Intelligent collection and management system and method for arbitration data based on large model

Abstract

The application discloses an arbitration data intelligent collection management system and method based on a large model, relates to the field of arbitration data management, and comprises the following steps: constructing an original log record table according to original images of an accident scene; performing pseudo-focal length marking on the original images based on the original log record table to obtain candidate pseudo-focal length images; collecting reference targets in the candidate pseudo-focal length images, and constructing triplets based on the reference targets and the candidate pseudo-focal length images; identifying real pseudo-focal length images based on the triplets, and performing scaling coefficient calculation on the real pseudo-focal length images to obtain a correction proportion coefficient. The application provides a management scheme that is accurate, controllable, and tamper-proof in the whole process from data collection to evidence storage, and provides reliable technical support for traffic accident liability division.

Description

Technical Field

[0001] This invention relates to the field of arbitration data management technology, and in particular to an intelligent arbitration data acquisition and management system and method based on a large model. Background Technology

[0002] In traffic accident arbitration and liability determination scenarios, visual evidence from the accident scene is crucial for reconstructing the collision process and estimating the collision angle and speed. Currently, most scene images are taken by the parties involved using various mobile phone models. Legal teams need to use these images to reconstruct 3D scenes through multi-view modeling technology, and then quantitatively analyze key parameters such as vehicle deformation depth and brake mark length.

[0003] However, due to differences in mobile phone hardware characteristics and algorithm design, there is a significant hidden distortion problem in the EXIF ​​metadata of images: on the one hand, different manufacturers have different implementations of recording EXIF ​​focal length; on the other hand, dual or triple-camera phones will automatically switch from the physical telephoto lens to the main camera and digital crop in low light, close-up and other scenarios, but still write the telephoto equivalent focal length in EXIF, resulting in a mismatch between the nominal parameters of the image and the actual imaging field of view.

[0004] In existing technologies, modeling software typically relies directly on EXIF ​​focal length for parameter initialization. If the aforementioned distortion issues exist, it can lead to scale drift in the 3D reconstruction model, resulting in collision angle and velocity calculation errors exceeding 15%, far exceeding the threshold requirements for liability determination. Furthermore, the industry generally lacks a unified calibration mechanism and evidence preservation methods for multi-source images, and issues such as image tampering and the untraceability of parameter correction processes further exacerbate disputes over liability determination. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies in traffic accident arbitration images, such as metadata focal length distortion, 3D reconstruction scale drift, and data untraceability, which make it difficult to quantify the division of responsibility. The invention proposes an intelligent acquisition and management system and method for arbitration data based on a large model.

[0006] To achieve the above objectives, this invention provides an intelligent data acquisition and management method for arbitration based on a large model, comprising:

[0007] S1. Construct the original log record table based on the original images of the accident scene;

[0008] S2. Based on the original log record table, perform pseudo-focal length marking on the original image to obtain candidate pseudo-focal length images;

[0009] S3. Select an object with a standard physical length from the candidate pseudofocal length image as a reference target, and construct a triplet based on the reference target and the candidate pseudofocal length image;

[0010] S4. Based on the true field of view and the nominal field of view of the triplet, identify the true pseudo-focal distance image among the candidate pseudo-focal distance images, and calculate the scaling factor of the true pseudo-focal distance image to obtain the correction scaling factor.

[0011] S5. Based on the correction scaling factor, calibrate the original image to obtain the correction camera intrinsic parameter matrix. Package the correction camera intrinsic parameter matrix, the correction scaling factor, and the hash value of the original image into a calibration file.

[0012] S6. Write the calibration documents and original images into the compliant evidence storage system, generate an arbitration data collection report of the accident scene based on the evidence storage information, and form a management plan for arbitration data.

[0013] Preferably, an original log record table is constructed based on the original images of the accident scene, including:

[0014] The original images of the accident scene are hashed to obtain their hash values.

[0015] Extract metadata from the original image, including: equivalent focal length, ISO, exposure time, digital zoom ratio, and device model.

[0016] The original log table is constructed based on the original image, the hash value of the original image, and the metadata of the original image.

[0017] Preferably, the original image is pseudo-focal length marked based on the original log record table to obtain candidate pseudo-focal length images, including:

[0018] Read the information recorded in the original log record table;

[0019] Low light determination is performed on the original image in the recorded information; if the photosensitivity of the original image in the recorded information is greater than the preset photosensitivity threshold and the exposure time is greater than the preset exposure time threshold, then it is determined that the original image in the recorded information has low light phenomenon.

[0020] The original image in the recorded information is determined to be telephoto; if the equivalent focal length of the original image in the recorded information is greater than the preset focal length threshold, then the original image in the recorded information is determined to have a telephoto phenomenon.

[0021] The original image in the recorded information is cropped; if the digital zoom ratio of the original image in the recorded information is greater than the preset zoom ratio threshold, it is determined that the original image in the recorded information has been cropped.

