Vehicle structure accident damage detection method and system based on visual recognition

Abstract

The application discloses a vehicle structure accident damage detection method and system based on visual recognition and relates to the technical field of image recognition.The vehicle structure accident damage detection method and system based on visual recognition comprises the following steps: S1, acquiring structure accident damage detection data and performing vehicle model matching binding, time sequence alignment, field extraction and standardization processing; S2, performing structure positioning and bearing analysis based on the structure accident damage detection data to form a candidate area and a structure damage representation; S3, performing cross-view homologous analysis on the effective structure candidate area to form a damage evidence group and a repeated damage suppression result; and S4, performing structure accident damage determination and evidence chain archiving.The problems that the existing detection methods are difficult to cross-view collect the same structure damage representation in multi-angle images, cause the same damage to be repeatedly counted or associated damage to be split, and affect the closed-loop determination of structure accident damage evidence are solved.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, specifically to a method and system for detecting accident damage to vehicle structural components based on visual recognition. Background Technology

[0002] With the rapid popularization of new energy vehicles, the continuous expansion of the used car market, and the increasing digitalization of scenarios such as auto finance, insurance claims settlement, and after-sales service, vehicle quality assessment has gradually evolved from traditional offline manual inspection to intelligent, remote, and data-driven approaches. Vehicle quality status not only affects transaction pricing and consumer decisions but also relates to risk control by financial institutions, claims settlement by insurance companies, and quality traceability by OEMs. The existing auto appraisal industry has begun to introduce technologies such as artificial intelligence, computer vision, big data analytics, and industry knowledge models to assist in identifying vehicle appearance, historical repair history, accident risks, and value trends. At the same time, vehicle structures are becoming increasingly complex, with new energy vehicles, intelligent connected vehicles, and vehicles of various models and years coexisting, making vehicle appraisal objects non-standardized, differentiated, and dynamic. Traditional assessment methods relying on human experience, single inspections, or single information sources are no longer sufficient to meet the requirements of efficiency, accuracy, consistency, and traceability in large-scale vehicle circulation scenarios. Therefore, technologies related to vehicle image recognition, intelligent vehicle condition analysis, accident risk assessment assistance, and reliable output of assessment results have become important technological directions for the intelligent upgrading of the automotive aftermarket and the high-quality development of the automotive industry.

[0003] For example, the invention patent with announcement number CN114359717B discloses a vehicle damage recognition method based on multi-view correlation deep learning. The method includes: acquiring labeled distant and close-up images of vehicle damage; extracting distant and close-up image features using a Siamese deep convolutional neural network; obtaining distant vehicle part region proposal boxes and close-up vehicle damage region proposal boxes through a region proposal network, and unifying the image feature scale using ROI Align; constructing a relationship module to extract the relationship features between part regions and damage regions, and fusing the relationship features with image features; finally, performing damage segmentation, classification, and regression based on the fused features to obtain the damage category, part category, and damage region bounding box coordinates, thus realizing the correlation recognition between distant vehicle part regions and close-up vehicle damage regions.

[0004] For example, the invention patent with announcement number CN115115611B discloses a vehicle damage identification method, device, electronic device, and storage medium. The method includes: dividing the overall appearance of the target vehicle into N predetermined blocks, and acquiring N original images according to an image acquisition model; performing vehicle component identification on each original image to obtain the vehicle component location identification result; then cutting each original image into M sub-images according to a cutting model, and performing damage identification on the original images and sub-images respectively; subsequently performing coordinate transformation on the local damage identification result and fusing it with the overall damage identification result, and finally combining the vehicle component location identification result to obtain the vehicle component damage result of the target vehicle, thus realizing vehicle appearance partition acquisition, local damage identification, and vehicle component damage fusion output.

[0005] Existing vehicle damage visual recognition technologies can identify vehicle exterior damage, part categories, and damage types through multi-view images, vehicle exterior zoning, region proposal boxes, image segmentation, and component location fusion. However, their focus remains on damage detection and result fusion under a single image or single viewpoint. For multi-view images in remote vehicle inspection caused by differences in shooting angle, shooting distance, lighting conditions, and visible range, existing methods lack a cross-view homogeneous aggregation mechanism for the same structural component. This makes it difficult to determine whether damage representations in different images point to the same structural component, easily leading to duplicate counting of the same damage or fragmentation of related damage. Furthermore, insufficient closed-loop verification between visual recognition results and vehicle model structural benchmarks, maintenance records, and accident records results in an incomplete chain of evidence for structural component accident damage, affecting the stability and traceability of accident risk assessment.

[0006] Therefore, in order to address the above problems, there is an urgent need for a visual recognition-based method and system for detecting accident damage to vehicle structural components. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method and system for detecting accident damage to vehicle structural components based on visual recognition. This solves the problem that existing detection methods struggle to collect damage representations of the same structural component across multiple perspectives in multi-angle images, leading to repeated counting of the same damage or fragmentation of related damages, which affects the closed-loop determination of evidence related to structural component accident damage.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a vehicle structural component accident damage detection method based on visual recognition, comprising: S1, acquiring structural component accident damage detection data from a vehicle detection task and performing vehicle model matching and binding, temporal alignment, field extraction, and standardization; S2, performing structural component location and connection analysis based on the structural component accident damage detection data, and forming effective structural component candidate regions and structural component damage characterization based on the results of the structural component location and connection analysis; S3, performing cross-perspective homology analysis on the effective structural component candidate regions, and forming damage evidence groups and repeated damage suppression results based on the results of the cross-perspective homology analysis; S4, performing accident closed-loop credibility analysis on the damage evidence groups, and determining structural component accident damage and archiving the evidence chain based on the results of the accident closed-loop credibility analysis.

[0009] Furthermore, the specific process for obtaining structural component accident damage detection data in vehicle inspection tasks is as follows: Under the same vehicle inspection task, structural component accident damage detection data is obtained through a remote shooting terminal, a surround-view video acquisition terminal, a vehicle business database, and a vehicle model structure database. This data includes: obtaining vehicle identification code, brand and model name, vehicle series and year, multi-view images of the vehicle, image timestamp, image width, image height, focal length, exposure time, ISO, shooting distance, shooting orientation label, and vehicle part shooting label through the remote shooting terminal; obtaining surround-view video files, video frame timestamps, and video frame sequences through the surround-view video acquisition terminal; obtaining repair record text, insurance claim record text, inspection report text, and record generation time through the vehicle business database; and retrieving the corresponding structural component name table, structural component reference orientation table, structural component reference outline diagram, structural component reference position coordinates, structural component reference projected area table, and structural component adjacency relationship table from the vehicle model structure database.

[0010] Further, the specific process of vehicle model matching and binding, time sequence alignment, field extraction, and standardization is as follows: Vehicle model matching and binding is performed based on the vehicle identification code, brand and model name, and vehicle year; video frame extraction processing is performed on the surround-view video file to obtain surround-view video frame images; the vehicle multi-view images, surround-view video frame images, image timestamps, video frame timestamps, and video frame sequences are time-bound for the same task; the location field, repair action field, and accident direction field in the repair record text, insurance claim record text, and inspection report text are extracted using named entity recognition methods; the structural component accident damage detection data are standardized using a zero-mean unit variance standardization algorithm; according to the acquisition order of the same vehicle inspection task, adjacent differences are calculated for the image timestamp, video frame timestamp, and video frame sequence number to obtain the image timestamp difference, video frame timestamp difference, and video frame sequence number difference, which are then normalized using a maximum-minimum-value normalization algorithm.

[0011] Furthermore, the specific process of structural component location and acceptance analysis based on structural component accident damage detection data is as follows: Based on the structural component name table, structural component reference contour diagram, structural component reference position coordinates, and structural component reference projected area table, candidate box localization, mask segmentation, and reference constraint matching are performed on the vehicle multi-view images and surround-view video frame images, and the order of generation is determined. The first structural component candidate region; inputting multi-view vehicle images, surround-view video frame images, structural component name list, and structural component baseline contour map into the object detection model and semantic segmentation model for structural component region matching, outputting the region matching quantity; obtaining the shooting consistency quantity by mapping the shooting orientation label, vehicle part shooting label, and structural component baseline orientation table; for the first... The number of pixels in the candidate region mask corresponding to each structural component candidate region is counted to obtain the candidate region mask area. The area carrying capacity is calculated by comparing the candidate region mask area with the projected area of ​​the corresponding structural component in the structural component reference projection area table. The occlusion and reflection amount is obtained by detecting reflective areas and occlusion areas in multi-view images of the vehicle and circumferential video frames. The acquisition adaptation amount is obtained by mapping the image width, image height, focal length, exposure time, sensitivity, and shooting distance values ​​through the acquisition condition interval. The area carrying capacity is calculated by summing the area carrying capacity and the numerical protection constant and taking the square root to obtain the area carrying capacity root value. The product of the region matching amount, the shooting consistency amount, and the area carrying capacity root value is calculated and then added to obtain the positioning enhancement term. The occlusion and reflection amount is calculated and then added to obtain the occlusion suppression term. The ratio of the positioning enhancement term to the occlusion suppression term is calculated and the arctangent is taken to obtain the angle compression term. The negative number of the acquisition adaptation amount is calculated and the exponent is taken, and then the exponent result is subtracted by one to obtain the acquisition adaptation term. The product of the angle compression term and the acquisition adaptation term is calculated to obtain the positioning carrying capacity value.

