A method and equipment for intelligent securing of accident vehicles based on towing services

Abstract

This invention discloses an intelligent vehicle securing method and equipment for towing operations, belonging to the field of intelligent towing securing technology. It addresses the problem that current towing operations often rely on human experience for vehicle securing, lacking scientific loading and securing techniques for vehicles with special damage needs, and also presenting technical issues related to secondary safety hazards during vehicle transport. The method includes: comparing the real-time 3D architecture model of the accident vehicle with a standard vehicle model's 3D model to determine the loading risk level of the accident vehicle under relevant damage structures; identifying damage characteristics at multiple scales from a set of multi-angle images of the entire vehicle's damage, determining key damage structure data; performing factor influence contribution correlation processing based on damage characteristics on the key damage structure data to obtain structural safety data; and performing local area correction processing on the initial vehicle securing position data to determine reference vehicle securing position data.

Description

Technical Field

[0001] This application relates to the field of intelligent trailer securing, and more particularly to a method and device for intelligent securing of accident vehicles based on trailer services. Background Technology

[0002] As vehicles become increasingly common, the number of vehicles on the road is also increasing, which greatly increases the probability of traffic accidents. In some serious traffic accidents, vehicles are severely damaged, rendering them unable to drive, or even severely deformed and immobile, with serious damage to the vehicle frame and body structure. In such cases, to avoid disrupting normal traffic, towing services are provided, using tow trucks or flatbed trucks to remove the immobile accident vehicles as quickly as possible for subsequent repairs.

[0003] In current towing services, severely damaged or structurally deformed accident vehicles are often secured to flatbed trucks manually using chains, ropes, and other tools. This method suffers from low securing efficiency (requiring multiple people to work together), poor stability (prone to loosening due to bumps), and safety hazards (operators must be in close contact with the damaged vehicle). In some cases with significant damage or substantial deformation, statistical data shows that manual securing can lead to secondary damage (such as uneven stress at the securing points causing further deformation). Furthermore, manual securing is generally based on the operator's experience and judgment, which is insufficient for properly securing severely deformed vehicles.

[0004] Although adjustable pallets (such as those that expand the size of the main / sub pallet via a hydraulic controller) or wheel hub fixing devices (such as trailer locks with elastic connecting plates) exist, they do not specifically address the special needs of damaged vehicles. For example, they cannot accommodate vehicles with deformed wheel hubs or damaged suspension; they lack an automatic compensation mechanism for vehicle center of gravity shift; and the fixing process still requires manual assistance in positioning.

[0005] Traditionally, the method for securing severely deformed accident vehicles towed is based on human experience and judgment, combined with securing devices. However, current methods often rely on manual judgment, which is difficult for novice tow truck operators to master. It lacks effective auxiliary positioning tools for quickly securing severely damaged vehicles, and improper loading and securing can easily cause secondary damage during transport. Relying primarily on the experience of the loading operator to determine suitable securing points for the accident vehicle cannot provide customized safety measures for severely deformed vehicles, posing certain safety hazards during accident transport and hindering the safe and efficient execution of towing services. There is an urgent need for a loading and securing method that can assist tow truck operators in efficiently and quickly securing severely deformed vehicles. Summary of the Invention

[0006] This application provides an intelligent method and device for securing accident vehicles based on towing services, which addresses the following technical problems: In existing towing services, the securing of accident vehicles often relies on human experience to determine the securing position, making it difficult to reasonably and safely complete the loading and securing of deformed accident vehicles. It also lacks scientific loading and securing for vehicles with special damage needs and poses secondary safety hazards during the transportation of accident vehicles.

[0007] The embodiments of this application adopt the following technical solutions:

[0008] On one hand, this application provides an intelligent vehicle fixation method based on towing services, comprising: identifying damage characteristics at multiple scales from a set of vehicle damage images taken from multiple angles, and determining key damage structure data in each damage area; performing factor influence contribution correlation processing on the key damage structure data in each damage area based on damage characteristics, and evaluating and obtaining structural safety data for each damage area; performing fixation force balance calculation on the accident vehicle based on a real-time three-dimensional architecture model to obtain initial vehicle fixation position data; and using the structural safety data of each damage area, performing local area correction processing on the initial vehicle fixation position data based on the entire vehicle body to determine reference vehicle fixation position data, and displaying the reference vehicle fixation position data in the form of a three-dimensional graph, thereby providing a reference for the fixation method of the accident vehicle.