[0022] If the original image in the recorded information simultaneously exhibits low light, telephoto, and cropping phenomena, then the original image in the recorded information is marked with a pseudo-focal length, and the original image in the recorded information is merged into the candidate pseudo-focal length image set.

[0023] Preferably, an object with a standard physical length is selected as a reference target from the candidate pseudofocal length image, and a triplet is constructed based on the reference target and the candidate pseudofocal length image, including:

[0024] The trained target detection model is used to perform target detection on the candidate pseudofocal images in the candidate pseudofocal image set to obtain a reference target set for the candidate pseudofocal images.

[0025] Select the target with the highest confidence level from the reference target set as the reference target;

[0026] The physical length of the reference target is retrieved from the standard length library, and the pixel length of the reference target in the candidate pseudofocal length image is calculated.

[0027] Construct triples based on candidate pseudofocal length images, physical length, and pixel length.

[0028] Preferably, based on the true field of view and nominal field of view of the triplet, the true pseudo-focal distance image among the candidate pseudo-focal distance images is identified, and the scaling factor of the true pseudo-focal distance image is calculated to obtain the correction scaling factor, including:

[0029] The true field of view of the candidate pseudofocal image is calculated based on the physical length, pixel length, and device model in the metadata from the triplet.

[0030] The nominal field of view of the candidate pseudofocal image is calculated based on the equivalent focal length of the candidate pseudofocal image.

[0031] The difference between the actual field of view and the nominal field of view is obtained.

[0032] If the difference in field of view is greater than the preset difference threshold, the candidate pseudo-focal image is determined to be a real pseudo-focal image.

[0033] The scaling factor is calculated for the true field of view and nominal field of view of the real pseudo-focal length image to obtain the correction scaling factor.

[0034] Preferably, the original image is calibrated based on the correction scaling factor to obtain the corrected camera intrinsic parameter matrix. The corrected camera intrinsic parameter matrix, the correction scaling factor, and the hash value of the original image are packaged into a calibration file, including:

[0035] Image matching is performed between the real pseudo-focal length image and the original log record table to obtain the original image in the original log record table that corresponds to the real pseudo-focal length image.

[0036] The equivalent focal length in the original image is calibrated using a correction scaling factor to obtain the corrected camera intrinsic parameter matrix;

[0037] The corrected camera intrinsic matrix, the corrected scaling factor, and the hash value of the original image are packaged into a calibration file.

[0038] Preferably, the calibration documents and original images are written into a compliant evidence storage system, and an arbitration data collection report for the accident scene is generated based on the stored evidence information, forming a management scheme for arbitration data, including:

[0039] Evidence storage information is generated based on the hash value of the original image, the calibration file, the hash value of the calibration file, and the timestamp, and then stored in a compliant evidence storage system.

[0040] An arbitration data collection report for the accident scene is generated based on the evidence information.

[0041] The arbitration data collection report was analyzed and integrated to obtain a management plan for the arbitration data.

[0042] To address the aforementioned problems, this invention also provides an intelligent arbitration data acquisition and management system based on a large model, the system comprising:

[0043] The image log construction module is used to build an image data log table based on the original images of the accident scene, realize the extraction and structured management of unique image identifiers, and form an accident scene image log table.

[0044] The pseudo-focal length labeling module, based on the image log table, performs visual feature mining and focal length labeling inference on the original image to obtain a set of candidate pseudo-focal length images for subsequent judgment and correction.

[0045] The reference object extraction module is used to identify target reference objects with calibrated physical properties in the candidate pseudo-focal length image set and to fuse pseudo-focal length images to construct triples.

[0046] The field of view calculation module is used to calculate the difference between the real field of view and the nominal field of view based on the triplet, identify real pseudo-focal length images, and calculate the scaling correction coefficient.

[0047] The inverse camera geometry correction module is used to perform inverse reconstruction of the camera intrinsic parameter matrix on the real pseudo-focal length image based on the obtained correction scale coefficient, output the corrected camera intrinsic parameter matrix, and generate a calibration file together with the original image hash value.

[0048] The evidence storage module is used to generate evidence information based on calibration files, the hash value of calibration files, timestamps, and the hash value of original images, and store it in the compliant evidence storage system. It also generates an arbitration data collection report of the accident scene based on the evidence information.

[0049] Compared with the prior art, the beneficial effects of the present invention are:

[0050] 1. In this invention, an original log record table is constructed using the original images of the accident scene. The original images in the log record table are marked with pseudo-focal distance to obtain candidate pseudo-focal distance images. Trustworthy original metadata is locked and suspected pseudo-focal distance images are preliminarily screened based on this, forming a dual defense line of source quality control and risk prediction for arbitration data collection.

