A method for rapid evaluation of road and bridge earthquake damage and traffic capacity based on traffic monitoring images

Abstract

The application provides a road and bridge earthquake damage and traffic capacity rapid evaluation method based on traffic monitoring images, and belongs to the cross field of earthquake engineering, traffic engineering and computer vision technology. The technical problems of slow response, limited coverage, poor adaptability to complex imaging conditions of image recognition method, and lack of unified evaluation framework in the existing post-earthquake road and bridge damage evaluation are solved. A road and bridge earthquake damage grading rule and a traffic capacity mapping relationship are established. Real image data sets are collected after the earthquake and are labeled according to the grading rule to construct the real image data sets. A damage three-dimensional model is constructed, and a synthetic image data set is generated through domain randomization. The real and synthetic images are integrated into a joint data set. An evaluation model is established, and the joint data set is used for training and evaluation. The image to be evaluated is input into the model, the damage type and degree are output, the damage type and degree are corresponded to the grading, and the traffic state recognition result is output.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of earthquake engineering, traffic engineering and computer vision technology, and in particular relates to a rapid assessment method for earthquake damage and traffic capacity of roads and bridges based on traffic monitoring images. Background Technology

[0002] Following an earthquake, road infrastructure may suffer various forms of damage. Road sections may experience surface cracking, collapse, misalignment, uplift, and slope collapse; bridge joints may suffer damage such as deck system failure, beam misalignment, bearing abnormalities, cracking and tilting of piers or abutments, and partial collapse. Roads and bridges together determine the continuity and accessibility of the post-earthquake transportation network, and their damage status and remaining capacity directly affect the efficiency of personnel evacuation, emergency rescue, and material transportation.

[0003] Current post-earthquake damage assessment of road traffic systems mainly relies on manual on-site inspections, expert interpretation, or comprehensive judgment based on multi-source information. While these methods can reflect the condition of local facilities, they typically suffer from slow response times, high labor costs, strong subjectivity, and limited coverage under complex conditions such as road blockages, communication restrictions, aftershocks, and severe weather. Consequently, they are insufficient to meet the needs for rapid investigation of large-scale road and bridge nodes.

[0004] With the construction of urban road monitoring networks and key bridge monitoring equipment, traffic cameras have been widely deployed at road intersections, bridges, tunnels, and key road sections. These devices feature continuous online data acquisition, clear spatial location, timely data transmission, and convenient centralized management. They can quickly acquire image or video information after an earthquake, providing a data foundation for post-earthquake damage identification and traffic capacity assessment of roads and bridges based on monitoring images. However, traffic monitoring images often suffer from problems such as changing viewpoints, inconsistent scales, occlusion, low illumination, rain and fog, video compression, and background interference. This leads to a decline in the recognition performance of existing models in real post-disaster scenarios, or a serious underestimation of damage.

[0005] In summary, a rapid assessment method for seismic damage and traffic capacity of roads and bridges based on traffic monitoring images is proposed. Summary of the Invention

[0006] In view of this, the present invention aims to propose a rapid assessment method for road and bridge earthquake damage and traffic capacity based on traffic monitoring images, in order to solve the technical problems of slow response and limited coverage of manual inspections in the existing assessment of road and bridge damage after earthquakes, poor adaptability of image recognition methods to complex imaging conditions, and lack of a unified assessment framework.

[0007] To achieve the above objectives, the present invention adopts the following technical solution to provide a method for rapid assessment of seismic damage and traffic capacity of roads and bridges based on traffic monitoring images, including: Includes the following steps: S1. Based on the type and degree of road and bridge damage, establish a mapping relationship between seismic damage and traffic capacity: Level I road damage corresponds to normal traffic, Level II road damage corresponds to restricted traffic, and Level III road damage corresponds to no traffic; intact bridges correspond to normal traffic, and damaged bridges correspond to no traffic. S2. Collect real images of road and bridge damage after the earthquake, and label the images according to the classification rules in S1 to construct a real image dataset. S3. Construct a three-dimensional model of post-earthquake damage to roads and bridges, simulate different monitoring scenarios through domain randomization, and generate a synthetic image dataset; S4. Construct a joint dataset containing the real image dataset and the synthetic image dataset, and divide it into a training set, a validation set, and a test set; S5. Establish a rapid assessment model for earthquake damage and traffic capacity of roads and bridges. Use the joint dataset of S4 to train and evaluate the model. Input the monitoring image to be evaluated into the trained model and output the damage type and degree of the image, which corresponds to the damage classification of S1. Then, output the traffic status recognition result according to the mapping relationship of S1.