[0012] Furthermore, the specific process for forming effective candidate regions for structural components and structural component damage characterization based on the structural component location and acceptance analysis results is as follows: Real-time comparison of location acceptance values ​​and location acceptance thresholds: When the location acceptance value is greater than or equal to the location acceptance threshold, the corresponding region is marked as an effective candidate region for structural components. Based on the structural component name table, structural component reference contour map, structural component reference position coordinates, structural component reference orientation table, shooting orientation labels, and vehicle part shooting labels, the semantic labels and vehicle body orientation labels for structural components are determined. The effective candidate regions for structural components are processed by the scale-invariant feature transformation algorithm, edge contour extraction processing, and structural component damage characterization detection processing to generate structural component location anchor points and structural component damage characterization sets for cross-view homology analysis. When the location acceptance value is less than the location acceptance threshold, the region is marked as a candidate region for structural components to be supplemented. Based on the corresponding shooting orientation labels, vehicle part shooting labels, and structural component reference orientation table, the missing viewpoint set and the shooting parts to be supplemented are determined, and a supplementary shooting task is generated. Before the supplementary shooting is completed, the region is only used as auxiliary evidence to be verified.

[0013] Furthermore, the specific process of performing cross-perspective source analysis on the candidate regions of effective structural components is as follows: taking the first... The first effective structural component candidate region and the first Each valid candidate structural component region constitutes a pair of regions to be matched. Within each pair, the semantic labels of structural components are matched using the structural component name table and the structural component adjacency table to obtain the semantic consistency quantity. The anchor points of structural components are matched using a scale-invariant feature transformation algorithm, and the matching quantity of the anchor points is obtained through homography estimation. The damage representation similarity of the structural component damage representation set is calculated using cosine similarity. The vehicle body orientation deviation is obtained by mapping the shooting orientation label, vehicle part shooting label, body orientation label, and structural component reference orientation table using orientation difference mapping. The image width, image height, focal length, exposure time, ISO, shooting distance, image timestamp difference, video frame timestamp difference, and video frame... The sequence difference is used to obtain the image acquisition condition difference through Chebyshev distance mapping; the occlusion conflict is calculated by the absolute value of the difference for the occlusion reflection amount; the product of the location acceptance value of the a-th effective structural component candidate region and the location acceptance value of the b-th effective structural component candidate region is calculated and the square root is taken to obtain the dual-region location acceptance root value; the product of the structural component semantic consistency amount, the structural component location anchor point matching amount, the damage representation similarity amount, and the dual-region location acceptance root value is calculated to obtain the cross-view homology support term; the sum of the vehicle body orientation deviation amount, the image acquisition condition difference amount, and the occlusion conflict amount is calculated and added to obtain the cross-view deviation suppression term; the ratio of the cross-view homology support term to the cross-view deviation suppression term is calculated and the hyperbolic tangent is taken to obtain the cross-view homology value.

[0014] Furthermore, the specific process for forming damage evidence sets and repeated damage suppression results based on cross-perspective homology analysis results is as follows: real-time comparison of cross-perspective homology values ​​and cross-perspective homology thresholds: when the cross-perspective homology value is greater than or equal to the cross-perspective homology threshold, the first... The first effective structural component candidate region and the first The candidate regions of each valid structural component are grouped into the same damage evidence group, and a cross-view aggregation relationship is generated. For repeated damage characterization regions within the same damage evidence group, which are determined by region overlap and characterization similarity verification, a repeated damage suppression label is generated. When the cross-view homology value is less than the cross-view homology threshold, the first... The first effective structural component candidate region and the first The independent attribution relationship of each valid structural component candidate region is defined. When no cross-view aggregation relationship is formed for the same valid structural component candidate region, it is marked as isolated from a single view.

[0015] Furthermore, the specific process of conducting a closed-loop credibility analysis of the damage evidence set is as follows: In the first... Within each damage evidence group, the cross-view support quantity is obtained by deduplicating the number of candidate regions for different shooting orientation labels, vehicle part shooting labels, and effective structural components. The structural component damage intensity is obtained by calculating the structural component damage characterization set through probability union. The repair and accident record support quantity is obtained by matching the part field, repair action field, accident direction field, and record generation time in the repair record text, insurance claim record text, and inspection report text with the structural component semantic label and vehicle body orientation label through semantic matching methods. The boundary of the candidate region mask corresponding to the effective structural component candidate region is extracted and the mask center point is calculated to obtain the candidate region mask boundary and candidate region center coordinates. The contour difference and position offset are calculated with the structural component reference contour map and the structural component reference position coordinates to obtain the vehicle model reference conflict quantity. The occlusion reflection amount, the image acquisition mismatch amount obtained by subtracting the acquisition adaptation amount from one, and the amount of occlusion reflection, the amount of image acquisition mismatch obtained by subtracting the acquisition adaptation amount from one, and the amount of image acquisition mismatch obtained by subtracting the acquisition adaptation amount from one are also considered. The proportion of candidate structural components to be supplemented matching each damage evidence group is merged to obtain the insufficient image evidence quantity; the ratio of the number of valid structural component candidate regions with single-view isolation markers to the total number of valid structural component candidate regions is calculated to obtain the single-view isolation quantity; the arithmetic mean of cross-view homologous values ​​is taken to obtain the homologous mean of the evidence group; the exponential result of the support quantity of repair accident records is calculated to obtain the repair accident support amplification term; the product of the cross-view support quantity plus one, the structural component damage intensity plus one, and the repair accident support amplification term is calculated to obtain the accident closed-loop support term; the sum of the vehicle model benchmark conflict quantity, the insufficient image evidence quantity, and the single-view isolation quantity is calculated and then added to one to obtain the accident closed-loop suppression term; the ratio of the accident closed-loop support term to the accident closed-loop suppression term is calculated, added to one, and then the natural logarithm is taken to obtain the closed-loop evidence logarithm term; the negative of the homologous mean of the evidence group is calculated, the exponent is taken, and then the exponent is subtracted by one to obtain the homologous mean inheritance term; the product of the closed-loop evidence logarithm term and the homologous mean inheritance term is calculated to obtain the accident closed-loop confidence value.

[0016] Furthermore, the specific process for determining structural component accident damage and archiving the evidence chain based on the accident closed-loop credibility analysis results is as follows: Real-time comparison of the accident closed-loop credibility value and the accident closed-loop credibility threshold: When the accident closed-loop credibility value is greater than or equal to the accident closed-loop credibility threshold, the corresponding structural component accident damage is determined to be established, and the structural component accident damage detection result is output; at the same time, the structural component accident damage evidence chain is generated, created, and archived in the vehicle quality assessment database; when the accident closed-loop credibility value is less than the accident closed-loop credibility threshold, the conclusion that the structural component accident damage is established is not directly output, the corresponding damage evidence group is marked as the result of insufficient accident damage evidence, and a manual review task is generated based on the amount of insufficient image evidence, the amount of isolated single-view evidence, and the candidate areas of structural components to be supplemented; after the review is completed, the review conclusion is resubmitted to the structural component positioning and acceptance analysis.

[0017] The second aspect of this invention provides a vehicle structural component accident damage detection system based on visual recognition, comprising: a data acquisition and preprocessing module for acquiring structural component accident damage detection data in a vehicle inspection task and performing vehicle model matching and binding, time sequence alignment, field extraction, and standardization processing; a structural component candidate region identification module for performing structural component location and connection analysis based on the structural component accident damage detection data, and forming effective structural component candidate regions and structural component damage characterizations based on the structural component location and connection analysis results; a cross-view homogeneous aggregation module for performing cross-view homogeneous analysis on the effective structural component candidate regions, and forming damage evidence groups and repeated damage suppression results based on the cross-view homogeneous analysis results; and an accident damage closed-loop determination and output module for performing accident closed-loop credibility analysis on the damage evidence groups, and determining structural component accident damage and archiving the evidence chain based on the accident closed-loop credibility analysis results.

[0018] The present invention has the following beneficial effects:

[0019] (1) This invention performs candidate box localization, mask segmentation and structural component localization analysis on multi-view images of vehicles and surround video frame images, so that the visual recognition process can be extended from ordinary appearance damage recognition to vehicle structural component area recognition, thereby realizing the effect of visual localization and intelligent recognition of vehicle structural component accident damage, effectively solving the problem that in the prior art, vehicle visual detection is mainly limited to appearance damage recognition and it is difficult to accurately locate the structural component accident damage area.

[0020] (2) In this invention, the candidate regions of structural components are matched by a structural component name table, a structural component baseline outline diagram, a structural component baseline position coordinate table and a structural component baseline orientation table. This enables the structural component regions under different vehicle models to be verified according to the vehicle model structure data, thereby achieving the effect of accurate acceptance of structural component regions under different vehicle models. This effectively solves the problem that structural component misidentification or omission is easily caused by structural differences between different vehicle models in the prior art.