[0009] This application's embodiments, through multi-angle imaging and damage characteristic identification, can more accurately determine vehicle damage, thereby improving the assessment of vehicle structural safety during the fixation process. Furthermore, it can determine the loading risk level of the accident vehicle, especially for vehicles with high-risk loading levels, allowing for more stringent fixation measures to reduce the risk of accidents. It can also utilize deformable convolutional networks for multi-scale damage characteristic identification, more accurately identifying key damaged structures in each damaged area, providing a basis for subsequent fixation measures. Simultaneously, the SHAP algorithm is used to evaluate the factor contribution of key damaged structure data in each damaged area, enabling a more comprehensive assessment of the vehicle's structural safety data. Based on real-time 3D architecture models and structural safety data, the optimal vehicle fixation position can be calculated, ensuring optimal fixation effectiveness and stability. It can also periodically monitor the trailer fixation structure corresponding to the reference vehicle fixation position data, continuously generating loading and driving safety data, facilitating the monitoring and maintenance of vehicle safety status. Moreover, it can automate damage identification, risk level assessment, and optimal fixation position determination, reducing manual intervention and improving work efficiency.

[0010] In one feasible implementation, the method further includes the step of obtaining the real-time three-dimensional architecture model. This step specifically includes: acquiring and processing continuous image frames of the accident vehicle from all angles using a handheld visual sensor device to obtain an image set of the accident vehicle and a corresponding visual odometry; wherein the visual odometry is the cumulative pose estimation position and motion trajectory during the image frame acquisition process; constructing an initial three-dimensional structural model of the accident vehicle based on the visual odometry and the vehicle shape features in the image set; performing stereo matching cost aggregation calculation on the whole vehicle damage image set related to left and right visual differences to obtain a fitting matching cost contour curve of the whole vehicle damage image set; performing global loop detection on the fitting matching cost contour curve related to inter-frame pixel positioning changes, and constructing a three-dimensional map of damage transformation points corresponding to the whole vehicle damage image set based on constraint node parameters; and performing structural fusion processing between the three-dimensional map of damage transformation points and the initial three-dimensional structural model based on the coordinate positions of the same damage area in the whole vehicle damage image set to obtain the real-time three-dimensional architecture model.

[0011] This application embodiment improves the accuracy of data acquisition of the damaged structural area by comparing the real-time 3D model of the accident vehicle with the 3D model of a standard vehicle. Then, based on the non-overlapping area corresponding to the damaged structural area, the scale comparison between the 3D models is completed. Based on the volume of the non-overlapping area, the size of the actual damaged area is determined. Only then can it be finally calculated whether the accident vehicle belongs to the high-risk loading level and whether the optimal fixed position point should be calculated.

[0012] In one feasible implementation, a set of vehicle damage images captured from multiple angles of the accident vehicle is used to identify damage characteristics at multiple scales, determining key damage structure data in each damage region. Specifically, this includes: uploading the acquired set of vehicle damage images to a deformable convolutional network; wherein the set of vehicle damage images is a collection of images from several angles corresponding to each damage region; using the deformable convolutional network, the images in the damage image set under each damage region are adjusted to a uniform size to obtain damage images of the same size; the pre-acquired vehicle accident damage type images are processed to capture key points related to the damage type, determining the trend of graphic line changes; and based on the trend of graphic line changes, the blurred features in the damage images of the same size are effectively analyzed. Robustness enhancement calculations under graphic changes are performed to obtain a data-enhanced damage image. Type identification calculations related to vehicle structural damage deformation characteristics are then performed on the data-enhanced damage image to determine the damage characteristic data for each damage region. The damage characteristic data includes both visual and hidden damage characteristic data. Using a preset hybrid loss function, the focus type of the damage characteristic data for each damage region is evaluated based on key damage structures. Based on the location, segmentation, and classification of the damage target, key damage structure data for each damage region is identified and determined. The key damage structure data includes dent damage, crack damage, wrinkle damage, perforation damage, and structural loss damage. The hybrid loss function includes classification loss, bounding box loss, and mask loss.