[0008] Furthermore, in S1, the road damage type includes at least one of road cracking, road collapse, road misalignment, road uplift, and slope collapse and accumulation; the bridge damage type includes at least one of bridge deck cracking, beam end displacement, beam misalignment, bearing abnormality, pier cracking and spalling, pier tilting, and local collapse.

[0009] Furthermore, in S1, road damage level I is basically intact or slightly damaged, road damage level II is moderately damaged, and road damage level III is severely damaged or destroyed; the criteria for judging the integrity of the bridge are that the bridge deck is smooth, the positional relationship between the beam and the pier is normal, and there is no obvious damage to the railings and ancillary facilities; the criteria for judging the damage to the bridge are that there are visible structural abnormalities in the main structure or ancillary facilities of the bridge.

[0010] Furthermore, the criteria for determining whether the slope is basically intact or slightly damaged are: crack width < 5cm, collapse diameter < 3cm, misalignment height < 3cm, bulge height < 3cm, or the slope does not intrude into the driving lane. The criteria for determining moderate damage are: crack width of 5-10cm, collapse diameter of 3-15cm, misalignment height of 3-15cm, uplift height of 3-15cm, or slope intrusion into the road surface but with a remaining continuous passable strip width ≥3.5m. The criteria for determining severe damage or destruction are: crack width > 10cm, collapse diameter > 15cm, misalignment height > 15cm, bulge height > 15cm, or slope intrusion into the road surface resulting in a continuous passable strip width < 3.5m.

[0011] Furthermore, in S2, when multiple types of road damage coexist in an image, the overall damage level of the image is determined based on the maximum level criterion; when a single image is insufficient to clearly determine the safety status of a bridge, the safety priority principle is adopted, and suspicious structural anomaly samples are classified into the bridge damage category.

[0012] Furthermore, S3 specifically includes the following steps: S31. Establish the basic geometric model of the road and bridge, and assign materials to them; S32. Construct road damage models with different types of road damage and bridge damage models with different types of bridge damage on the basic geometric model. S33. Introduce domain randomization parameters to batch render and generate synthetic image datasets.

[0013] Furthermore, in S33, the domain randomization parameters include at least camera position and attitude randomization, time and lighting condition randomization, weather condition randomization, object material randomization, background image randomization, and foreground interference randomization.

[0014] Furthermore, in S4, the joint dataset is constructed by verifying the labels of real and synthetic images based on the damage classification in S1, so that they are all converted into five image-level labels: bridge damage, bridge intact, road severe damage, road moderate damage, and road minor damage. Then, it is divided into training set, validation set, and test set.

[0015] Furthermore, in S5, the model uses ConvNeXtV2 as the backbone network and introduces a multi-scale damage perception attention module and a multi-scale damage feature fusion classification head.

[0016] Furthermore, in S5, the model employs a transfer learning strategy of pre-training with synthetic data and fine-tuning with real data, and incorporates low-light enhancement samples during the training phase.

[0017] Beneficial effects: 1. This invention uses traffic monitoring images as the main information source and incorporates road sections and bridge nodes into the same post-earthquake damage assessment framework, which can quickly transform the apparent damage information of roads and bridges in the monitoring images into engineering-usable traffic status results.

[0018] 2. This invention simultaneously establishes a three-level damage classification rule for roads and a two-classification damage label for bridges, and achieves consistent expression of road and bridge assessment results through a unified traffic capacity mapping, which is suitable for post-earthquake emergency route screening, rapid investigation, and traffic control auxiliary decision-making.

