, the controller will connect the spare infrared lamp beads, the spare infrared lamp beads will start working, and adjust the constant current source of the spare infrared lamp beads until the illumination E is within the set illumination range.

[0056] 8. When the target illumination E exceeds the set illumination range, an early warning message is issued and transmitted to the background through the wireless communication module, providing decision support for the maintenance unit to repair and replace the target, thereby reducing the risk of structural safety operation.

[0057] Combination Figure 3-7 The present invention provides the following algorithm based on the image acquisition model. If the traffic conditions on the bridge deck are judged by a camera, photos of overloaded vehicles and traffic congestion are screened out through a large number of snapshots of the bridge deck and the snapshots of the weighing system, and the photos are marked. Because the vehicle is a moving object, the present invention embeds an attention mechanism on the basis of the traditional YoloV8 network, and integrates the displacement information of the vehicle into the channel attention. Compared with the original YoloV8 network, a feature map with direction perception and position sensitivity is formed. Based on this, this patent selects a large amount of bridge deck vehicle photo data sampled by the existing monitoring system as a training set (including large traffic conditions and conventional vehicles), and forms a discrimination model through the improved YoloV8 network training and embeds it into the image processing unit to work with the camera.

[0058] A bridge abnormal event recognition method based on YOLOV8, the implementation steps of which include:

[0059] Step 1: Collect a large number of photos of vehicles on the bridge deck captured by the bridge health monitoring system, select and match them to form a dataset, expand the dataset by means of rotation, contrast enhancement, and brightness enhancement, and label the dataset to obtain a pre-training dataset and a training dataset.

[0060] Step 2: Construct an improved YOLOV8 object detection network model. Use spatial and channel reconstruction convolutions to replace the traditional convolutions in the original feature extraction network module. Adopt a skip connection method in the feature extraction network module to meet the recognition requirements of vehicle images at different scales due to distance differences.

[0061] The improved YOLOV8 object detection network model includes: a backbone feature extraction network, a feature fusion network, and a prediction network. The backbone feature extraction network is used to extract feature information from a large number of photos of vehicles on the bridge deck. The feature fusion network fuses the features extracted by the feature extraction network, fusing features at different scales. The prediction network is used to train the prediction results.

[0062] Among them, in the present invention, the traditional convolution in the backbone feature extraction layer is changed to spatial and channel reconstruction convolution. Using spatial and channel reconstruction convolution can reduce redundant calculations and promote representative feature learning. The network structure is divided into a spatial reconstruction unit and a channel reconstruction unit.

[0063] Spatial reconstruction unit: First, perform group normalization processing. Then is the separation operation. Specifically, through a series of weights, the features are weighted. These weights are calculated through the channels of the input features, after normalization and non-linear activation functions. Finally is the reconstruction. The two weighted parts of the features are each transformed, and finally reconstructed through addition and splicing operations to obtain spatially refined features.

[0064] The specific steps of the separation operation are as follows:

[0065] The normalization processing uses the group normalization scaling factor to evaluate the information content of different feature maps, separates the feature maps with large information content from those with small information content, corresponding to the spatial content.

[0066]

[0067] W = Gate(Sigmoid(W γ (GN(X)))))

[0068] Among them, μ, σ are the mean and standard deviation, γ, β are trainable variables, ε is a constant used to ensure stability, W γis the normalized weight; then, the weight values of the feature map reweighted by Wγ are mapped to the range (0, 1) through the sigmoid function and gated by a threshold; the weights higher than the threshold are set to 1 to obtain the information-rich weight W1, while the weights higher than the threshold are set to 0 to obtain the information-poor weight W2; further, the input feature X is multiplied by W1 and W2 respectively to obtain two weighted features, which are the information-rich feature and the information-poor feature

[0069] The reconstruction steps are as follows:

[0070] Add the feature with more information and the feature with less information to generate a feature with even more information and save space. The specific operation is cross-reconstruction, where the two different information features after weighting are combined to obtain X ω1 and X ω2 , and after connecting them, the spatially refined feature map Xω is obtained. The mathematical calculation is expressed as:

[0071]

[0072] Unnecessary information is reduced through spatial and channel reconstruction convolutions, effectively improving the representation efficiency of features, while reducing the number of model parameters and computational costs, meeting the purpose of real-time monitoring.