[0013] This application embodiment identifies damage characteristics at multiple scales using a whole vehicle damage image set. Specifically, it utilizes a deformable convolutional network to further analyze the acquired damage structure images. Based on the trend of image line changes and the combined identification effect of intuitive and hidden damage characteristic data, it completes the localization, segmentation, and classification of damage targets under key damage structures. Ultimately, it identifies and determines the key damage structure data in each damage region, that is, the detailed damage structure type. This helps to clarify the severity of the damage region and the specific damage type, providing an accurate data foundation for subsequent data evaluation calculations.

[0014] In one feasible implementation, the data-enhanced damage image is subjected to type recognition calculations related to the vehicle structural damage deformation characteristics to determine the damage characteristics of each damage region. Specifically, this includes: performing feature capture processing on the data-enhanced damage image in the whole vehicle damage image related to dynamic convolution kernels and dynamic receptive fields, and performing secondary bounding box feature capture on the data-enhanced damage image based on an irregular vehicle damage shape template to obtain the intuitive damage characteristic data; performing score high and low marking processing on the score boundary overlapping boxes in the data-enhanced damage image, and performing hierarchical feature capture on all bounding boxes based on the marked score of each score boundary overlapping box to obtain the hidden damage characteristic data.

[0015] In one feasible implementation, before performing factor influence contribution correlation processing based on damage features on the key damage structure data in each damage region, and evaluating and obtaining the structural safety data of each damage region, the method further includes: performing feature capture processing on the data-enhanced damage image in the whole vehicle damage image related to dynamic convolution kernels and dynamic receptive fields, and performing secondary bounding box feature capture on the data-enhanced damage image based on an irregular vehicle damage shape template to obtain the intuitive damage characteristic data; performing score high and low marking processing on the score boundary overlapping boxes in the data-enhanced damage image in the whole vehicle damage image, and performing hierarchical feature capture on all bounding boxes based on the marked score of each score boundary overlapping box to obtain the hidden damage characteristic data.

[0016] This application's embodiments, by labeling the damage types of key damaged structural data in each damaged region, ensure accurate damage classification, facilitating subsequent detailed analysis. Numerical information corresponding to each damage type can also be recorded, providing foundational data for quantitative analysis. Then, the factor contribution value for each damage type is calculated, which helps understand the degree of impact of different damage types on overall structural safety. Simultaneously, by using the SHAP algorithm to calculate the weighted average of marginal contributions under feature combinations, the impact of each factor on structural safety can be measured more accurately, even when these factors interact. Average segmentation of damage types based on SHAP interaction values ​​allows for a fairer allocation of influence contributions, avoiding the possibility of a single damage type excessively affecting the assessment results. Finally, by obtaining the SHAP value for each damaged region, the impact of each damage type on structural safety can be more comprehensively assessed, providing a basis for fixed strategies.

[0017] In one feasible implementation, the key damaged structural data in each damaged area undergoes factor influence contribution correlation processing based on damage characteristics to assess and derive structural safety data for each damaged area. Specifically, this includes: establishing a structural safety assessment and prediction model; wherein the structural safety assessment and prediction model is used to perform comprehensive factor assessment processing on the key damaged structural data; using the structural safety assessment and prediction model, the contribution of the total SHAP value in each damaged area is calculated by summing the predicted values ​​of relevant model scores to obtain the percentage of accident damage factors for each damaged area; wherein the percentage of accident damage factors is positively correlated with the total SHAP value; and converting the percentage of accident damage factors into a numerical score to obtain structural safety data for each damaged area; wherein a higher score in the structural safety data corresponds to a more severe degree of damage.