[0019] 3. This invention constructs training samples by jointly using real image data and synthetic image data, and introduces a domain randomization mechanism in the process of generating synthetic images. This helps to alleviate the problem of insufficient real post-earthquake image samples and the domain gap between synthetic data and real data, and improves the model's adaptability to complex monitoring scenarios.

[0020] 4. This invention introduces a low-light enhancement strategy during model training, enabling the model to better adapt to imaging degradation conditions such as low illumination at night, noise, blur, video compression, and local shadows, thereby reducing the risk of underestimating bridge damage and severe road damage. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention; Figure 2 Typical seismic damage diagrams for roads and bridges; Figure 3 A flowchart for generating synthetic images of roads and bridges based on domain randomization; Figure 4 Example image comparing real images and synthetic damage images of roads and bridges; Figure 5 A unified earthquake damage assessment model for roads and bridges; Figure 6 Comparison of Macro-F1 under different perturbation conditions. Detailed Implementation

[0022] Referring to the accompanying drawings, this embodiment provides a method for rapid assessment of seismic damage and traffic capacity of roads and bridges based on traffic monitoring images, including: S1. Based on the type and degree of road and bridge damage, establish a mapping relationship between seismic damage and traffic capacity: Level I road damage corresponds to normal traffic, Level II road damage corresponds to restricted traffic, and Level III road damage corresponds to prohibited traffic; intact bridges correspond to normal traffic, and damaged bridges correspond to prohibited traffic. Road earthquake damage types and classification rules: Based on historical earthquake damage data, road earthquake damage is mainly classified into five categories: road cracking, road collapse, road misalignment, road uplift, and slope landslide accumulation. Road cracking manifests as linear, network, or continuous cracks in the road surface; road collapse manifests as localized or large-scale depressions in the road surface; road misalignment manifests as significant vertical displacement of the roadbed on both sides of the crack; road uplift manifests as arching and lifting of the road surface after compression; slope landslide accumulation manifests as the collapse or sliding of slope soil and rock that intrudes into the road surface, causing road blockage or traffic interruption, such as... Figure 2 The image shown is a breakdown of road earthquake damage.

[0023] Road damage is classified into three levels. Level I corresponds to basically intact or minor damage, and the road is passable. Level II corresponds to moderate damage, and the road is still passable but traffic is inconvenient, requiring traffic restrictions or restoration after repair. Level III corresponds to severe damage or destruction, and the road is unsafe to pass or requires a long period of repair or reconstruction before restoration. For road cracking, crack width is used as the main quantitative indicator; for road collapse, collapse diameter and settlement are used as quantitative indicators; for road misalignment and road uplift, vertical height is used as the quantitative indicator; for slope landslides, whether soil and rock have intruded into the road surface and the width of the remaining continuous passable zone are used as criteria. When multiple types of road damage coexist, the highest level criterion is used, that is, the highest level among the individual damage assessment results is used as the comprehensive damage level of the road section, as shown in Table 1, which illustrates the road seismic damage classification rules.

[0024]

[0025] Bridge Seismic Damage Types and Grading Rules: As a key node in the road traffic system, the post-earthquake condition of bridges directly affects the connectivity of the transportation network. Bridge damage can occur in the bridge deck system, superstructure, bearing system, piers, abutments, and overall spatial relationships. Typical damage includes bridge deck cracking, pavement damage, railing damage, beam end displacement, beam misalignment, bearing misalignment or detachment, pier cracking and spalling, pier tilting, abutment cracking, and localized collapse. Considering the difficulty in reliably acquiring information on internal component damage, detailed bearing deformation, and hidden foundation damage from monitoring images, this invention classifies bridge image status into two categories: intact and damaged. An intact bridge indicates that the bridge deck, beams, bearings, piers, abutments, and ancillary facilities show no visible serious abnormalities, and traffic continuity remains normal. Damaged bridges indicate the presence of obvious cracks, spalling, misalignment, component deformation, bridge deck damage, bearing area anomalies, pier damage, or overall alignment abnormalities in the image, requiring manual review, traffic restrictions, or closure. Figure 2 The image shows common bridge earthquake damage images, and Table 2 shows the classification rules for bridge earthquake damage levels.