[0051] 2. In this invention, by selecting an object with a standard physical length as a reference target in the candidate pseudo-focal distance image, and constructing a triplet based on the reference target and the candidate pseudo-focal distance image, a direct correspondence between the image pixel space and the real physical space is established. Based on the real field of view and nominal field of view of the triplet, the real pseudo-focal distance image in the candidate pseudo-focal distance image is identified. The scaling factor of the real pseudo-focal distance image is calculated to obtain the correction scaling factor, accurately identifying the image that actually has a pseudo-focal distance, avoiding misjudgment, and providing directly usable correction parameters for subsequent calibration. This solves the core problem of dual-scale coordinate systems appearing at the same accident scene due to pseudo-focal distance from a data perspective.

[0052] 3. In this invention, the original image is calibrated by correcting the scaling factor to obtain the corrected camera intrinsic parameter matrix. The corrected camera intrinsic parameter matrix, the corrected scaling factor, and the hash value of the original image are packaged into a calibration file to avoid model scaling drift or point cloud islanding caused by camera parameter errors, thus significantly improving the accuracy of 3D reconstruction. The calibration file and the original image are written into a compliant evidence storage system, and an arbitration data collection report of the accident scene is generated based on the evidence storage information, forming an arbitration data management scheme. This achieves full-process traceability and tamper-proof protection from original image acquisition to parameter correction, ensuring data credibility. Attached Figure Description

[0053] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0054] Figure 1 A flowchart illustrating an intelligent data acquisition and management method for arbitration based on a large model, provided in an embodiment of the present invention;

[0055] Figure 2 This is a functional block diagram of an intelligent arbitration data acquisition and management system based on a large model, provided as an embodiment of the present invention. Detailed Implementation

[0056] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0057] Example: This example provides an intelligent data acquisition and management method for arbitration based on a large model. See [link to example]. Figure 1 Specifically, including:

[0058] S1. Construct the original log record table based on the original images of the accident scene;

[0059] Specifically, in traffic accident arbitration, the authenticity and uniqueness of the metadata of the original image are the core foundation for subsequent 3D reconstruction and liability determination. If the equivalent focal length and exposure parameters of the original image are tampered with or misread, it will directly lead to distortion in the calculation of key parameters such as collision angle and brake mark length. Therefore, the core purpose of constructing the original log record table is to solidify the feature information and metadata of the original image to form an immutable initial data archive, providing a traceable benchmark for pseudo-focal length detection and parameter correction throughout the entire process.

[0060] In an embodiment of the present invention, an original log record table is constructed based on the original images of the accident scene, including:

[0061] Obtain metadata of the original images of the accident scene;

[0062] Specifically, the metadata of the original image refers to the shooting parameter information embedded in the image file, and its integrity directly determines the accuracy of subsequent pseudo-focal length determination; the hash value is used to ensure that the image has not been tampered with during the transfer process, and is the key to data traceability.

[0063] Specifically, the first step is to collect all original images from the accident scene, in formats including but not limited to JPEG, PNG, and RAW. Then, core metadata is extracted using the professional image analysis tool ExifTool. This core metadata includes: equivalent focal length, ISO, exposure time, digital zoom ratio, and device model. Specifically, the equivalent focal length reflects the lens's actual angle of view, ISO reflects the sensor's sensitivity to light, exposure time reflects ambient light intensity, digital zoom ratio reflects the cropping ratio of the original image by the device, and the device model is used to match sensor parameters in a publicly available database.

[0064] During the extraction process, if the image is damaged or has format issues that result in missing metadata, it must be marked as metadata abnormal and supplemented by qualified personnel using the equipment manual and publicly available parameter tables to ensure the integrity of subsequent detection.

[0065] Perform a hash calculation on the original image to generate the hash value of the original image;

[0066] Specifically, a hash value is a unique string calculated from the complete binary data of an image using a cryptographic algorithm. Its core characteristic is that even a small change in the original data will result in a completely different hash value, which can be used to verify whether an image has been tampered with.

[0067] Specifically, the complete binary stream of the original image is read, and the binary stream is processed using the SHA-256 algorithm to generate a 64-bit hexadecimal string. The generated string is the unique hash value of the image. The validity of the hash value needs to be ensured by double verification. The hash value is calculated repeatedly for the same image. If the two results are inconsistent, it indicates that the image is corrupted during the reading process, and the original image needs to be obtained again.

[0068] The original image, hash value, and metadata are integrated into a read-only log record;

[0069] Specifically, read-only log records are structured mirrors of the original image data. Through access control and format solidification, the original information is prevented from being modified by humans. Its core function is to bind the image, metadata, and hash value to form a data chain that can be traced for life after a single collection.

[0070] Specifically, the original image, the generated hash value, and the read metadata are linked and integrated to form combined data containing the original image itself, the hash value, and the metadata. This combined data is written into the original log record, which is a JSON file. The file fields include: original image, original image hash value, equivalent focal length, ISO, exposure time, digital zoom ratio, device model, shooting timestamp, data collector number, and image ID. The image ID consists of the case number and the image sequence number.

[0071] After the logs are generated, they are locked to an uneditable state through operating system permission settings and synchronously stored in a distributed database to ensure that even if the data on a single node is damaged, the complete logs can still be recovered through other nodes.