[0018] This application's embodiments enable a comprehensive assessment of key structural damage data of accident vehicles, considering multiple influencing factors. The sum of the SHAP values ​​for each damage type in each damage region yields the total SHAP value, which helps identify the damage types with the greatest impact on structural safety. Furthermore, through interaction correlation analysis, the interactions between different damage types are understood, which is crucial for comprehending complex damage patterns and their impact on structural integrity. Simultaneously, the use of a TAN Bayesian network structure to represent factor interactions effectively captures complex dependencies between variables, improving the model's predictive ability. Based on the Bayesian network structure, the sum of the contribution of the total SHAP value for each damage region is calculated, helping to determine the importance of accident damage factors in structural safety and facilitating an understanding of the degree of impact of various damage factors on overall structural safety. Finally, through fractional numerical conversion, the percentages of accident damage factors are transformed into structural safety data, making the assessment results more intuitive and easier to understand.

[0019] In one feasible implementation, based on the real-time three-dimensional architecture model, a fixed force balance calculation is performed on the accident vehicle to obtain initial vehicle fixed position data. Specifically, this includes: performing a vehicle body attitude angle balance calculation on the real-time three-dimensional architecture model to obtain vehicle body attitude angle data; performing a static balance calculation on the real-time three-dimensional architecture model related to single-point load-bearing capacity to obtain single-point load-bearing capacity data; performing an automatic compensation calculation on the real-time three-dimensional architecture model related to vehicle center of gravity offset to obtain vehicle center of gravity balance data; determining the vehicle body attitude angle data, the single-point load-bearing capacity data, and the vehicle center of gravity balance data as overall constraints; and performing a nonlinear force balance calculation on the real-time three-dimensional architecture model related to fixed points of the vehicle body using the objective function corresponding to the overall constraints to obtain the initial vehicle fixed position data of the real-time three-dimensional architecture model.

[0020] In one feasible implementation, the initial vehicle fixed position data is corrected based on a local region correction across the entire vehicle body using the structural safety status data of each damaged area to determine reference vehicle fixed position data. Specifically, this includes: processing the structural safety status data using a dynamic grading strategy to obtain safety level regions; wherein the safety level regions include: red level regions, yellow level regions, and green level regions; discretizing the real-time 3D architecture model to establish a stiffness transfer function between nodes; and annotating the corresponding nodes in the stiffness transfer function based on the node positions corresponding to the safety level regions to obtain annotated mesh nodes; wherein... The labeled grid nodes contain three fixed meanings and correspond to the colors of the safety level areas; the initial vehicle fixed position data is loaded; a neighborhood scan of the global grid nodes is performed on each initial point in the initial vehicle fixed position data, and based on the node avoidance attributes of the labeled grid nodes in the local area, the corresponding neighboring nodes are calculated using the node Euclidean distance under the spherical correction domain to obtain the position offset vector of each initial point; the node avoidance attributes include: all avoidance attributes of red nodes and displacement compensation avoidance attributes of yellow nodes; through the position offset vector, all initial fixed points in the initial vehicle fixed position data are offset to obtain the corrected reference vehicle fixed position data.

[0021] This application embodiment uses structural safety data to perform localized correction of vehicle fixation positions, ensuring a direct correlation between the fixation position and vehicle structural safety, thus improving the safety of the fixation effect. The structural safety data is dynamically graded into red, yellow, and green level regions, facilitating rapid identification of regions with different safety risk levels. Furthermore, the entire vehicle is discretized using a NURBS surface mesh, which provides a precise geometric representation, beneficial for detailed mechanical analysis. Simultaneously, establishing a stiffness transfer function between nodes helps simulate the mechanical characteristics of the vehicle structure, providing a mechanical basis for fixation position correction. Annotated mesh nodes provide clear identification for regions of different safety levels, aiding operators in understanding and implementing fixation strategies. Avoidance attributes for red and yellow nodes are introduced; red nodes require complete avoidance, while yellow nodes allow for displacement compensation avoidance, enhancing the flexibility of the fixation strategy. Additionally, a global neighborhood scan of the initial vehicle fixation position data is performed, and local correction is conducted based on the avoidance attributes of the annotated mesh nodes, which helps optimize the fixation position. Finally, by calculating the position offset vector of each initial point, the fixed position can be precisely adjusted to ensure that the vehicle structure is not subjected to additional damage during the fixing process.