[0026]

[0027] Unified Traffic Capacity Mapping for Roads and Bridges: To facilitate model output for post-earthquake emergency response, this invention maps road and bridge conditions into three categories: normal traffic, restricted traffic, and prohibited traffic. Level I road damage corresponds to normal traffic, Level II road damage to restricted traffic, and Level III road damage to prohibited traffic. Intact bridges correspond to normal traffic, while damaged bridges correspond to prohibited traffic or at least require manual review, traffic restrictions, and structural inspection. Thus, the model can simultaneously serve the rapid assessment of road segments and bridge nodes under the same output system. Table 2 shows the unified rules for the earthquake damage levels and traffic conditions of roads and bridges.

[0028]

[0029] S2. Collect real images of road and bridge damage after the earthquake, and label the images according to the classification rules in S1 to construct a real image dataset.

[0030] The real image dataset is primarily collected from earthquake damage investigation reports of real earthquake events, scientific research databases, publicly available internet resources, and publicly released datasets of earthquake damage images of roads and bridges. For each image, the location, time, type of facility, damaged parts, damage phenomena, image source, and identifiable traffic obstruction are recorded to form a representative sample library of post-earthquake images of real roads and bridges.

[0031] For real road images, manual screening, classification, and labeling are performed according to the three-level road damage grading rules established in Step 1. If an image contains only a single type of road damage, it is graded based on its apparent damage and impact on traffic capacity; if multiple types of damage coexist in the image, the comprehensive damage level is determined according to the highest grade criterion. For real bridge images, they are labeled as either bridge intact or bridge damaged based on the presence of visible structural anomalies or potential traffic risks. If a single image cannot clearly determine the bridge's safety status, a safety-first principle is adopted, and samples with suspected structural anomalies are classified as bridge damaged or subject to manual review.

[0032] The collected real images undergo data cleaning, including removing duplicate samples, irrelevant scenes, severely blurred images, and images with unidentifiable subjects; images with significant differences in resolution, size, and composition are uniformly cropped, standardized in size, and have invalid regions removed; and category labels, directory structure, and file naming are standardized to ensure they can be used together with synthetic images to form training samples. S3. Construct a three-dimensional model of post-earthquake damage to roads and bridges, simulate different monitoring scenarios through domain randomization, and generate a synthetic image dataset.

[0033] To address the issues of limited real post-earthquake image samples, insufficient scene coverage, and high annotation costs, this invention further constructs a synthetic post-earthquake damage scene for roads and bridges, and generates a batch of synthetic image datasets based on a domain randomization method. The synthetic scene is constructed using Blender 3D modeling software, and damage morphology simulation and data output are completed by combining rigid body fracture, geometric deformation, Boolean cutting, material mapping, and scripted rendering. The overall process is shown in Figure 3.

[0034] The road damage model first establishes a basic model of the road's main geometry, lane lines, road surface texture, and the bearing surface of slopes or surrounding environments. A physically based rendering (PBR) material system is then used to assign color maps, normal maps, roughness maps, and displacement maps to the road surface, soil, and environmental objects. Subsequently, five typical post-earthquake road damage models are constructed: road cracking, collapse, misalignment, uplift, and slope collapse / deposition. Road cracking is simulated by generating irregular cracks through Boolean cutting and noise-driven boundaries; road collapse is simulated by rigid body decomposition and local depressions to depict road surface subsidence; road misalignment is simulated by vertical displacement dynamic triggers to depict fracture displacement; road uplift is simulated by local heave triggers to depict arching deformation; and slope collapse / deposition is simulated by rock fragmentation, gravity fall, and road surface accumulation to depict scenarios where the road is buried or partially blocked.