[0073] The process of the channel reconstruction unit includes:

[0074] Segmentation: The feature is segmented into two parts and subjected to different 1x1 convolution processes through paths with different ratios.

[0075] Transformation: The two parts of the feature are further transformed through global convolution and point convolution.

[0076] Fusion: The two transformed features are weighted and fused through pooling and non-linear activation functions to form the final channel-refined feature.

[0077] The present invention reduces unnecessary information through spatial and channel reconstruction convolutions, effectively improving the representation efficiency of features, while reducing the number of model parameters and computational costs, meeting the purpose of real-time monitoring.

[0078] The above method incorporates an image pyramid structure into the network, which satisfies the feature fusion of feature maps at different scales. Finer / more dense sampling can capture more details, while coarser / more sparse sampling can observe the overall trend. The two are connected by skip connections to achieve accurate recognition of objects at different scales in the image. The features extracted by the backbone feature extraction network have a high downsampling ratio and a large receptive field, making them suitable for detecting large objects. The feature maps with a low downsampling ratio have a small receptive field and are suitable for detecting small objects. The combination of the two meets the application scenarios of the present invention.

[0079] The present invention adopts the method of parallel infrared lamp beads, which greatly improves the service life and reliability of the target, reduces the number of operation and maintenance times, and lowers the operation and maintenance costs. Using the target illuminance rather than the luminous power as the evaluation index can effectively address the problem of lamp bead aging and improve the accuracy of control. It can monitor the target illuminance in real time, ensure that the target is always within the field of view of the image deflection meter camera, improve the continuity and accuracy of bridge deflection monitoring, reduce the problem of deflection test interruption caused by target failure, and lower the risk of structural safety operation, providing decision-making support for the maintenance and replacement of the target by the management and maintenance unit. At the same time, the target is linked with the bridge deck camera of the deflection meter. By using image recognition methods to determine whether there is a large traffic situation on the bridge deck, if a large traffic situation is judged, the target is triggered to emit light to monitor the current real-time deflection of the bridge, greatly extending the service life of the target and reducing the power required for the entire deflection meter system, meeting the low-power scenario requirements of the existing lightweight bridge monitoring system.

Claims

1. An infrared target system based on the linkage triggering of an image deflectometer and a camera, characterized in that: The system is used for health monitoring of a bridge, and the system comprises: An image acquisition module, including monitoring equipment and an image processing unit deployed on the bridge, is used to collect the traffic information of vehicles on the bridge deck and determine whether an abnormal event occurs on the bridge, wherein the abnormal event includes an overloaded vehicle or an overload caused by the gathering of several vehicles in a unit area; The infrared detection module is deployed on the bridge and includes a target infrared lamp bead, a constant current source and an image deflectometer. The target infrared lamp bead is used to select the reference area and the target area of ​​the image deflectometer, and the constant current source is used to provide a stable current output. The detection control module includes an illumination sensor and an illumination control module. When the image acquisition module detects an abnormal event, the illumination sensor monitors the illumination of the target in real time. When the illumination of the target exceeds a specified range, the illumination control module drives the constant current source to adjust the output current, thereby adjusting the illumination of the target. A power module is used to provide power to the electrical equipment and electrical components in the system; When the target illumination exceeds the set illumination range, the system issues a warning message, which is then sent to the visualization terminal by the wireless communication module.

2. The infrared target system according to claim 1, characterized in that: In the infrared detection module, the target infrared lamp beads are arranged in series or in parallel.