[0021] (3) In this invention, cross-perspective homology analysis is performed by using semantic tags of structural components, vehicle body orientation tags, structural component positioning anchor points and structural component damage characterization sets. Damage characterizations pointing to the same structural component under different shooting angles are grouped into the same damage evidence group, thereby achieving the effect of unified attribution of damage characterizations of the same structural component in multi-angle images. This effectively solves the problem that the same damage is easily counted repeatedly or related damages are separated in the prior art.

[0022] (4) In this invention, the homology relationship between candidate regions of effective structural components is determined by cross-view homology values, and a repeated damage suppression mark is generated for repeated damage characterization regions, thereby achieving the effect of deduplication and consistent aggregation of structural component damage evidence, effectively solving the problems of repeated superposition of multi-view vehicle image recognition results and inconsistent accident damage evidence in the prior art.

[0023] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0024] Figure 1 This is a flowchart of the vehicle structural component accident damage detection method based on visual recognition according to the present invention;

[0025] Figure 2 This is a structural diagram of the vehicle structural component accident damage detection system based on vision recognition according to the present invention;

[0026] Figure 3 This is a comparison chart of the positioning and bearing values ​​and the amount of light reflection blocked in this invention;

[0027] Figure 4 The thermal diagram shows the positioning and bearing parameters of the structural components of this invention.

[0028] Figure 5 This is a flowchart of the cross-perspective homogeneity grouping determination process of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Please see Figures 1-5 This invention provides a technical solution: a vehicle structural component accident damage detection method based on visual recognition, comprising the following steps: S1, acquiring structural component accident damage detection data from a vehicle detection task and performing vehicle model matching and binding, time sequence alignment, field extraction, and standardization; S2, performing structural component location and connection analysis based on the structural component accident damage detection data, and forming effective structural component candidate regions and structural component damage characterization based on the structural component location and connection analysis results; S3, performing cross-perspective homology analysis on the effective structural component candidate regions, and forming damage evidence groups and repeated damage suppression results based on the cross-perspective homology analysis results; S4, performing accident closed-loop credibility analysis on the damage evidence groups, and determining structural component accident damage and archiving the evidence chain based on the accident closed-loop credibility analysis results.

[0031] Specifically, the process of obtaining structural component accident damage detection data in vehicle inspection tasks is as follows: Under the same vehicle inspection task, structural component accident damage detection data is obtained through a remote shooting terminal, a vehicle surround video acquisition terminal, a vehicle business database, and a vehicle model structure database. This data includes: vehicle identification number, brand and model name, vehicle series and year, multi-view images of the vehicle, image timestamp, image width, image height, focal length, exposure time, ISO, shooting distance, shooting orientation label, and vehicle part shooting label obtained through the remote shooting terminal; among these, the multi-view images include left front side images, right front side images, engine compartment images, left front wheel arch images, right front wheel arch images, sill area images, trunk interior images, and partial detail images of structural components, along with image width, image height, focal length, exposure time, and other parameters. The sensitivity is synchronously written by the remote shooting terminal during image acquisition; the shooting orientation label is automatically written by the remote shooting terminal when shooting is triggered; the vehicle part shooting label is automatically written by the shooting guidance program according to the current shooting task; the shooting distance value is output by the distance measurement component of the remote shooting terminal; the circumferential vehicle video file, video frame timestamp, and video frame sequence number are obtained through the circumferential vehicle video acquisition terminal; the maintenance record text, insurance claim record text, inspection report text, and record generation time are obtained through the vehicle business database; the corresponding structural component name table, structural component reference orientation table, structural component reference outline diagram, structural component reference position coordinates, structural component reference projected area table, and structural component adjacency relationship table are retrieved through the vehicle model structure database. The structural component reference position coordinates and structural component reference projected area table maintain the same projection aperture under the same shooting orientation as the structural component reference outline diagram.

[0032] This implementation plan establishes a data foundation for structural component accident damage detection under the same vehicle inspection task, enabling the unified aggregation of vehicle identity information, image evidence, video evidence, business text evidence, and vehicle model structural benchmark data. This ensures that subsequent identification of candidate areas for structural components has clear vehicle objects, shooting sources, time sources, text sources, and vehicle model benchmark sources, improving the evidence integrity, task consistency, and process traceability of multi-view damage identification. It also provides original data support for the progressive calculation of location acceptance values, cross-view homogeneous values, and accident closed-loop credibility values.

[0033] Specifically, the process of vehicle model matching and binding, time sequence alignment, field extraction, and standardization is as follows: Vehicle model matching and binding is performed based on the vehicle identification code, brand and model name, and vehicle year. Specifically, the vehicle identification code is used as the priority matching field, and the brand and model name and vehicle year are used as vehicle model verification fields. The corresponding vehicle structure data is determined in the vehicle structure database, ensuring that multi-view images, surround-view video files, maintenance record texts, insurance claim record texts, inspection report texts, structural component name tables, structural component reference orientation tables, structural component reference outline diagrams, structural component reference position coordinates, structural component reference projected area tables, and structural component adjacency relationship tables belong to the same vehicle inspection task. Video frames are extracted from the surround-view video files to obtain surround-view video frame images, and the corresponding video frame timestamps and video frame sequences are retained. The multi-view images, surround-view video frame images, image timestamps, video frame timestamps, and video frame sequences are time-bound according to the same task, based on the timestamps. The rule of minimizing adjacent differences and ensuring continuous video frame numbers establishes the correspondence between static images and video frames, maintaining a traceable order. Named entity recognition is used to extract the location, repair action, and accident direction fields from maintenance record text, insurance claim record text, and inspection report text, converting unstructured business text into field-based records suitable for subsequent semantic matching. Zero-mean unit variance standardization is used to standardize structural component accident damage detection data, mitigating dimensional differences caused by different shooting equipment and conditions. Adjacent differences are calculated for image timestamps, video frame timestamps, and video frame numbers according to the acquisition order of the same vehicle inspection task, yielding image timestamp differences, video frame timestamp differences, and video frame number differences. These differences are then normalized using a maximum-minimum value normalization algorithm, ensuring that the continuous parameters used for subsequent location acceptance values, cross-view source values, and accident closed-loop reliability values ​​are dimensionless.

[0034] In this implementation plan, front-end data acquisition, vehicle surround video, business text, and vehicle structure data are organized into a unified data representation under the same vehicle inspection task. This ensures that static images, video frame images, maintenance record text, insurance claim record text, inspection report text, and vehicle structure benchmark data have consistent vehicle identification, time sequence, field-based semantic foundation, and numerical scale foundation. Video frame extraction and time binding ensure that multi-source visual evidence can form a traceable sequence according to the acquisition process. Text field processing ensures that business records can participate in subsequent structural component semantic matching. Standardization and normalization processing ensure that continuous parameters have dimensionless comparability, avoiding the impact of equipment differences, shooting condition differences, and recording format differences on subsequent judgment results. This provides standardized, structured, and traceable data support for the stable calculation of location acceptance values, cross-view homogeneous values, and accident closed-loop credibility values.

[0035] Specifically, the process of structural component location and acceptance analysis based on structural component accident damage detection data is as follows: Based on the structural component name table, structural component reference contour map, structural component reference position coordinates, and structural component reference projected area table, candidate box localization, mask segmentation, and reference constraint matching are performed on multi-view vehicle images and surround-view video frame images. Candidate box localization uses the structural component reference position coordinates to limit the search range; mask segmentation uses the candidate box boundaries as segmentation constraints; and reference constraint matching uses the structural component reference contour map and structural component reference projected area table as matching references, and determines the first [item name] according to the generation order. Candidate regions for structural components.

[0036] Vehicle multi-view images, surround-view video frames, structural component name tables, and structural component baseline contour maps are input into the object detection model and semantic segmentation model for structural component region matching. The object detection model outputs candidate box confidence results, and the semantic segmentation model outputs mask confidence results. Consistency fusion is performed on the candidate box confidence results, mask confidence results, and the degree of overlap between the candidate region mask and the structural component baseline contour map to output the region matching quantity. The shooting orientation labels, vehicle part shooting labels, and structural component baseline orientation tables are mapped using label consistency to obtain the shooting consistency quantity. The label consistency mapping is encoded and matched according to the orientation correspondence between the shooting orientation labels, vehicle part shooting labels, and structural component baseline orientation tables. The number of pixels in the candidate region mask corresponding to the i-th structural component candidate region is counted to obtain the candidate region. The mask area is calculated by comparing the candidate region mask area with the corresponding structural component's projected area in the structural component reference projection area table, and the area ratio is used to characterize the degree of scale correspondence between the candidate region mask and the corresponding structural component's reference projected area. For multi-view vehicle images and surround-view video frames, the amount of occlusion and reflection is obtained through reflective area detection and occlusion area detection. Reflective area detection is based on brightness saturation areas and bright continuous areas, while occlusion area detection is based on non-structural component occlusion blocks and structural component visible boundary interruption areas. The image width, image height, focal length, exposure time, sensitivity, and shooting distance are mapped to a unified acquisition adaptation range through acquisition condition interval mapping. This mapping maps imaging parameters from different shooting devices and shooting distances to a unified acquisition adaptation range.