[0035] In the basic model stage, the bridge damage model decomposes the bridge into components such as bridge deck, main beams, crossbeams, supports, piers, abutments, and guardrails, and assembles them according to the structural relationships of a beam bridge. For intact bridge samples, the bridge geometry, materials, camera angle, lighting, background, and foreground interference are randomly changed without introducing structural damage to avoid the model mistakenly learning a single appearance as an intact state. For damaged bridge samples, one or more damaged components are randomly selected based on the intact model, and the damage location, scale, and intensity are randomly set. Bridge damage includes at least visible seismic damage forms such as bridge deck cracking and misalignment, beam displacement, anomalous bearing areas, cracking and spalling of piers or abutments, pier tilting, and local collapse. A comparison between the actual damage model and the synthetic damage model is shown in Figure 4.

[0036] After the damage model is established, domain randomization parameters are introduced to expand the scene. These domain randomization parameters include at least camera position and orientation randomization, time and lighting condition randomization, weather condition randomization, object material randomization, background image randomization, and foreground interference object randomization. Specifically, camera position and orientation are used to simulate different surveillance camera installation heights, shooting distances, and pitch angles; lighting conditions are used to simulate morning, noon, night, and low-light scenes; weather conditions are used to simulate environments such as rain, fog, snow, and dust; material randomization is used to change the appearance of objects such as roads, soil, concrete, steel, and rubber bearings; and background and foreground object randomization is used to increase scene complexity.

[0037] RGB images are generated through batch rendering controlled programmatically, and category labels and metadata are saved simultaneously. Road sample labels are determined by the road damage level, corresponding to minor, moderate, and severe road damage; bridge sample labels are determined by whether visible damage has been applied, corresponding to intact and damaged bridges. Metadata includes information such as sample number, facility type, label category, damage type, damage location, main geometric parameters, material number, camera parameters, lighting parameters, weather parameters, background type, and random seed.

[0038] S4. Construct a joint dataset containing the real image dataset and the synthetic image dataset, and divide it into a training set, a validation set and a test set.

[0039] After obtaining the real image dataset and the synthetic image dataset respectively, the two types of data are integrated to construct a joint real-synthetic image dataset of post-earthquake damage to roads and bridges. During the integration process, the real and synthetic images are first verified for labels according to the unified labeling system in step one, transforming them into five image-level labels: bridge damage, bridge intact, road severely damaged, road moderately damaged, and road slightly damaged. Then, the image size, naming format, directory structure, label file format, and metadata fields are standardized to form a standardized joint sample set that can be directly used for deep learning training.

[0040] The joint dataset is divided into training, validation, and test sets. Each category includes both real and synthetic samples as much as possible. The training set is used for model parameter learning, the validation set is used for model selection and early stopping control, and the test set is used for final performance evaluation. For synthetic data, training, validation, and test samples are divided according to a predetermined ratio; for real data, a fixed partitioning method is used to ensure the comparability of test results from different experimental routes, as shown in Table 4.

[0041]

[0042] S5. Establish a rapid assessment model for earthquake damage and traffic capacity of roads and bridges. Use the joint dataset of S4 to train and evaluate the model. Input the monitoring image to be evaluated into the trained model and output the damage type and degree of the image, which corresponds to the damage classification of S1. Then, output the traffic status recognition result according to the mapping relationship of S1.

[0043] After constructing the real-synthetic joint dataset, a rapid evaluation model for unified road and bridge identification was established. The model takes road or bridge monitoring images and keyframe images extracted from the monitoring videos as input, which are then cropped, resized, and standardized before being fed into a deep learning network. The network outputs five probability distributions, corresponding to bridge damage, bridge integrity, severe road damage, minor road damage, and moderate road damage. Simultaneously, based on the traffic capacity mapping relationship from step one, the image-level recognition results are converted into engineering handling information such as normal traffic, restricted traffic, or prohibited traffic.

[0044] In this scheme, the model uses ConvNeXtV2 as the backbone network, and introduces a multi-scale damage perception attention module and a multi-scale damage feature fusion classification head on top of it. The multi-scale damage perception attention module is used to enhance the model's response to key areas such as cracks, misalignments, gravel accumulation, bridge deck misalignment, beam anomalies, and pier damage; the multi-scale damage feature fusion classification head is used to simultaneously fuse local details, mid-level damage textures, and high-level semantic relationships. This model is only one preferred implementation of the present invention, and other image classification networks, visual Transformer networks, or lightweight networks capable of performing unified damage identification for roads and bridges can also be used as alternative implementations.