[0072] The core of constructing the original log table is based on the principle of data solidification; it ensures parameter parsing by extracting metadata, ensures content immutability through hash calculation, and ensures record immutability through read-only storage. The log table formed by combining these three elements serves as the benchmark for subsequent pseudo-focal length detection.

[0073] S2. Based on the original log record table, perform pseudo-focal length marking on the original image to obtain candidate pseudo-focal length images;

[0074] Specifically, the pseudo-focal length phenomenon occurs when a mobile phone shoots in low light or with a telephoto lens, but the actual image cropped using the main camera is recorded as a telephoto focal length in the metadata, resulting in a discrepancy between the nominal parameters of the image and the actual image. If such images are directly used for 3D reconstruction, contradictions will arise between two scales of the same scene. For example, the length of a vehicle in the main camera image may be 5m, while the length of the same vehicle in the pseudo-telephoto image may be only 3m. Therefore, the core purpose of pseudo-focal length labeling is to analyze the metadata characteristics in the log record table to filter out images suspected of parameter deception, thus narrowing down the scope for subsequent accurate correction.

[0075] In an embodiment of the present invention, pseudo-focal length marking is performed on the original image based on the original log record table, including:

[0076] Read a single record from the original log record table;

[0077] Specifically, a single record is read from the original log record table using the image ID, and four metadata items related to pseudo-focal length determination are extracted: equivalent focal length, ISO, exposure time, and digital zoom ratio. If there are metadata anomaly markers in the record, manual verification should be triggered first, and the record should be supplemented before entering the marking process.

[0078] Determine low-light conditions in the original image;

[0079] Specifically, low light is the core environmental factor that triggers the phone's camera to automatically switch from physical telephoto to main camera cropping. Experiments show that when the ambient light is below EV4, the phone will automatically enable the main camera plus digital cropping mode to avoid image quality degradation caused by insufficient light entering the telephoto lens. However, the metadata will still record the telephoto parameters, forming a pseudo focal length.

[0080] Specifically, low-light conditions must be met simultaneously: ISO greater than 800 and exposure time greater than 1 / 33 second. Higher ISO indicates that the sensor compensates for insufficient light by increasing gain. When ISO is greater than 800, the light intensity is less than 5 lux, which is equivalent to nighttime when streetlights are insufficient. Longer exposure time indicates that the sensor needs to collect light for a longer period of time. 1 / 33 second is the critical value for the phone to balance image brightness and motion blur in low light.

[0081] If both of the above conditions are met, the original image is determined to have low light.

[0082] Determine the telephoto effect in the original image;

[0083] Specifically, the determination of telephoto phenomenon requires meeting one condition: the equivalent focal length is greater than or equal to 65 mm. The basis for setting this threshold is that the equivalent focal length of the physical telephoto lens of mainstream mobile phones is usually greater than 65 mm, while the equivalent focal length of the main camera is mostly 24 mm to 28 mm. Therefore, 65 mm can effectively distinguish between nominal telephoto and nominal main camera images.

[0084] If the equivalent focal length is greater than or equal to 65 mm, the original image is determined to have a telephoto effect.

[0085] Determine if the original image is cropped.

[0086] Specifically, the cropping phenomenon needs to meet one condition: the digital zoom ratio is greater than 1.1. The basis for setting this threshold is that when the digital zoom ratio is equal to 1.0, the image is the sensor's native size. When the digital zoom ratio is greater than 1.0, the image only retains the central area and crops the edge part. 1.1 is the minimum magnification that enables digital cropping when the mobile phone is in the nominal telephoto mode.

[0087] If the digital zoom ratio is greater than 1.1, the original image is determined to have been cropped.

[0088] Generate a set of candidate pseudofocal length images;

[0089] Specifically, if the original image in a single log record simultaneously exhibits low light, telephoto, and cropping phenomena, then the image is marked as a candidate pseudofocal image and included in the candidate pseudofocal image set.

[0090] S3. Select an object with a standard physical length from the candidate pseudofocal length image as a reference target, and construct a triplet based on the reference target and the candidate pseudofocal length image;

[0091] Specifically, the core problem of candidate pseudofocal length images is the mismatch between the nominal focal length and the actual field of view. The essence of the field of view is the mapping relationship between the physical spatial scale and the pixel spatial scale. Therefore, it is necessary to establish the relationship between the two through a reference target with known physical length.

[0092] In an embodiment of the present invention, an object with a standard physical length is selected as a reference target from the candidate pseudofocal length image, and a triplet is constructed based on the reference target and the candidate pseudofocal length image, including:

[0093] Identify reference targets in candidate pseudofocal length images using an object detection model;

[0094] Specifically, a reference target refers to an object at a traffic accident scene that has a legally standardized physical length. Its length is uniform and unambiguous throughout the country and can be used as an absolute scale benchmark. The main reference targets include two categories: close-up motor vehicle license plates, which are suitable for images taken within 3 meters, and distant standard lane lines, which are suitable for images taken from 3 meters away.