[0022] In one feasible implementation, after determining reference vehicle fixed position data by performing local area correction processing based on the structural safety status data of each damaged area and the initial vehicle fixed position data, the method further includes: loading and securing the accident vehicle to a trailer based on the reference vehicle fixed position data; wherein the trailer fixing structure includes at least: a binding fixing structure, auxiliary fixing buckles, and a vehicle limiter; placing pressure sensors at each fixing point in the reference vehicle fixed position data; periodically monitoring the pressure data of the trailer fixing structure through the pressure sensors, and obtaining feedback data from the trailer fixing structure; performing tabular data statistical processing on the feedback data from the trailer fixing structure to generate loading and driving safety data for monitoring the transportation process of the accident vehicle.

[0023] On the other hand, embodiments of this application also provide an intelligent vehicle securing device based on trailer services, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to execute an intelligent vehicle securing method based on trailer services as described in any of the above embodiments.

[0024] This application provides an intelligent method and equipment for securing accident vehicles based on towing services. Compared with the prior art, the embodiments of this application have the following beneficial technical effects:

[0025] 1. Improved safety: By using multi-angle photography and damage characteristic identification, the damage to the vehicle can be assessed more accurately, thereby improving the assessment of the vehicle's structural safety during the fixing process.

[0026] 2. Risk level classification: It can determine the loading risk level of accident vehicles, especially for vehicles with high loading risk level, and can take more stringent fixing measures to reduce the risk of accidents.

[0027] 3. Damage characteristic identification: Using deformable convolutional networks for multi-scale damage characteristic identification can more accurately determine the key damage structures in each damage region, providing a basis for subsequent fixation measures.

[0028] 4. Structural safety assessment: By using the SHAP algorithm to evaluate the contribution of key damage structural data in each damaged area, the structural safety status of the vehicle can be assessed more comprehensively.

[0029] 5. Intelligent fixed position determination: Based on real-time 3D architecture model and structural safety data, the optimal vehicle fixing position can be calculated to ensure the best fixing effect and stability.

[0030] 6. Periodic monitoring: Periodic monitoring of the trailer fixing structure corresponding to the reference vehicle fixed position data can continuously generate loading and driving safety data, which facilitates the monitoring and maintenance of vehicle safety status.

[0031] 7. Improve work efficiency: It can automatically identify damage, assess risk levels, and determine the optimal fixation location, reducing manual intervention and improving work efficiency. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0033] Figure 1 A flowchart illustrating an intelligent method for securing accident vehicles based on towing services, provided in this application embodiment;

[0034] Figure 2 This application provides a schematic diagram of the structure of a standard vehicle model in three dimensions, as shown in the embodiments of the present application.

[0035] Figure 3 This is a top view schematic diagram of the optimal fixing position of a trailer flatbed provided in an embodiment of this application;

[0036] Figure 4 This is a schematic diagram of the structure of an intelligent vehicle securing device for accident vehicles based on towing services, provided in an embodiment of this application. Detailed Implementation

[0037] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0038] This application provides an intelligent method for securing accident vehicles based on towing services, such as... Figure 1 As shown, the intelligent fixation method for accident vehicles based on towing services specifically includes steps S101-S106:

[0039] S101. By comparing the real-time three-dimensional architecture model of the accident vehicle with the three-dimensional model of the standard vehicle, the loading level under the relevant damage structure is determined to obtain the loading risk level of the accident vehicle.

[0040] It should be noted that visual SLAM has introduced a more advanced map form in the visual mapping process, namely, by constructing a 3D animated map that can display models of all objects within the field of view, assisting the robot in identification. Since the above-mentioned visual SLAM method is only a basic architecture, this application designs a binocular visual SLAM system and a 3D animation modeling method based on the SMA algorithm to further refine visual SLAM. Firstly, the visual sensor uses the STEREOLABS ZED2 binocular stereo camera, which can fully utilize the advantages of binocular cameras and has better applicability in a wide range of application scenarios. Because the camera introduces various optical components, distortion occurs during imaging; therefore, calibration is required. The efficient and practical Zhang Zhengyou calibration method is adopted, which overcomes the shortcomings of traditional calibration methods that require high-precision calibration objects. Calibration can be completed with only a printed checkerboard pattern. Compared with self-calibration, it improves accuracy and is easier to operate.