3. A working method of an infrared target system based on the linkage triggering of an image deflectometer and a camera, characterized in that: The steps include: S1. Set the target target and the reference target, adjust the focal length of the image deflectometer camera until the target target and the reference target can be clearly seen in the field of view, and the illumination sensor records the illumination at that moment, and the illumination is used as the upper limit E_max of the illumination control; adjust the constant current source until the target can be just recognized by the field of view of the image deflectometer camera, and the illumination sensor records the illumination at that moment, and the illumination is used as the lower limit E_min of the illumination control; S2. Monitor the vehicle conditions on the bridge deck through the image processing unit in the image acquisition module to determine whether any abnormal event occurs. The abnormal event triggers the image disturbance meter to start working, and the illumination sensor monitors the illumination E of the target infrared lamp bead; S3. If the current target illumination is within the set range (E_min, E_max), the target is in normal working condition and the bridge health is in normal condition; S4. If the current target illumination E is higher than E_max, the controller controls the constant current source to reduce the output current, thereby reducing the illumination of the infrared lamp beads until the illumination E is within the set illumination range; S5. If the current target illumination E is lower than E_min, the external infrared detection module is controlled to connect the spare infrared lamp bead, the spare infrared lamp bead starts to work, and the constant current source of the spare infrared lamp bead is adjusted until the illumination E is within the set illumination range. The spare lamp bead is the target infrared lamp bead connected in series at this position in the infrared detection module; S6. When the target illumination E exceeds the set illumination range, an early warning message is issued and transmitted to the background through the wireless communication module to provide decision support for the management and maintenance unit to repair and replace the target.

4. A bridge abnormal event recognition method based on YOLOV8, characterized in that: The method is used for an image processing module in the system, and the method includes constructing and improving a YOLOV8 target detection network model, using spatial and channel reconstruction convolutions to replace the convolutional layers in the original feature extraction network module, and using a pyramid jump connection method in the feature extraction network module to meet the recognition requirements of vehicle images of different scales due to differences in distance; The implementation process of spatial reconstruction convolution includes: first, using the group normalization scaling factor to evaluate the information content of different feature maps, separating the feature maps with large information content from the feature maps with relatively small information content, and corresponding to the spatial content; then adding the features with more information and the features with less information to generate features with more information and save space; The implementation process of channel reconstruction convolution includes: dividing the features into two parts, performing different 1x1 convolution processing through paths of different proportions, and then performing transformation operations. The two parts of the features are further transformed through global convolution and point convolution, and finally fused. The two transformed features are weighted fused through pooling and nonlinear activation functions to form the final channel refined features.

5. The bridge abnormal event recognition method based on YOLOV8 according to claim 4 is characterized in that: In the implementation of spatial reconstruction convolution, the mathematical representation of group normalization is: W=Gate(Sigmoid(W γ (GN(X))))) Among them, μ, σ are the mean and standard deviation, γ, β are trainable variables, ε is a constant used to ensure stability, and W γ is the normalized weight; the weight value of the feature map reweighted by Wγ is then mapped to the range (0,1) by the sigmoid function and gated by the threshold; the weight above the threshold is set to 1 to obtain the informative weight W1, and the weight above the threshold is set to 0 to obtain the informative weight W2; the input feature X is further multiplied by W1 and W2 respectively to obtain two weighted features, and the distribution is the informative feature and less informative features Spatial reconstruction refers to cross reconstruction, which specifically combines the two weighted different information features to obtain X ω1 and X ω2 , after connecting them, we get the spatial refinement feature map Xω, which can be expressed mathematically as follows: Through spatial and channel reconstruction convolution, unnecessary information is reduced to improve the representation efficiency of features, while reducing the number of model parameters and computational cost to meet the purpose of real-time monitoring.

6. The bridge abnormal event recognition method based on YOLOV8 according to claim 4 is characterized in that: The YOLOV8 target detection network model includes a backbone feature extraction network, a feature fusion network and a prediction network. The backbone feature extraction network is used to extract feature information from bridge vehicle photos, including the number of yellow-plate vehicles and the vehicle traffic volume per unit time. The feature fusion network fuses the features extracted by the feature extraction network and fuses features of different scales. The prediction network is used to train the prediction results.

7. The bridge abnormal event recognition method based on YOLOV8 according to claim 4 or 6, characterized in that: The sample data set used for training the YOLOV8 target detection network model includes the following processing: first, a data set is obtained by collecting bridge deck vehicle image information, and the data set is preprocessed, and then the data set is labeled to construct a pre-training data set and training data. The preprocessing includes image rotation, contrast enhancement, brightness enhancement and data set expansion to obtain a sample data set.

Images