[0037] The sum of the area acceptance quantity and the numerical protection constant is calculated, and the square root is taken to obtain the root value of the area acceptance quantity. The square root is used to compress the amplitude differences in the area acceptance quantity, preventing area variations from dominating the positioning acceptance value. The product of the region matching quantity, the shooting consistency quantity, and the root value of the area acceptance quantity is calculated, and then one is added to obtain the positioning enhancement term. The product relationship preserves the synergistic enhancement effect between region matching, shooting consistency, and area acceptance. The addition of one is used to maintain the positive reference of the positioning enhancement term. The amount of occlusion reflection is calculated and one is added to obtain the occlusion suppression term. The addition of one avoids the denominator imbalance when the amount of occlusion reflection is zero, and ensures that the amount of occlusion reflection participates in the calculation in a suppressed form. The positioning enhancement term and occlusion... The ratio of the suppression term is taken and its arctangent is used to obtain the angle compression term. Bounded compression is then performed on the comparison result using the arctangent function to reduce the disturbance of abnormal extreme values ​​on the positioning acceptance value. The negative number of the acquisition adaptation quantity is calculated and its exponent is taken. Subtracting the exponent from one yields the acquisition adaptation term. A negative exponent mapping transforms the acquisition adaptation quantity into a monotonically increasing adaptation acceptance factor. The product of the angle compression term and the acquisition adaptation term is calculated to obtain the positioning acceptance value. This product structure ensures that the structural component positioning result is simultaneously constrained by positioning evidence and acquisition conditions. Furthermore, the region matching quantity, shooting consistency quantity, area acceptance quantity, occlusion reflection quantity, and acquisition adaptation quantity in the formula are all dimensionless quantities. The specific calculation formulas are as follows:

[0038] ;

[0039] In the formula, This represents the location acceptance value, used to determine whether the candidate region of the i-th structural component has the location basis to enter the cross-view homology analysis; The region matching quantity is used to characterize the degree of visual matching between the candidate region and the structural component region; This indicates the consistency of the shooting, used to characterize the degree of correspondence between the shooting label and the reference orientation of the structural component; This represents the area bearing capacity, used to characterize the bearing relationship between the mask area of ​​the candidate region and the reference projected area of ​​the structural component; This indicates the amount of light obstruction and reflection, used to characterize the degree of interference between obstruction and reflection on the visibility of structural components; represents the acquisition adaptation amount, used to characterize the degree of support of image acquisition conditions for structural component localization calculation; i represents the sequence number of the structural component candidate region according to the generation order under the same vehicle detection task. The i-th structural component candidate region is determined by candidate box localization, mask segmentation and benchmark constraint matching, and is used to distinguish different structural component candidate regions in the same vehicle detection task. This represents the numerical protection constant, determined by the lower limit of numerical accuracy after processing the structural component accident damage detection data using a maximum-minimum-value normalization algorithm. Its value range is [value range missing]. This is used to maintain the numerical stability of square root calculations.

[0040] In this embodiment, Table 1 is a data table for calculating the positioning acceptance value of structural components. It details the region matching amount, image consistency amount, area acceptance amount, occlusion and reflection amount, and the final calculated positioning acceptance value for different candidate regions of structural components during the positioning acceptance analysis process. This data is used to quantify the positioning basis for each candidate region of structural components to enter the cross-view common source analysis. Specifically: Candidate region 1 has a region matching amount of 0.95, an image consistency amount of 1.00, an area acceptance amount of 1.20, an occlusion and reflection amount of 0.05, and a positioning acceptance value of 0.963; Candidate region 2 has a region matching amount of 0.80, an image consistency amount of 0.80, an area acceptance amount of 0.50, an occlusion and reflection amount of 0.30, and a positioning acceptance value of 0.539; Candidate region 3 has a region matching amount of 0.60, an image consistency amount of 0.50, an area acceptance amount of 0.20, an occlusion and reflection amount of 0.85, and a positioning acceptance value of 0.14. 4; Candidate region 4 has a region matching value of 0.90, a shooting consistency value of 0.90, an area coverage value of 0.10, an occlusion reflection value of 0.10, and a positioning coverage value of 0.758; Candidate region 5 has a region matching value of 0.70, a shooting consistency value of 0.70, an area coverage value of 0.80, an occlusion reflection value of 0.60, and a positioning coverage value of 0.070; Candidate region 6 has a region matching value of 0.98, a shooting consistency value of 0.95, an area coverage value of 0.05, an occlusion reflection value of 0.02, and a positioning coverage value of 0.582.

[0041] Table 1. Calculation Data for Structural Component Positioning and Bearing Values

[0042] 1 0.95 1.00 1.20 0.05 0.963 2 0.80 0.80 0.50 0.30 0.539 3 0.60 0.50 0.20 0.85 0.144 4 0.90 0.90 0.10 0.10 0.758 5 0.70 0.70 0.80 0.60 0.070 6 0.98 0.95 0.05 0.02 0.582

[0043] like Figure 3 The figure shows a comparison between the positioning acceptance value and the amount of occlusion reflection. Combined with Table 1, it can be seen that there is a clear inhibitory relationship between the positioning acceptance value and the amount of occlusion reflection for different candidate structural component areas. Candidate area 1 has a positioning acceptance value of 0.963 and an occlusion reflection of 0.05, indicating that its structural component area matching, shooting consistency, and visibility conditions can jointly support structural component positioning. Candidate area 3 has an occlusion reflection of 0.85 and a positioning acceptance value of 0.144, showing that occlusion reflection interference significantly inhibits the structural component positioning acceptance result. Although candidate area 5 has an area acceptance of 0.80, its occlusion reflection is 0.60, resulting in a positioning acceptance value of only 0.070, indicating that area acceptance alone cannot directly support the admission of structural component candidate areas. Candidate area 4 has a positioning acceptance value of 0.758, demonstrating the supporting role of area matching and shooting consistency in the positioning acceptance result. Overall, the comparison chart of positioning acceptance value and occlusion reflection amount intuitively reflects the impact of occlusion reflection factors on the admission judgment of structural component candidate areas, and can be used as the basis for screening effective structural component candidate areas and triggering reshoot tasks.

[0044] like Figure 4 The image shown is a heatmap of structural component positioning and bearing parameters, illustrating the distribution of candidate areas for each structural component across four parameters: area matching, image consistency, area bearing capacity, and occlusion / reflection. Figure 4 It can be seen that candidate region 1 has sufficient support in terms of region matching, image consistency, and area acceptance, while the occlusion reflection is 0.05, resulting in a positioning acceptance value of 0.963. Candidate region 3 is insufficient in terms of region matching, image consistency, and area acceptance, while the occlusion reflection is 0.85, resulting in a positioning acceptance value of 0.144. Candidate region 6 has region matching and image consistency of 0.98 and 0.95 respectively, but its area acceptance is only 0.05, indicating that although this region has a good visual matching foundation, the acceptance relationship between the candidate region mask and the reference projected area is insufficient. Candidate region 5 has an area acceptance of 0.80, but the occlusion reflection is 0.60, which suppresses its positioning acceptance value. Overall, the heatmap of structural component positioning acceptance parameters can simultaneously present the differences between multiple parameters, providing a visual basis for judging the formation of positioning acceptance values ​​and identifying candidate regions for supplementary structural components.

[0045] In this implementation scheme, the visible areas of structural components in multi-view images and surround-view video frames are transformed into candidate structural component areas with vehicle model reference constraints. This allows for a unified quantitative expression of the visual matching degree, shooting orientation adaptation degree, scale acceptance degree, visibility interference degree, and acquisition condition support degree of the candidate areas. The positioning acceptance value comprehensively judges the positioning reliability of the candidate structural component areas, filtering out areas with obvious occlusion and reflection, mismatched shooting conditions, and insufficient vehicle model reference acceptance, while retaining effective candidate structural component areas with stable positioning foundations. This provides a clear, calculable, and traceable basis for structural component positioning for subsequent cross-view homogeneous aggregation.