[0045] In terms of training strategy, a transfer learning process of "pre-training on synthetic data - fine-tuning on real data" is preferred. The first stage utilizes large-scale synthetic images to learn the basic visual features of road damage, bridge damage, and changes in traffic conditions. The second stage transfers the pre-trained parameters to real image data for fine-tuning, allowing the model to further adapt to imaging noise, background interference, shooting angles, and scene distribution in real surveillance images. To improve stability in all-weather surveillance scenarios, low-light enhancement samples such as low exposure, gamma darkening, low-light noise, low-light blur compression, and local shadows can be added during the training phase.

[0046] Example This embodiment first constructs a synthetic image dataset of post-earthquake damage to roads and bridges, and a real image dataset of post-earthquake damage to roads and bridges. The synthetic dataset is generated using the Blender domain randomization method and contains 4475 images; the real dataset is obtained by collecting, cleaning, and manually verifying real images of road and bridge earthquake damage, and contains 1697 images. The dataset categories include bridge damage, bridge integrity, severe road damage, minor road damage, and moderate road damage. The synthetic data is divided into training, validation, and test sets in approximately an 8:1:1 ratio, while the real data is divided in approximately a 0.7:0.15:0.15 ratio, with the real test set fixed for the final evaluation.

[0047] This embodiment calculates Accuracy, Balanced Accuracy, Macro-F1, Weighted-F1, and a high-risk underestimation metric on the real test set. Accuracy represents the model's overall classification accuracy across all test samples; Balanced Accuracy represents the average recall of each class, mitigating the impact of imbalanced sample sizes across classes on the evaluation results; Macro-F1 represents the arithmetic mean of the F1 scores for each class, reflecting the model's ability to identify a minority of classes; and Weighted-F1 represents the F1 score weighted by the sample size of each class, reflecting the model's overall classification performance across the entire sample distribution.

[0048] To further evaluate whether the model underestimates the high-risk state after the earthquake, this embodiment defines a high-risk underestimation index. This indicator mainly counts two types of high-risk underestimation errors: first, misjudging "bridge damage" as "bridge intact"; second, misjudging "severe road damage" as "moderate road damage" or "minor road damage".

[0049] set up To determine the number of samples in the test set whose true category is "bridge damage", To determine the number of samples in the test set whose true category is "severe road damage"; let... This represents the number of samples whose true category is "bridge damage" but were misclassified as "bridge intact". This represents the number of samples whose true category is "severe road damage" but were misclassified as "moderate road damage". Given the number of samples whose true category is "severe road damage" but were misclassified as "minor road damage," the underestimation rates for the three high-risk categories are defined as follows:

[0050]

[0051]

[0052] The high-risk, low-risk comprehensive index is defined as follows:

[0053] In the formula, The lower the value of the comprehensive index, the less likely the model is to misjudge bridge damage or severe road damage as a low-risk state.

[0054] In terms of model construction, this embodiment uses the MSDANeXt five-class classification model for roads and bridges. The training parameters are set as shown in Table 5. The model input size is uniformly set to 384×384 pixels. The backbone network uses ConvNeXtV2, followed by a multi-scale damage perception attention module and a multi-scale damage feature fusion classification head, ultimately outputting five probability distributions. During model training, a combination of cross-entropy loss and Focal Loss is used to reduce the impact of class imbalance on the training results, and to focus on reducing the underestimation error of high-risk categories such as bridge damage and severe road damage.

[0055]

[0056] To verify the rationality of the model structure selection in this invention, ResNet50, MobileNetV3-Large, EfficientNet-B0, YOLOv8n-cls, and MSDANeXt were compared on a fixed real test set. All models used the same real training, validation, and testing partitions, and various metrics were calculated on the real test set.