[0095] Specifically, an object detection model is used to identify objects in candidate pseudo-focal distance images. The training and deployment process of the object detection model includes: collecting 100,000 images of traffic accident scenes containing license plates and lane lines, covering daytime, nighttime, sunny, rainy, frontal, and oblique scenes; using the LabelMe tool to annotate the bounding boxes of reference targets; employing the YOLOv8 algorithm, with the input image size set to 640×640 pixels, and outputting the target category (license plate or lane line), bounding box coordinates, and confidence score; using the bounding box intersection-union ratio (IUU) as the core indicator, through 50 rounds of iterative training, ensuring that the IUU of license plate detection bounding boxes is greater than or equal to 0.85 and the IUU of lane line detection bounding boxes is greater than or equal to 0.8; inputting candidate pseudo-focal distance images into the trained model, filtering out license plate targets with a confidence score greater than or equal to 0.85 or lane line targets with a confidence score greater than or equal to 0.8, and taking the target with the highest confidence score in the same image as the reference target;

[0096] Obtain the standard physical length and pixel length of the reference target;

[0097] Specifically, the corresponding value is retrieved from the standard length library. According to the regulations for motor vehicle license plates, the corresponding length for a small car license plate is 440 mm; according to the regulations for road traffic signs and markings, the length of a solid line segment at the edge of the roadway is 3750 mm.

[0098] Based on the bounding box coordinates of the reference target output by the target detection model, the horizontal pixel length of the reference target is obtained by subtracting the horizontal coordinate of the upper left corner from the horizontal coordinate of the lower right corner.

[0099] If there is no matching reference target in the image, manual intervention is triggered, and the accident investigator manually marks other standard targets to supplement the physical length and pixel length data.

[0100] Construct and store triples;

[0101] Specifically, a triple is a structured association of an image, physical length, and pixel length. Its function is to transform the intangible field of view into a computable scale relationship. A triple is formed by associating the candidate pseudo-focal length image, the standard physical length of the reference target, and the pixel length. The format of the triple is as follows:<P,L,X> Where P is the ID of the candidate pseudofocal image, L is the standard physical length of the reference target, and X is the pixel length of the reference target;

[0102] After the triplet is generated, it is written to the triplet table. The table structure includes the candidate pseudofocal length image ID, physical length, pixel length, generation time, and operator ID fields to ensure data traceability.

[0103] S4. Based on the true field of view and the nominal field of view of the triplet, identify the true pseudo-focal distance image among the candidate pseudo-focal distance images, and calculate the scaling factor of the true pseudo-focal distance image to obtain the correction scaling factor.

[0104] Specifically, the field of view is a core parameter reflecting the width of a lens's field of view. The shorter the focal length, the larger the field of view; the longer the focal length, the smaller the field of view. The essence of pseudo-focal length is that the field of view corresponding to the equivalent focal length deviates significantly from the field of view of the actual image. Therefore, by calculating the true field of view and the nominal field of view of the triplets, the existence of pseudo-focal length can be accurately determined. The scaling factor is the proportional benchmark for correcting the equivalent focal length to the true focal length, providing a basis for subsequent camera intrinsic parameter calibration.

[0105] In embodiments of the present invention, based on the true field of view and nominal field of view of the triplet, identifying the true pseudo-focal length image and calculating the correction scaling factor includes:

[0106] Obtain triplet data and device parameters to calculate the true field of view;

[0107] Specifically, read triples from the triple table, query the device model of the image in the triple in the raw log record table, and query the sensor width and native resolution width from the public database;

[0108] Divide the sensor width by the native resolution width to obtain the pixel physical size; multiply the pixel length in the triplet by the pixel physical size to obtain the image length of the reference target on the sensor; when the lens is imaging, the physical length of the reference target and its image length on the sensor will form two sets of similar triangles. According to the principle of similar triangles and the Gaussian lens law, the shooting distance is equal to the actual focal length multiplied by the ratio of the physical length of the reference target to the image length plus 1.

[0109] Based on the geometric relationship between the physical length of the reference target and the imaging length, the true field of view can be calculated using the following formula:

[0110]

[0111] In the formula, F is the true field of view angle. For reference target physical length, For shooting distance, It is the arctangent function;

[0112] Calculate the nominal field of view based on the image's equivalent focal length;

[0113] Specifically, the triplet table is read, and the equivalent focal length of the image in the triplet is retrieved from the original log record table; based on the geometric relationship of lens imaging, the nominal field of view can be calculated using the following formula:

[0114]

[0115] In the formula, Nominal field of view For sensor width, For equivalent focal length, It is the arctangent function;

[0116] Identify real pseudo-focal length images and calculate correction scaling factors;

[0117] Specifically, the difference between the true field of view and the nominal field of view is obtained, and the absolute value of the difference is taken as the field of view difference value. If the field of view difference value is greater than 8 degrees, the candidate pseudo-focal distance image in the triplet is determined to be a true pseudo-focal distance image. The threshold value of the field of view difference value is obtained by statistically analyzing 1000 sets of experimental data and taking a 95% confidence interval. The correction ratio coefficient of the true pseudo-focal distance image is obtained by dividing the nominal field of view by the true field of view.