[0046] Specifically, the process of forming effective candidate regions for structural components and characterizing structural component damage based on the structural component location and bearing analysis results is as follows: Real-time comparison of location bearing values ​​and location bearing thresholds:

[0047] When the location acceptance value is greater than or equal to the location acceptance threshold, the corresponding area is marked as a valid structural component candidate area and written into the structural component candidate area set. The corresponding candidate area mask, candidate box coordinates, and source image identifier are retained to support subsequent cross-view matching and evidence tracing. Based on the structural component name table, structural component baseline contour map, structural component baseline position coordinates, structural component baseline orientation table, shooting orientation label, and vehicle part shooting label, semantic labels and vehicle body orientation labels for the structural component are determined. Scale-invariant feature transformation algorithm, edge contour extraction processing, and structural component damage characterization detection processing are performed on the valid structural component candidate areas to generate structural component location anchors and a structural component damage characterization set containing damage characterization areas, damage characterization types, and characterization vectors. Specifically, the valid structural component candidate areas are processed using... Scale-invariant feature transformation algorithm and edge contour extraction processing are used to obtain local key points, edge intersection points, bolt center points, and weld center points. These local key points, edge intersection points, bolt center points, and weld center points are then identified as structural component positioning anchor points. For valid structural component candidate regions, weld region identification, wrinkle edge extraction, cutting and splicing line detection, bolt disassembly and assembly trace detection, edge misalignment detection, and paint coverage detection are performed to obtain a structural component damage characterization set. This set includes the damage characterization region, damage characterization type, and characterization vector corresponding to weld anomalies, local wrinkles, cutting and splicing, bolt disassembly and assembly, edge misalignment, and paint coverage features. This provides a feature basis for subsequent judgment on whether the same structural component has consistent damage characterization under different viewpoints, thus entering cross-viewpoint homology analysis.

[0048] When the location acceptance value is less than the location acceptance threshold, the corresponding area is not used as the main collection object for cross-view same source. Instead, the area is marked as a candidate area for supplementary structural components, and its shooting orientation label, vehicle part shooting label, and source image identifier are retained to avoid the use of insufficient evidence areas directly for accident conclusion determination. Based on the corresponding shooting orientation label, vehicle part shooting label, and structural component reference orientation table, the missing viewpoint set and the shooting parts to be supplemented are determined, and a supplementary shooting task is generated. The supplementary shooting task points to the corresponding structural component orientation and vehicle part to supplement the visibility evidence of the current structural component area. Before the supplementary shooting is completed, the area is only used as auxiliary evidence to be reviewed for subsequent calculation of insufficient image evidence.

[0049] In this implementation plan, by real-time determination of the location acceptance value and the location acceptance threshold, evidence diversion and quality control are performed on the candidate areas of structural components. Areas with stable positioning foundations are included in the set of candidate areas of structural components, forming a set of semantic tags, vehicle body orientation tags, structural component positioning anchors, and structural component damage representations. This provides matching structural component identity, spatial positioning, and damage feature evidence for cross-view homogeneous aggregation. Areas with insufficient positioning foundations are transferred to the supplementary candidate areas of structural components, retaining source information and shooting direction. This avoids areas with insufficient image evidence or insufficient visibility from directly participating in the determination of accident damage conclusions. This ensures that subsequent calculation of insufficient image evidence, generation of reshoot tasks, and manual review have clear data sources and processing basis, improving the reliability, interpretability, and evidence loop integrity of structural component accident damage detection results.

[0050] Specifically, the process of performing cross-perspective homology analysis on candidate regions of effective structural components is as follows: taking the first... The first effective structural component candidate region and the first A pair of valid structural component candidate regions is formed to be matched. These pairs are generated by combining valid structural component candidate regions from different shooting angles according to their region numbers. Within each pair, the semantic labels of structural components are matched using a structural component name table and a structural component adjacency table to obtain a semantic consistency quantity. Consistency matching is hierarchically mapped according to relationships with the same structural component, adjacent structural components, and no structural association. Structural component anchor points are matched using a scale-invariant feature transformation algorithm, and the matching quantity is obtained through homography estimation. Feature point matching is based on the descriptor similarity of the structural component anchor points, and homography estimation is used to eliminate matching point pairs that do not meet geometric consistency. The damage characterization set of structural components is calculated using cosine similarity to obtain a damage characterization similarity quantity. Cosine similarity is based on weld anomalies, local wrinkles, cutting and splicing, and screw... The representation vectors corresponding to bolt disassembly / reassembly, edge misalignment, and paint coverage features are used as the calculation objects. The vehicle body orientation deviation is obtained through orientation difference mapping of the shooting orientation label, vehicle part shooting label, body orientation label, and structural component reference orientation table. The orientation difference mapping is numerically encoded according to the directional deviation relationship between the shooting orientation and the structural component reference orientation. The image acquisition condition difference is obtained through Chebyshev distance mapping of the standardized and normalized image width, image height, focal length, exposure time, ISO, shooting distance, image timestamp difference, video frame timestamp difference, and video frame sequence number difference. The maximum normalized difference among the acquisition condition parameters is used as the region pair difference representation. The occlusion and reflection amount is calculated using the absolute value of the difference to obtain the occlusion conflict amount. The absolute value of the difference is used to measure the inconsistency in the degree of occlusion and reflection interference between the regions to be matched.

[0051] The product of the location acceptance value of the a-th effective structural component candidate region and the location acceptance value of the b-th effective structural component candidate region is calculated and its square root is taken to obtain the dual-region location acceptance root value. The square root is used to compress the amplitude difference between the two location acceptance values, allowing the location basis of the region pair to participate in the calculation in a collaborative manner. The product of the structural component semantic consistency quantity (increased by one), the structural component location anchor point matching quantity (increased by one), the damage representation similarity quantity (increased by one), and the dual-region location acceptance root value is calculated to obtain the cross-view homogeneous support term. The increment is used to retain the positive reference of each support quantity, and the product structure is used to reflect the common support relationship between semantics, anchor points, damage representation, and location acceptance. The vehicle body orientation deviation and image acquisition bar are calculated. The sum of component differences and occlusion conflicts is added to obtain the cross-view deviation suppression term. Adding one is used to avoid denominator imbalance and to ensure that orientation deviation, acquisition differences, and occlusion conflicts participate in the calculation in a suppressed form. The ratio of the cross-view homology support term to the cross-view deviation suppression term is calculated and its hyperbolic tangent is taken to obtain the cross-view homology value. The hyperbolic tangent function is used to perform bounded compression on the homology discrimination results, reducing the disturbance of the results to extreme image differences or local anomaly matching. Furthermore, the component semantic consistency, component positioning anchor point matching, damage characterization similarity, positioning acceptance value, vehicle body orientation deviation, image acquisition condition differences, and occlusion conflict in the formula are all dimensionless quantities. The specific calculation formulas are as follows:

[0052] ;

[0053] In the formula, This indicates a cross-perspective homology value, used to determine whether the regions to be matched can be grouped into the same group of damage evidence. This represents the semantic consistency of structural components, used to characterize the degree of consistency between the regions to be matched in terms of structural component names and structural component adjacency relationships. This indicates the matching amount of the structural component positioning anchor points, used to characterize the degree of matching between the area to be matched and the stable positioning features; This represents the damage characterization similarity measure, used to characterize the degree of similarity between the regions to be matched on the set of structural damage characterizations. This represents the location acceptance value of the a-th valid structural component candidate region, used to characterize the location validity of the a-th valid structural component candidate region; This represents the location acceptance value of the b-th valid structural component candidate region, used to characterize the location validity of the b-th valid structural component candidate region; This indicates the vehicle's orientation deviation, used to characterize the degree of deviation of the area to be matched from the vehicle's orientation. It represents the difference in image acquisition conditions, used to characterize the degree of difference between the regions to be matched in terms of imaging conditions and temporal conditions; The value represents the amount of occlusion conflict, used to characterize the degree of conflict between the regions to be matched in terms of occlusion and reflective interference; a and b represent the indices of the two valid structural candidate regions in the region to be matched, used to distinguish the two region objects participating in cross-view homogeneous analysis under the same vehicle detection task.

[0054] In this implementation plan, valid structural component candidate regions under different shooting angles are transformed into comparable pairs of regions to be matched. This makes structural component semantics, structural component positioning anchor points, structural component damage characterization, positioning support basis, vehicle body orientation differences, image acquisition condition differences, and occlusion and reflection conflicts form a unified basis for homology discrimination. Cross-view homology values ​​can comprehensively measure whether two valid structural component candidate regions point to the same structural component damage, avoiding erroneous aggregation based solely on a single visual similarity result. It also avoids the same structural component damage being split into multiple independent damage evidences due to differences in shooting angle, imaging conditions, or occlusion status. This provides a stable, interpretable, and traceable cross-view aggregation basis for damage evidence group generation, repeated damage suppression marker generation, and subsequent accident closure confidence value calculation.

[0055] Specifically, the process of forming damage evidence sets and repeated damage suppression results based on cross-perspective homology analysis results is as follows: Figure 5 The flowchart shown is for cross-view homology grouping determination, which compares cross-view homology values ​​and cross-view homology thresholds in real time:

[0056] When the cross-view homology value is greater than or equal to the cross-view homology threshold, the a-th and b-th valid structural component candidate regions are grouped into the same damage evidence group, and a cross-view grouping relationship is generated. Specifically, the region numbers of the two valid structural component candidate regions are used as binding objects, and the source image identifier, shooting orientation label, structural component semantic label, vehicle body orientation label, structural component damage representation set, and cross-view homology value are written under the same damage evidence group number to form a correspondence between the damage of the same structural component under different shooting perspectives. The cross-view grouping relationship records the corresponding region number, source image identifier, shooting orientation label, and structural component semantic label to form a multi-view evidence link for the same structural component. For repeated damage representation regions determined by region overlap and representation similarity verification within the same damage evidence group, a repeated damage suppression mark is generated to prevent the damage of the same structural component from being counted repeatedly.