[0057]

[0058] As shown in Table 6, MSDANeXt achieved the highest results in Accuracy, Balanced Accuracy, and Macro-F1, at 91.73%, 90.50%, and 90.40%, respectively. Meanwhile, its high-risk underestimation index was only 18.39%, significantly lower than other comparative models. These results indicate that MSDANeXt achieves a good balance between overall classification performance and secure identification capability, making it suitable as the preferred identification model for this invention.

[0059] This embodiment employs a training approach combining synthetic pre-training, real-world fine-tuning, and low-light enhancement. First, pre-training is performed on a synthetic training set, enabling the model to learn basic damage characteristics such as road cracking, collapse, misalignment, bulges, slope accumulation, and anomalies in bridge decks, beams, supports, and piers. Then, fine-tuning is performed on a real training set to adapt the model to the textures, backgrounds, and acquisition noise in real monitoring images. During the real-image fine-tuning stage, low-light enhancement samples are introduced simultaneously. Low-light enhancement does not regenerate Blender data; instead, it generates additional image-level perturbation samples based on the original training images, including low exposure, gamma darkening, low-light noise, low-light blur compression, and local shadows. The validation and test sets remain constant to ensure the fairness of the evaluation results.

[0060] To evaluate the robustness of the model under the low-light enhancement training strategy, this embodiment constructs various perturbation test scenarios, including low brightness, high brightness, low contrast, high contrast, Gaussian noise, Gaussian blur, JPEG compression, and composite low-light perturbations. Among them, the composite low-light perturbations include Gamma low light, low light noise, low light blur compression, and shadow low light, used to simulate nighttime surveillance and low-quality video images after disasters.

[0061] The test results, as shown in Figure 6 and Table 6, demonstrate that after employing the synthetic pre-training-realistic fine-tuning + low-light enhancement strategy, the model achieved an average Macro-F1 score of 92.66% across all perturbation test sets. Regarding high-risk underestimation, the model's high-risk underestimation index on the original test set was 9.82%, a significant improvement compared to the baseline model's 17.14%; and on the low-light perturbation test set, the high-risk underestimation index was 9.79%, a significant improvement compared to the baseline model's index of 20.36% on the general test set. These results indicate that synthetic pre-training can supplement prior knowledge of road and bridge damage morphology and scene changes, while low-light enhancement can supplement degraded samples in low-light imaging. The combination of these two methods maintains high recognition stability under complex monitoring imaging conditions and reduces the risk of underestimating severe damage.

[0062]

[0063] In summary, this embodiment demonstrates that the present invention can utilize traffic monitoring images to uniformly identify the post-earthquake state of roads and bridges, and map the identification results into traffic capacity information. Through real-synthetic joint data, synthetic pre-training-real fine-tuning, and low-light enhancement training strategies, the model can maintain stable identification capabilities under complex imaging conditions, making it suitable for rapid investigation of post-earthquake road traffic infrastructure, auxiliary judgment of traffic restrictions and closures, and emergency rescue route formulation.

[0064] The embodiments of the present invention disclosed above are merely illustrative of the invention. These embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A rapid assessment method for seismic damage and traffic capacity of roads and bridges based on traffic monitoring images, characterized in that, Includes the following steps: S1. Based on the type and degree of road and bridge damage, establish a mapping relationship between seismic damage and traffic capacity: Level I road damage corresponds to normal traffic, Level II road damage corresponds to restricted traffic, and Level III road damage corresponds to prohibited traffic; intact bridges correspond to normal traffic, and damaged bridges correspond to prohibited traffic. S2. Collect real images of road and bridge damage after the earthquake, and label the images according to the classification rules in S1 to construct a real image dataset. S3. Construct a three-dimensional model of post-earthquake damage to roads and bridges, simulate different monitoring scenarios through domain randomization, and generate a synthetic image dataset; S4. Construct a joint dataset containing the real image dataset and the synthetic image dataset, and divide it into a training set, a validation set, and a test set; S5. Establish a rapid assessment model for earthquake damage and traffic capacity of roads and bridges. Use the joint dataset of S4 to train and evaluate the model. Input the monitoring image to be evaluated into the trained model and output the damage type and degree of the image, which corresponds to the damage classification of S1. Then, output the traffic status recognition result according to the mapping relationship of S1.