[0041] Specifically, the process begins with using modern visual SLAM algorithms and a handheld visual sensor device (either a dedicated industrial camera or a fixed camera app on a mobile phone) to continuously acquire and process image frames of the accident vehicle from all angles, resulting in an image set and corresponding visual odometry. The visual odometry estimates the cumulative pose and motion trajectory during the image frame acquisition process.

[0042] Furthermore, based on visual odometer data and vehicle shape features from the accident vehicle image set, an initial three-dimensional structural model of the accident vehicle was constructed.

[0043] In one embodiment, a worker in a towing operation holds a handheld visual sensor device, which can be a camera or other visual sensor. The visual sensor then captures continuous image frames of the accident vehicle from all angles. The captured image frames are then preprocessed, such as denoising and color correction. A modern visual SLAM algorithm is then applied to extract feature points from the image frames and calculate pose changes between consecutive frames to obtain a visual odometry. During the construction of the 3D structural model, the vehicle's shape features need to be extracted from the image set first. Then, based on the visual odometry and the vehicle's shape features, an initial 3D structural model of the accident vehicle can be constructed.

[0044] Furthermore, by using a pre-set binocular camera, the stereo matching cost aggregation calculation is performed on the whole vehicle damage image set of the accident vehicle under the relevant left and right visual differences, and the fitting matching cost contour curve of the whole vehicle damage image set is obtained.

[0045] Furthermore, global loop detection is performed on the fitted matching cost contour curve under the relevant inter-frame pixel positioning changes, and a three-dimensional map of damage transformation points corresponding to the whole vehicle damage image set is constructed based on the constraint node parameters.

[0046] In one embodiment, a pre-defined binocular camera system is also required. Next, stereo matching is performed on the entire vehicle damage image set based on left-right visual difference. Then, the cost of stereo matching is calculated and aggregated to obtain a fitted matching cost contour curve. Next, global loop detection is performed on the fitted matching cost contour curve to identify inter-frame pixel location changes. Finally, based on the constraint node parameters, a 3D map of damage transformation points corresponding to the entire vehicle damage image set is constructed.

[0047] As a feasible implementation method, after the above process is completed, stereo matching is required, which is an extremely dense task. It requires obtaining the disparity of almost all pixels in the left view image based on the corresponding points of the left and right eyes, and then obtaining a dense disparity map. The SMA algorithm modeling process used in this study mainly consists of the following four parts: calculating the matching cost, matching cost aggregation, disparity calculation and optimization, and disparity refinement. First, in the part of calculating the cost of stereo matching, it is essentially calculating the disparity grayscale similarity between two images. This study uses the absolute value of grayscale difference (AVGD) for calculation. Second, matching cost aggregation is calculated. Currently, there are two main types: global cost clustering methods and local methods. The former obtains an energy function through the initial matching cost and then optimizes the above function to the global minimum. The latter improves the reliability of the matching cost clustering operation by superimposing processing windows. Then, stereo vision disparity acquisition and optimization operations are performed, usually by two types: local SMA algorithms based on window methods and global SMA algorithms. The former picks the optimal point within the window frame after stereo matching cost aggregation. The latter requires first designing an energy evaluation function, and then calculating the minimum energy value using an optimization method to obtain the best disparity matching, which is the pixel matching relationship that minimizes the above function.

[0048] As a feasible implementation method, the trailer operator controls the camera movement to determine whether the estimated pose error exceeds a set threshold. If it does, a repositioning operation is required; otherwise, the subsequent steps can proceed. A randomized fern algorithm is used to perform global loop detection. If a global loop is found, the repositioning and tracking algorithm from the previous step estimates the pose changes between frames. Simultaneously, a small number of points are uniformly and randomly sampled, and the node parameters in the graph are optimized by optimizing the constraint equations. Otherwise, the system continues to detect local loops. If a local loop is detected, a relative pose estimation step is performed, constraints are constructed, and node parameters are optimized. Finally, the transformed pose is determined, i.e., a 3D map of the damage transformation points corresponding to the vehicle damage image set is constructed.