[0118] Write the real pseudo-focal length image and the correction scaling factor into the scaling factor table. The table structure includes: real pseudo-focal length image ID, real field of view, nominal field of view, correction scaling factor, and decision time.

[0119] S5. Based on the correction scaling factor, calibrate the original image to obtain the correction camera intrinsic parameter matrix. Package the correction camera intrinsic parameter matrix, the correction scaling factor, and the hash value of the original image into a calibration file.

[0120] Specifically, the camera intrinsic parameter matrix is ​​the core parameter in 3D reconstruction that maps three-dimensional spatial points to two-dimensional image pixels. Its core element, focal length, directly determines the scale accuracy. If the intrinsic parameter matrix of a pseudo-focal length image is calculated based on an incorrect nominal focal length, it will lead to a distortion in the scale of the reconstructed model. Therefore, the core purpose of this step is to adjust the focal length parameter in the intrinsic parameter matrix by correcting the scaling coefficient, generate a calibration file that reflects the true imaging, and ensure its unique association with the original image through hash binding.

[0121] In an embodiment of the present invention, calibrating the original image based on a correction scaling factor and generating a calibration file includes:

[0122] Calculate the corrected camera intrinsic parameter matrix based on the original image and the correction scaling factor;

[0123] Specifically, the real / pseudofocal length image ID and correction ratio coefficient are read from the scaling factor table. The real / pseudofocal length image ID is matched with the corresponding original image in the original log record table to extract the equivalent focal length and principal point coordinates of the original image. The origin of the coordinate axis is the top left corner of the image plane. The x-axis extends horizontally to the right along the image plane and the y-axis extends vertically downward along the image plane. The principal point coordinates are the intersection of the lens optical axis and the image plane. The equivalent focal length is divided by the pixel physical size to obtain the pixel unit focal length. The correction ratio coefficient is multiplied by the pixel unit focal length to obtain the adjusted focal length. The corrected focal length and principal point coordinates are filled into the camera intrinsic parameter matrix to obtain the corrected camera intrinsic parameter matrix. The format of the corrected camera intrinsic parameter matrix is ​​[[f_x,0,c_x], [0,f_y,c_y], [0,0,1]], where f_x and f_y are the focal lengths in the x and y directions, respectively, and c_x and c_y are the principal point coordinates.

[0124] Package the calibration information and generate a calibration file;

[0125] Specifically, the calibration file is the solidified carrier of the corrected intrinsic parameter matrix. Its core function is to force subsequent 3D reconstruction software to use the corrected parameters and avoid scale drift through hard constraints.

[0126] Specifically, the original image hash value, correction scaling factor, correction camera intrinsic parameter matrix, and timestamp are integrated to form a set containing all the above information. Then, the set is encoded in JSON format to generate the initial calibration file. After the calibration file is generated, it is locked to an uneditable state through file system permission settings and stored in the directory associated with the original image. The file name format is: imageID_calib.json.

[0127] S6. Write the calibration documents and original images into the compliance evidence storage system, generate an arbitration data collection report of the accident scene based on the evidence storage information, and form a management plan for arbitration data;

[0128] Specifically, the legal validity of arbitration data depends on the traceability and tamper-proof nature of the entire process; if calibration documents or original images are tampered with during the transfer, the conclusion on the division of responsibility will become invalid; therefore, the core purpose of this step is to solidify key data through a compliant evidence storage system, generate an arbitration report containing information from the entire process, form a closed-loop management scheme for collection, correction and evidence storage, and ensure the credibility of data in the arbitration process.

[0129] In an embodiment of the present invention, writing the calibration file and the original image into a compliant evidence storage system and generating an arbitration report includes:

[0130] Integrate the evidence storage information and write it into a compliant evidence storage system;

[0131] Specifically, the hash value of the original image corresponding to the calibration file is obtained from the original log record table; the correction scaling factor, correction intrinsic parameter matrix, and timestamp are extracted from the calibration file; SHA-256 hash calculation is performed on the calibration file to obtain its hash value; the above information is used to construct evidence storage information, where the evidence storage information fields include: case number, hash value of the original image, calibration file hash, correction scaling factor, correction intrinsic parameter matrix, and timestamp; the data is written to the compliant evidence storage system (implemented based on blockchain technology), and the above evidence storage information is stored in <case number, block index> format via the blockchain API. The blockchain automatically records the writing node and timestamp, generating an immutable evidence record. The blockchain architecture is a consortium blockchain, with nodes including 3 arbitration institutions, 2 notary offices, and 1 technology provider. It employs a practical Byzantine fault-tolerant algorithm, with a consensus latency of less than or equal to 3 seconds, ensuring that the evidence information is uploaded to the chain in real time and is tamper-proof. Each block contains the hash value of the original image, the hash of the calibration file, a timestamp, and the node signature. The original image and calibration file are encrypted using the national cryptographic algorithm SM4 when stored locally, with the key automatically generated by the system and bound to the case number; only the hash of the evidence information is uploaded to the chain.