[0057] When the cross-view homogeneity value is less than the cross-view homogeneity threshold, the independent attribution relationship between the a-th and b-th effective structural component candidate regions is maintained, and they are temporarily not written into the same damage evidence group to avoid erroneous merging caused by differences in shooting angle, occlusion differences, or similar textures of non-homogeneous structural components. When the same effective structural component candidate region does not form any cross-view aggregation relationship, it is marked as a single-view isolated region and retained as single-view evidence to be verified for use in the calculation of the single-view isolated quantity in the subsequent accident closed-loop confidence value.

[0058] In this implementation plan, valid candidate regions of structural components from different perspectives are aggregated and adjudicated based on cross-perspective homology values. Regions that reach the cross-perspective homology threshold are grouped into the same damage evidence group. The region number, source image identifier, shooting orientation label, and structural component semantic label are retained through cross-perspective aggregation relationship to establish a multi-perspective evidence link for damage to the same structural component. The duplicate damage suppression marker can prevent the damage to the same structural component from being counted repeatedly in multiple image sources. The independent attribution relationship and single-perspective isolation marker can prevent non-homology regions from being incorrectly merged. The cross-perspective aggregation results take into account the accuracy of merging, the completeness of evidence, and the traceability of verification, providing a clear basis for the calculation of cross-perspective support, single-perspective isolation, and homology mean of evidence groups in the subsequent accident closure credibility value.

[0059] Specifically, the process of conducting a closed-loop credibility analysis of the damage evidence set is as follows: in the first... Within each damage evidence group, deduplication statistics are performed on different shooting orientation labels, vehicle part shooting labels, and the number of valid structural component candidate regions to obtain cross-view support. The deduplication statistics use the source image identifier, shooting orientation label, and valid structural component candidate region number as the joint key to avoid duplicate regions being counted repeatedly within the same viewpoint. The structural component damage intensity is obtained by probabilistic union calculation of the structural component damage characterization set. The probabilistic union calculation fuses the confidence probabilities of each damage characterization output from the structural component damage characterization detection and processing to retain the common support relationships among multiple types of damage characterizations. The location field, repair action field, accident direction field, and record generation time in maintenance record text, insurance claim record text, and inspection report text are matched with the structural component semantic labels and vehicle body orientation labels using semantic matching methods to obtain dimension. The repair accident record support quantity, the semantic matching method uses part correspondence, orientation consistency and time correlation as matching constraints to establish the correspondence between business text evidence and image damage evidence; the boundary extraction and mask center point calculation of the candidate region mask corresponding to the effective structural component candidate region are performed to obtain the candidate region mask boundary and candidate region center coordinates, and the contour difference and position offset calculation are performed with the structural component reference contour map and structural component reference position coordinates respectively. The contour difference and position offset are normalized to obtain the vehicle model reference conflict quantity. The contour difference calculation is used to compare the shape deviation of the mask boundary and the reference contour, and the position offset calculation is used to compare the spatial deviation of the candidate region center coordinates and the reference position coordinates; the occlusion reflection amount, the image acquisition mismatch amount obtained by subtracting the acquisition adaptation amount from one, and the amount of occlusion reflection, the amount of image acquisition mismatch obtained by subtracting the acquisition adaptation amount from one, and the amount of image acquisition mismatch obtained by subtracting the acquisition adaptation amount from one, are all considered. The proportion of candidate regions for supplementary structural components matching each damage evidence group is merged to calculate the insufficient image evidence. The proportion of candidate regions for supplementary structural components is then calculated according to the ratio of the first group of damage evidence. The number of candidate structural components matching the semantic labels of structural components and the vehicle body orientation labels in each damage evidence group is calculated. The single-view isolation quantity is calculated by proportionally dividing the number of valid structural component candidate regions with single-view isolation markers by the total number of valid candidate regions. This proportional calculation is limited to the [number of] [items / sections]. Within each damage evidence group; the arithmetic mean of the evidence group is obtained by taking the arithmetic mean of the cross-perspective homologous values. The arithmetic mean is used to characterize the first damage evidence group. The overall homology of cross-perspective aggregation relationships within each damage evidence group.

[0060] The exponential result of the normalized maintenance accident record support quantity is calculated to obtain the maintenance accident support amplification term. The exponential result is the result of the exponential operation with the natural constant as the base. The exponential function amplifies the supporting role of maintenance records, insurance accident records, and inspection report texts in the closed loop of accident evidence. The product of cross-view support quantity plus one, structural component damage intensity plus one, and maintenance accident support amplification term is calculated to obtain the accident closed loop support term. The plus one is used to retain the positive benchmark of cross-view support quantity and structural component damage intensity. The product structure is used to reflect the synergistic support relationship between visual evidence, damage characterization, and textual evidence. The system calculates the sum of the vehicle model baseline conflict quantity, the insufficient image evidence quantity, and the single-view isolation quantity, and adds one to obtain the accident closure suppression term. Adding one is used to maintain the stability of the denominator and ensure that the vehicle model baseline conflict, insufficient image evidence, and single-view isolation participate in the calculation in a suppressed form. The ratio of the accident closure support term to the accident closure suppression term is calculated, added one, and then the natural logarithm is taken to obtain the closure evidence logarithm term. The natural logarithm is used to compress the differences in evidence support and reduce the disturbance of outliers to the accident closure confidence value. The negative of the common source mean of the evidence group is calculated, the exponent is taken, and then one is subtracted from the exponent to obtain the common source mean support term. Following this, the negative exponential mapping is used to transform the homologous mean of the evidence group into a monotonically increasing homologous support factor; the product of the logarithmic term of the closed-loop evidence and the homologous mean support term is calculated to obtain the accident closed-loop confidence value. Through the product structure, the accident closed-loop confidence value is simultaneously constrained by the evidence closed-loop strength and cross-view homologous stability. Moreover, the cross-view support quantity, structural component damage strength, repair accident record support quantity, vehicle model benchmark conflict quantity, insufficient image evidence quantity, single-view isolation quantity, and homologous mean of the evidence group in the formula are all dimensionless quantities. The visual damage confidence information obtained from the structural component damage characterization detection and processing, and the cross-view aggregation are used to... The resulting multi-perspective evidence relationships, the matching results between business text fields and structural component semantic labels and vehicle body orientation labels are all incorporated into the calculation of the accident closed-loop credibility value. This ensures that accident damage determination no longer relies on a single image recognition result, but is supported by image damage evidence, cross-perspective homogeneous evidence, vehicle model structural benchmark constraints, and business text evidence. Simultaneously, the vehicle model benchmark conflict quantity, insufficient image evidence quantity, and single-perspective isolation quantity are used to suppress insufficient evidence or structural conflict situations, improving the technical credibility, evidence closed-loop, and traceability of structural component accident damage detection results. The specific calculation formula is as follows:

[0061] ;

[0062] In the formula, Indicates the closed-loop confidence value of the accident, used to determine the first... Whether a single set of damage evidence is sufficient to support the establishment of structural component accident damage; Represents cross-view support quantity, used to characterize the first The extent of multi-perspective evidence coverage in each injury evidence group; Indicates the damage strength of structural components, used to characterize the first The degree of comprehensive manifestation of structural component damage characterization sets within a damage evidence group; This indicates the amount of support from repair accident records, used to characterize the degree of textual support from business records for structural component accident damage; This represents the vehicle model reference conflict amount, used to characterize the degree of deviation between the candidate region and the vehicle model structural reference. Indicates insufficient image evidence, used to characterize the first The extent of insufficient image visibility and acquisition completeness in each group of injury evidence; Represents a single-view isolated quantity, used to characterize the first... The percentage of areas lacking cross-perspective support within each damage evidence set; Represents the common source mean of the evidence group, used to characterize the first... Average credibility of cross-perspective homology within each group of injury evidence; This indicates the sequence number of the damage evidence group generated according to the cross-perspective aggregation relationship, numbered... Each damage evidence group consists of valid structural component candidate regions aggregated to the same structural component and their structural component damage characterization sets, used to distinguish different damage evidence groups under the same vehicle inspection task.

[0063] In this implementation plan, the closed-loop strength assessment of accident evidence is conducted on the g-th damage evidence group. This establishes a unified judgment criterion for the coverage of multi-view visual evidence, the degree of structural component damage manifestation, the textual support of repair and accident records, the degree of deviation from vehicle structural benchmarks, the completeness of image evidence, the isolated state of a single view, and the stability of cross-view homogeneity. The accident closed-loop credibility value can comprehensively distinguish between structural component accident damage with sufficient evidence support and suspected damage with insufficient evidence. This avoids relying solely on single-view image anomalies, local damage characterization, or single business records to form accident conclusions. It improves the closed-loop credibility, evidence consistency, and traceability of structural component accident damage detection results, providing a clear quantitative basis for subsequent output of structural component accident damage detection results, structural component accident damage evidence chains, and manual review tasks.