2. The method for rapid assessment of road and bridge seismic damage and traffic capacity based on traffic monitoring images according to claim 1, characterized in that: In S1, the road damage type includes at least one of road cracking, road collapse, road misalignment, road uplift, and slope collapse accumulation; the bridge damage type includes at least one of bridge deck cracking, beam end displacement, beam misalignment, bearing abnormality, pier cracking and spalling, pier tilting, and local collapse.

3. The method for rapid assessment of road and bridge seismic damage and traffic capacity based on traffic monitoring images according to claim 2, characterized in that: In S1, road damage level I is basically intact or slightly damaged, road damage level II is moderately damaged, and road damage level III is severely damaged or destroyed; the criteria for judging the integrity of the bridge are that the bridge deck is smooth, the positional relationship between the beam and the pier is normal, and there is no obvious damage to the railings and ancillary facilities; the criteria for judging the damage of the bridge are that there are visible structural abnormalities in the main structure or ancillary facilities of the bridge.

4. The method for rapid assessment of road and bridge seismic damage and traffic capacity based on traffic monitoring images according to claim 3, characterized in that: The criteria for determining whether the slope is basically intact or slightly damaged are: crack width < 5cm, collapse diameter < 3cm, misalignment height < 3cm, bulge height < 3cm, or the slope does not intrude into the driving lane. The criteria for determining moderate damage are: crack width of 5-10cm, collapse diameter of 3-15cm, misalignment height of 3-15cm, uplift height of 3-15cm, or slope intrusion into the road surface but with a remaining continuous passable strip width ≥3.5m. The criteria for determining severe damage or destruction are: crack width > 10cm, collapse diameter > 15cm, misalignment height > 15cm, bulge height > 15cm, or slope intrusion into the road surface resulting in a continuous passable strip width < 3.5m.

5. The method for rapid assessment of road and bridge seismic damage and traffic capacity based on traffic monitoring images according to claim 4, characterized in that: In S2, when multiple types of road damage coexist in an image, the comprehensive damage level of the image is determined according to the maximum level criterion; when a single image is insufficient to clearly determine the safety status of a bridge, the safety priority principle is adopted, and suspicious structural anomaly samples are classified into the bridge damage category.

6. The method for rapid assessment of road and bridge seismic damage and traffic capacity based on traffic monitoring images according to claim 5, characterized in that: S3 specifically includes the following steps: S31. Establish the basic geometric model of the road and bridge, and assign materials to them; S32. Construct road damage models with different types of road damage and bridge damage models with different types of bridge damage on the basic geometric model. S33. Introduce domain randomization parameters to batch render and generate synthetic image datasets.

7. The method for rapid assessment of road and bridge seismic damage and traffic capacity based on traffic monitoring images according to claim 1, characterized in that: In S33, the domain randomization parameters include at least camera position and attitude randomization, time and lighting condition randomization, weather condition randomization, object material randomization, background image randomization, and foreground interference randomization.

8. The method for rapid assessment of road and bridge seismic damage and traffic capacity based on traffic monitoring images according to claim 1, characterized in that: In S4, the joint dataset is constructed by labeling real and synthetic images according to the damage classification in S1, so that they are all converted into five image-level labels: bridge damage, bridge intact, road severe damage, road moderate damage, and road minor damage. Then, it is divided into training set, validation set, and test set.

9. The method for rapid assessment of road and bridge seismic damage and traffic capacity based on traffic monitoring images according to claim 1, characterized in that: In S5, the model uses ConvNeXtV2 as the backbone network and introduces a multi-scale damage perception attention module and a multi-scale damage feature fusion classification head.

10. The method for rapid assessment of road and bridge seismic damage and traffic capacity based on traffic monitoring images according to claim 1, characterized in that: In S5, the model adopts a transfer learning strategy of pre-training with synthetic data and fine-tuning with real data, and adds low-light enhancement samples during the training phase.

Images