[0132] Generate arbitration data collection reports based on evidence storage information;

[0133] Specifically, the arbitration data collection report is a visual presentation of evidence information. It needs to clearly show the original data, the correction process, and the evidence, so that the arbitration tribunal can quickly understand the validity of the data.

[0134] Specifically, the process involves retrieving the block index information corresponding to the target case from the on-chain evidence storage system. This block index information includes the case number, the hash value of the original image, the block hash of the calibration file, and a timestamp. Next, based on the calibration file block hash in the block index, the corresponding calibration file is obtained from the evidence storage system. The calibration file contains the hash value of the original image, the scaling factor, the corrected camera intrinsic parameter matrix, and the generation time. Then, the scaling factor and the corrected camera intrinsic parameter matrix are extracted from the calibration file, and the case number, the hash value of the original image, the timestamp, and the block hash of the calibration file are extracted from the block index. Finally, the above... The extracted information is integrated according to a pre-defined structure. The case number is placed at the beginning of the report to clearly identify the corresponding case. The hash value of the original image is used to identify the original image associated with the report. The timestamp is used to record the time nodes of data collection and calibration. The scaling factor and the corrected camera intrinsic parameter matrix are used to explain the specific parameters of focal length correction. The block hash of the calibration file is used to verify the integrity of the calibration file. Finally, the integrated information is formed into a structured text, namely the arbitration data collection report. This report fully presents the key information of the entire process from original image collection to parameter calibration and evidence storage, providing traceable and verifiable evidence for arbitration review.

[0135] Through the above process, a full-chain management solution is finally formed, which includes original image acquisition, pseudo-focal length marking, scale correction, compliant evidence storage, and arbitration reports. This ensures the authenticity, traceability, and legal validity of traffic accident arbitration data, and provides reliable technical support for liability determination.

[0136] like Figure 2 The diagram shown is a functional block diagram of an intelligent arbitration data acquisition and management system based on a large model, provided by an embodiment of the present invention.

[0137] In this embodiment, the functions of each module / unit are as follows:

[0138] The image log construction module is used to build an image data log table based on the original images of the accident scene, realize the extraction and structured management of unique image identifiers, and form an accident scene image log table.

[0139] The pseudo-focal length labeling module, based on the image log table, performs visual feature mining and focal length labeling inference on the original image to obtain a set of candidate pseudo-focal length images for subsequent judgment and correction.

[0140] The reference object extraction module is used to identify target reference objects with calibrated physical properties in the candidate pseudo-focal length image set and to fuse pseudo-focal length images to construct triples.

[0141] The field of view calculation module is used to calculate the difference between the real field of view and the nominal field of view based on the triplet, identify real pseudo-focal length images, and calculate the scaling correction coefficient.

[0142] The inverse camera geometry correction module is used to perform inverse reconstruction of the camera intrinsic parameter matrix on the real pseudo-focal length image based on the obtained correction scale coefficient, output the corrected camera intrinsic parameter matrix, and generate a calibration file together with the original image hash value.

[0143] The evidence storage module is used to generate evidence information based on calibration files, the hash value of calibration files, timestamps, and the hash value of original images, and store it in the compliant evidence storage system. It also generates an arbitration data collection report of the accident scene based on the evidence information.

[0144] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent acquisition and management of arbitration data based on a large model, characterized in that, include: S1. Construct the original log record table based on the original images of the accident scene; S2. Based on the original log record table, perform pseudo-focal length marking on the original image to obtain candidate pseudo-focal length images; S3. Select an object with a standard physical length from the candidate pseudofocal length image as a reference target, and construct a triplet based on the reference target and the candidate pseudofocal length image; S4. Based on the true field of view and the nominal field of view of the triplet, identify the true pseudo-focal distance image among the candidate pseudo-focal distance images, and calculate the scaling factor of the true pseudo-focal distance image to obtain the correction scaling factor. S5. Based on the correction scaling factor, calibrate the original image to obtain the correction camera intrinsic parameter matrix. Package the correction camera intrinsic parameter matrix, the correction scaling factor, and the hash value of the original image into a calibration file. Specifically, the following steps are performed: image matching is performed between the real pseudo-focal length image and the original log record table to obtain the original image corresponding to the real pseudo-focal length image in the original log record table; the equivalent focal length in the original image is calibrated using the correction scaling factor to obtain the corrected camera intrinsic parameter matrix; the corrected camera intrinsic parameter matrix, the correction scaling factor, and the hash value of the original image are packaged into a calibration file. S6. Write the calibration documents and original images into the compliant evidence storage system, generate an arbitration data collection report of the accident scene based on the evidence storage information, and form a management plan for arbitration data.