[0064] Specifically, the process of determining structural component accident damage and archiving the evidence chain based on the accident closed-loop credibility analysis results is as follows: Real-time comparison of the accident closed-loop credibility value and the accident closed-loop credibility threshold:

[0065] When the accident closure confidence value is greater than or equal to the accident closure confidence threshold, it indicates that the damage evidence group simultaneously meets the requirements of cross-view aggregation support, damage characterization support, and maintenance accident record support. The corresponding structural component accident damage is determined to be established, and the structural component accident damage detection result is output. The structural component accident damage detection result includes the damaged structural component name, vehicle body orientation, damage characterization type, cross-view support image, maintenance accident record support fields, and accident risk assessment result, enabling the detection result to simultaneously present visual evidence, textual evidence, and structural component orientation evidence. Simultaneously, a structural component accident damage evidence chain is generated, specifically using the damage evidence group as the evidence chain. The main index sequentially encapsulates the valid structural component candidate regions, cross-view aggregation relationships, source image identifiers, shooting orientation labels, structural component semantic labels, structural component damage characterization sets, repeated damage suppression markers, maintenance accident record supporting fields, and accident closure confidence values ​​within the same damage evidence group. This forms a structural component accident damage evidence chain that can correspond to the image source, structural component orientation, damage characterization, and text record. This chain is created and archived in the vehicle quality assessment database. The evidence chain stores the damage evidence group, cross-view aggregation relationships, repeated damage suppression markers, and maintenance accident record supporting fields for subsequent review, pricing, and risk tracing.

[0066] When the accident closure confidence value is less than the accident closure confidence threshold, the conclusion that the structural component accident damage has been established is not directly output. Instead, the corresponding damage evidence group is marked as insufficient accident damage evidence to avoid forming an accident damage conclusion based solely on single-view anomalies or insufficient image quality. Based on the amount of insufficient image evidence, the amount of single-view isolation, and the candidate areas of the structural component to be supplemented, a manual review task is generated. The manual review task carries the name of the structural component to be reviewed, the vehicle body orientation, the source image, and the location to be supplemented. After the review is completed, the review conclusion is re-submitted to the structural component location acceptance analysis so that the supplementary evidence can participate again in the progressive calculation of the location acceptance value, cross-view source value, and accident closure confidence value.

[0067] In this implementation plan, the damage evidence group is finally adjudicated based on the accident closed-loop credibility value, so that the structural component accident damage that meets the evidence closed-loop requirements forms a clear detection result, and the structural component accident damage evidence chain is simultaneously accumulated, ensuring that the name of the damaged structural component, vehicle body location, damage characterization type, cross-view support images, maintenance accident record support fields, and accident risk assessment results have complete output basis. Damage evidence groups with insufficient evidence will not directly form a conclusion of accident damage, but will be transferred to the manual review task. Supplementary evidence will be added and then re-enter the progressive calculation process of location acceptance value, cross-view homology value, and accident closed-loop credibility value after review. This avoids misjudgment caused by single-view anomalies, insufficient image quality, or incomplete evidence chains, and improves the credibility, closed-loop, verifiability, and archiving traceability of structural component accident damage detection results.

[0068] like Figure 2 As shown, the second aspect of this invention provides a vehicle structural component accident damage detection system based on visual recognition, comprising: a data acquisition and preprocessing module, used to acquire structural component accident damage detection data in vehicle inspection tasks and perform vehicle model matching and binding, time sequence alignment, field extraction, and standardization processing, providing a unified data foundation for subsequent calculation of location acceptance values, cross-view homogeneous values, and accident closed-loop confidence values; and a structural component candidate region identification module, used to perform structural component location acceptance analysis based on the structural component accident damage detection data, and form effective structural component candidate regions and structural component damage characterization based on the structural component location acceptance analysis results. The system provides structural component location and damage characteristic basis for cross-perspective homogeneous aggregation; the cross-perspective homogeneous aggregation module is used to perform cross-perspective homogeneous analysis on valid structural component candidate regions, and form damage evidence groups and repeated damage suppression results based on the cross-perspective homogeneous analysis results, reducing the risk of repeated damage counting of the same structural component and erroneous merging of non-homogeneous regions; the accident damage closed-loop judgment and output module is used to perform accident closed-loop credibility analysis on the damage evidence groups, and perform structural component accident damage judgment and evidence chain archiving based on the accident closed-loop credibility analysis results, improving the closed-loop credibility and traceability integrity of structural component accident damage detection results.

[0069] This implementation plan establishes a complete processing chain from acquiring structural component accident damage detection data, identifying structural component candidate regions, cross-view homogeneous aggregation, to accident damage closed-loop determination. This enables image evidence, video evidence, business text evidence, and vehicle model structural benchmark data under the same vehicle inspection task to participate in the calculation at each level. Valid structural component candidate regions with stable positioning are screened through location acceptance values. Damage representations pointing to the same structural component from different perspectives are aggregated into damage evidence groups through cross-view homogeneous values. Finally, the damage evidence groups are reliably determined through accident closed-loop credibility values. This reduces duplicate counting of damage to the same structural component, incorrect merging of non-homogeneous regions, and misjudgment of single-view evidence, thereby improving the accuracy, closed-loop credibility, and evidence traceability of structural component accident damage detection results.

[0070] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0071] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, 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. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for detecting accident damage to vehicle structural components based on visual recognition, characterized in that, Includes the following steps: S1: Obtain structural component accident damage detection data from vehicle inspection tasks and perform vehicle model matching and binding, time sequence alignment, field extraction and standardization processing; S2, Based on the structural component accident damage detection data, perform structural component location and bearing analysis, and form effective structural component candidate areas and structural component damage characterization based on the structural component location and bearing analysis results; S3, perform cross-view homology analysis on candidate regions of effective structural components, and form damage evidence groups and repeated damage suppression results based on the cross-view homology analysis results; S4. Conduct a closed-loop reliability analysis of the damage evidence group, and determine the structural component accident damage and archive the evidence chain based on the results of the closed-loop reliability analysis. The specific process of performing structural component positioning and bearing analysis based on structural component accident damage detection data is as follows: Based on the structural component name table, structural component datum contour diagram, structural component datum position coordinates, and structural component datum projected area table, candidate box localization, mask segmentation, and datum constraint matching are performed on the vehicle multi-view images and surround-view video frame images, and the generation order is determined. Candidate regions for structural components; Multi-view vehicle images, surround-view video frame images, structural component name list, and structural component baseline contour maps are input into the object detection model and semantic segmentation model for structural component region matching, outputting the region matching quantity; the shooting orientation labels, vehicle part shooting labels, and structural component baseline orientation table are mapped using label consistency to obtain the shooting consistency quantity; for the 1st The number of pixels in the candidate region mask corresponding to each structural component candidate region is counted to obtain the area of ​​the candidate region mask. The area carrying capacity is calculated by comparing the area of ​​the candidate region mask with the projected area of ​​the corresponding structural component in the structural component reference projection area table. The amount of occlusion and reflection is obtained by detecting reflective areas and occlusion areas in multi-view images of vehicles and surround-view video frames; the acquisition adaptation amount is obtained by mapping the image width, image height, focal length, exposure time, sensitivity, and shooting distance values ​​through acquisition condition interval mapping. Calculate the sum of the area carrying capacity and the numerical protection constant, and take the square root to obtain the area carrying capacity root value; calculate the product of the region matching amount, the shooting consistency amount, and the area carrying capacity root value, and add one to obtain the positioning enhancement term; calculate the occlusion reflection amount and add one to obtain the occlusion suppression term; calculate the ratio of the positioning enhancement term to the occlusion suppression term and take the arctangent to obtain the angle compression term; calculate the negative of the acquisition adaptation amount and take the exponent, and then subtract the exponent result from one to obtain the acquisition adaptation term; calculate the product of the angle compression term and the acquisition adaptation term to obtain the positioning carrying capacity value. The specific process of performing cross-view homology analysis on candidate regions of effective structural components is as follows: With the first The first effective structural component candidate region and the first Each valid candidate structural component region constitutes a pair of regions to be matched. Within each pair, the semantic labels of structural components are matched using the structural component name table and the structural component adjacency table to obtain the semantic consistency quantity of the structural components. The anchor points of structural components are matched using the scale-invariant feature transformation algorithm, and the anchor point matching quantity of the structural components is obtained through homography estimation. The damage representation similarity quantity of the structural component damage representation set is calculated using cosine similarity. The vehicle body orientation deviation quantity is obtained by mapping the shooting orientation label, vehicle part shooting label, body orientation label, and structural component reference orientation table through orientation difference mapping. The image acquisition condition difference quantity is obtained by mapping the image width, image height, focal length, exposure time, ISO, shooting distance, image timestamp difference, video frame timestamp difference, and video frame sequence number difference through Chebyshev distance mapping. The occlusion conflict quantity is obtained by calculating the absolute value of the difference between the occlusion reflection quantity and the occlusion reflection quantity. Calculate the product of the positioning and bearing value of the a-th effective structural component candidate region and the positioning and bearing value of the b-th effective structural component candidate region, and take the square root to obtain the dual-region positioning and bearing root value; The cross-view homogeneous support term is obtained by multiplying the semantic consistency of structural components by one, the matching of structural component positioning anchor points by one, the similarity of damage characterization by one, and the root value of dual-region positioning by one. The sum of vehicle orientation deviation, image acquisition condition difference, and occlusion conflict is calculated and added to obtain the cross-view deviation suppression term; the ratio of the cross-view homogeneous support term to the cross-view deviation suppression term is calculated and the hyperbolic tangent is taken to obtain the cross-view homogeneous value.