2. The intelligent acquisition and management method for arbitration data based on a large model according to claim 1, characterized in that, An original log record table was constructed based on the original images of the accident scene, including: The original images of the accident scene are hashed to obtain their hash values. Extract metadata from the original image, including: equivalent focal length, ISO, exposure time, digital zoom ratio, and device model. The original log table is constructed based on the original image, the hash value of the original image, and the metadata of the original image.

3. The intelligent acquisition and management method for arbitration data based on a large model according to claim 1, characterized in that, Based on the original log record table, the original image is pseudo-focal length labeled to obtain candidate pseudo-focal length images, including: Read the information recorded in the original log record table; Low light determination is performed on the original image in the recorded information: if the sensitivity of the original image in the recorded information is greater than the preset sensitivity threshold and the exposure time is greater than the preset exposure time threshold, then it is determined that the original image in the recorded information has low light phenomenon. Long-focus determination is performed on the original image in the recorded information: if the equivalent focal length of the original image in the recorded information is greater than the preset focal length threshold, it is determined that the original image in the recorded information has a long-focus phenomenon. Cropping determination of the original image in the recorded information: If the digital zoom ratio of the original image in the recorded information is greater than the preset zoom ratio threshold, it is determined that the original image in the recorded information has been cropped. If the original image in the recorded information simultaneously exhibits low light, telephoto, and cropping phenomena, then the original image in the recorded information is marked with a pseudo-focal length, and the original image in the recorded information is merged into the candidate pseudo-focal length image set.

4. The intelligent acquisition and management method for arbitration data based on a large model according to claim 1, characterized in that, In the candidate pseudofocal length image, an object with a standard physical length is selected as the reference target. A triplet is constructed based on the reference target and the candidate pseudofocal length image, including: The trained target detection model is used to perform target detection on the candidate pseudofocal images in the candidate pseudofocal image set to obtain a reference target set for the candidate pseudofocal images. Select the target with the highest confidence level from the reference target set as the reference target; The physical length of the reference target is retrieved from the standard length library, and the pixel length of the reference target in the candidate pseudofocal length image is calculated. Construct triples based on candidate pseudofocal length images, physical length, and pixel length.

5. The intelligent acquisition and management method for arbitration data based on a large model according to claim 1, characterized in that, Based on the true and nominal field of view of the triplet, the true pseudo-focal distance image is identified among the candidate pseudo-focal distance images. The scaling factor of the true pseudo-focal distance image is then calculated to obtain the correction scaling factor, including: The true field of view of the candidate pseudofocal image is calculated based on the physical length, pixel length, and device model in the metadata from the triplet. The nominal field of view of the candidate pseudofocal image is calculated based on the equivalent focal length of the candidate pseudofocal image. The difference between the actual field of view and the nominal field of view is obtained. If the difference in field of view is greater than the preset difference threshold, the candidate pseudo-focal image is determined to be a real pseudo-focal image. The scaling factor is calculated for the true field of view and nominal field of view of the real pseudo-focal length image to obtain the correction scaling factor.

6. The intelligent acquisition and management method for arbitration data based on a large model according to claim 1, characterized in that, The calibration documents and original images are written into the compliant evidence storage system. Based on the evidence storage information, an arbitration data collection report of the accident scene is generated, forming a management plan for arbitration data, including: Evidence storage information is generated based on the hash value of the original image, the calibration file, the hash value of the calibration file, and the timestamp, and then stored in a compliant evidence storage system. An arbitration data collection report for the accident scene is generated based on the evidence information. The arbitration data collection report was analyzed and integrated to obtain a management plan for the arbitration data.

7. A large-scale model-based intelligent data acquisition and management system for arbitration, characterized in that, The intelligent data acquisition and management system includes: The image log construction module is used to build an image data log table based on the original images of the accident scene, realize the extraction and structured management of unique image identifiers, and form an accident scene image log table. The pseudo-focal length labeling module, based on the image log table, performs visual feature mining and focal length labeling inference on the original image to obtain a set of candidate pseudo-focal length images for subsequent judgment and correction. The reference object extraction module is used to identify target reference objects with calibrated physical properties in the candidate pseudo-focal length image set and to fuse pseudo-focal length images to construct triples. The field of view calculation module is used to calculate the difference between the real field of view and the nominal field of view based on the triplet, identify real pseudo-focal length images, and calculate the scaling correction coefficient. The inverse camera geometry correction module is used to perform inverse reconstruction of the camera intrinsic parameter matrix on the real pseudo-focal length image based on the obtained correction scale coefficient, output the corrected camera intrinsic parameter matrix, and generate a calibration file together with the original image hash value. The evidence storage module is used to generate evidence storage information based on calibration files, the hash value of calibration files, timestamps, and the hash value of original images, and store it in the compliant evidence storage system. It also generates an arbitration data collection report of the accident scene based on the evidence storage information.

Images