2. The vehicle structural component accident damage detection method based on visual recognition according to claim 1, characterized in that: The specific process for obtaining structural component accident damage detection data in vehicle inspection tasks is as follows: Under the same vehicle inspection task, structural component accident damage detection data is obtained through remote shooting terminals, surround vehicle video acquisition terminals, vehicle business databases, and vehicle model structure databases. This data includes: vehicle identification number, brand and model name, vehicle year, multi-view images, image timestamp, image width, image height, focal length, exposure time, ISO, shooting distance, shooting orientation label, and vehicle part shooting label obtained through the remote shooting terminal; surround vehicle video files, video frame timestamps, and video frame sequences obtained through the surround vehicle video acquisition terminal; repair record text, insurance claim record text, inspection report text, and record generation time obtained through the vehicle business database; and corresponding structural component name tables, structural component reference orientation tables, structural component reference outline diagrams, structural component reference position coordinates, structural component reference projected area tables, and structural component adjacency relationship tables retrieved from the vehicle model structure database.

3. The method for detecting accident damage to vehicle structural components based on visual recognition according to claim 1, characterized in that: The specific process for performing vehicle model matching and binding, time sequence alignment, field extraction, and standardization is as follows: Vehicle model matching and binding are performed based on vehicle identification code, brand and model name, and vehicle year; video frame extraction processing is performed on the surround vehicle video file to obtain surround vehicle video frame images; vehicle multi-view images, surround vehicle video frame images, image timestamps, video frame timestamps, and video frame sequences are bound to the same task time; and the part field, repair action field, and accident direction field are extracted from the repair record text, insurance claim record text, and inspection report text using named entity recognition methods. The structural component accident damage detection data are standardized using a zero-mean unit variance standardization algorithm. Following the acquisition order of the same vehicle detection task, adjacent differences are calculated for image timestamps, video frame timestamps, and video frame sequence numbers to obtain image timestamp differences, video frame timestamp differences, and video frame sequence number differences. These differences are then normalized using a maximum-minimum-maximum normalization algorithm.

4. The method for detecting accident damage to vehicle structural components based on visual recognition according to claim 1, characterized in that: The specific process for forming effective candidate regions for structural components and characterizing structural component damage based on the structural component positioning and bearing analysis results is as follows: Real-time comparison of location acceptance value and location acceptance threshold: When the positioning acceptance value is greater than or equal to the positioning acceptance threshold, the corresponding area is marked as a valid structural component candidate area. Based on the structural component name table, structural component reference contour map, structural component reference position coordinates, structural component reference orientation table, shooting orientation label and vehicle part shooting label, the semantic label of the structural component and the vehicle body orientation label are determined. The effective structural component candidate area is processed by the scale invariant feature transformation algorithm, edge contour extraction processing and structural component damage characterization detection processing to generate structural component positioning anchor points and structural component damage characterization set for cross-view homology analysis. When the location acceptance value is less than the location acceptance threshold, the area is marked as a candidate area for the structural component to be supplemented. Based on the corresponding shooting orientation label, vehicle part shooting label and structural component reference orientation table, the missing viewpoint set and the shooting part to be supplemented are determined, and a supplementary shooting task is generated. Before the supplementary shooting is completed, the area is only used as auxiliary evidence to be reviewed.

5. The method for detecting accident damage to vehicle structural components based on visual recognition according to claim 1, characterized in that: The specific process for forming damage evidence sets and repeated damage suppression results based on cross-perspective homology analysis results is as follows: Real-time comparison of cross-view homology values ​​and cross-view homology thresholds: When the cross-view homology value is greater than or equal to the cross-view homology threshold, the first... The first effective structural component candidate region and the first Each valid structural component candidate region is grouped into the same damage evidence group, and a cross-view grouping relationship is generated; for repeated damage characterization regions identified by region overlap and characterization similarity verification within the same damage evidence group, repeated damage suppression markers are generated. When the cross-view homology value is less than the cross-view homology threshold, maintain the first The first effective structural component candidate region and the first The independent attribution relationship of each valid structural component candidate region is defined. When no cross-view aggregation relationship is formed for the same valid structural component candidate region, it is marked as isolated from a single view.

6. The method for detecting accident damage to vehicle structural components based on visual recognition according to claim 1, characterized in that: The specific process for conducting a closed-loop reliability analysis of the damage evidence set is as follows: In the Within each damage evidence group, the cross-view support quantity is obtained by deduplicating the labels of different shooting directions, the labels of vehicle parts, and the number of candidate areas of effective structural components; the structural component damage intensity is obtained by calculating the structural component damage characterization set through probability union. The repair record text, insurance claim record text, and inspection report text contain fields for part, repair action, claim direction, and record generation time. These fields are matched with the semantic tags of structural components and the vehicle body orientation tags using a semantic matching method to obtain the support quantity of the repair claim record. The candidate region mask corresponding to the effective structural component candidate region is subjected to boundary extraction and mask center point calculation to obtain the candidate region mask boundary and candidate region center coordinates. The contour difference and position offset are calculated with the structural component reference contour map and the structural component reference position coordinates to obtain the vehicle model reference conflict quantity. The amount of occlusion reflection, the image acquisition mismatch obtained by subtracting the acquisition adaptation amount from one, and the amount of occlusion reflection are considered. The proportion of candidate structural components to be supplemented that match each damage evidence group is merged to obtain the insufficient image evidence quantity; the ratio of the number of valid structural component candidate regions with single-view isolation marks to the total number of valid structural component candidate regions is calculated to obtain the single-view isolation quantity; the arithmetic mean of cross-view homologous values ​​is taken to obtain the homologous mean of the evidence group. Calculate the index result of the support amount of maintenance accident records to obtain the maintenance accident support amplification term; The accident closed-loop support term is obtained by multiplying the cross-view support quantity by one, the structural component damage intensity by one, and the maintenance failure support amplification term. The sum of the vehicle model baseline conflict quantity, the insufficient image evidence quantity, and the single-view isolation quantity is added by one to obtain the accident closure suppression term; the ratio of the accident closure support term to the accident closure suppression term is calculated, added by one, and then the natural logarithm is taken to obtain the closure evidence logarithm term; the negative of the common source mean of the evidence group is calculated, the exponent is taken, and then the exponent is subtracted by one to obtain the common source mean continuation term; the product of the closure evidence logarithm term and the common source mean continuation term is calculated to obtain the accident closure confidence value.

7. The method for detecting accident damage to vehicle structural components based on visual recognition according to claim 1, characterized in that: The specific process for determining structural component accident damage and archiving the evidence chain based on the closed-loop reliability analysis results of the accident is as follows: Real-time comparison of the accident closure confidence value and the accident closure confidence threshold: When the closed-loop confidence value of the accident is greater than or equal to the closed-loop confidence threshold of the accident, the corresponding structural component accident damage is determined to be established, and the structural component accident damage detection result is output; at the same time, the structural component accident damage evidence chain is generated, created and archived in the vehicle quality assessment database. When the confidence value of the closed loop of the accident is less than the confidence threshold of the closed loop of the accident, the conclusion that the structural component accident damage is established is not directly output. Instead, the corresponding damage evidence group is marked as the result of insufficient accident damage evidence. Based on the amount of insufficient image evidence, the amount of isolated single-view evidence, and the candidate areas of the structural component to be supplemented, a manual review task is generated. After the review is completed, the review results will be resubmitted to the structural component positioning and bearing analysis.

8. A vehicle structural component accident damage detection system based on vision recognition, applied to the vehicle structural component accident damage detection method based on vision recognition according to any one of claims 1-7, comprising: The data acquisition and preprocessing module is used to acquire structural component accident damage detection data in vehicle inspection tasks and perform vehicle model matching and binding, time sequence alignment, field extraction and standardization processing. The structural component candidate region identification module is used to perform structural component location and bearing analysis based on structural component accident damage detection data, and to form effective structural component candidate regions and structural component damage characterization based on the structural component location and bearing analysis results. The cross-perspective homology aggregation module is used to perform cross-perspective homology analysis on candidate regions of effective structural components, and to form damage evidence groups and repeated damage suppression results based on the cross-perspective homology analysis results. The accident damage closed-loop determination and output module is used to perform accident closed-loop credibility analysis on the damage evidence group, and to determine the structural component accident damage and archive the evidence chain based on the results of the accident closed-loop credibility analysis.

Images