[0049] As a feasible implementation method, data acquisition utilizes a visual sensor: an Intel RealSense D455 binocular depth camera (global shutter, resolution 1280×720@30fps, depth range 0.5-6m, FOV 86°×57°) and an inertial measurement unit (IMU): a built-in 6-axis IMU (accelerometer ±4g, gyroscope ±1000dps). Workers walk around the accident vehicle at a constant speed (≤0.5m / s), maintaining the camera 1.2-1.8m from the vehicle surface. A spiral path is used to cover the entire surface of the vehicle (pitch angle ±30°, yaw angle 360° continuous rotation). The acquisition time is 120-180 seconds, obtaining approximately 3600-5400 frames of RGB-D images. Next, nonlocal mean filtering (parameter h=7, search_window=21) is used for distortion correction, using the Brown-Conrady model (k1=-0.15, k2=0.03, p1=p2=0). Additionally, the SLAM algorithm is used: 1) ORB feature extraction: 1000 feature points are extracted per frame (scale pyramid 8 layers, scale factor 1.2).

[0050] 2) Pose estimation: based on PnP+RANSAC (500 iterations, reprojection error threshold of 2.5 pixels).

[0051] 3) Loop closure detection: DBoW2 bag-of-words model (ORB vocabulary tree with 10^6 nodes).

[0052] As a feasible implementation method, semantic segmentation is first performed in shape feature extraction: DeepLabV3+ (ResNet101 backbone, pre-trained Cityscapes model). Then, a vehicle mask (IoU ≥ 0.85) is extracted, and model reconstruction is performed using the Canny operator (dual thresholds 50 / 150). Specifically, TSDF (Truncation Signed Distance Function) fusion is first performed, combined with a voxel resolution of 5mm×5mm×5mm and a truncation distance of δ = 15cm. The input is a damaged area image acquired by a stereo camera (resolution 1280×720, baseline 75mm). Cost calculation is performed using Census transform (9×9 window). Cost aggregation is then performed using SGM (Semi-Global Matching), combined with parameters P1 = 10, P2 = 120 (penalty coefficients), and 8 path directions. Disparity optimization is then performed using sub-pixel fitting (quadratic curve interpolation). Finally, a matching cost contour curve is fitted, and the disparity-cost function and polynomial fitting are calculated for each pixel. The minimum point is then extracted. Finally, using the damage transformation point map: sparse 3D point set (density ≥ 200 points / ㎡) and damage quantification indicators: maximum indentation depth and deformation volume (accuracy ± 2mm), the three-dimensional map of damage transformation points is constructed.

[0053] Furthermore, based on the coordinates of the same damage area in the whole vehicle damage image set, the three-dimensional map of the damage transformation point is fused with the initial three-dimensional structural model to obtain a real-time three-dimensional architecture model.

[0054] Furthermore, the vehicle database is used to query the vehicle model information corresponding to the accident vehicle. Based on the vehicle model information, a corresponding standard vehicle 3D model is obtained.

[0055] Furthermore, the real-time 3D architecture model is compared with the standard vehicle 3D model at the same scale to determine the volume of the non-overlapping area. Finally, based on the proportion of the non-overlapping area volume to the standard vehicle 3D model, the damage structure of the accident vehicle is manually assessed to obtain the loading risk level. The loading risk levels include: ordinary loading risk level, slightly dangerous loading risk level, and high-risk loading risk level.

[0056] In one embodiment of the present invention, the loading risk level of the accident vehicle can also be obtained by comparing the real-time three-dimensional architecture model of the accident vehicle with the three-dimensional model of the standard vehicle to determine the loading level under the relevant damage structure.

[0057] In one embodiment, Figure 2 This is a structural schematic diagram of a standard vehicle model 3D model provided in an embodiment of this application, such as... Figure 2As shown, based on the coordinates of the same damaged area in the whole vehicle damage image set, the 3D map of the damage transformation points is structurally fused with the initial 3D structural model to obtain a real-time 3D architecture model. That is, deformation fusion is performed on the corresponding areas of the initial 3D structural model, and the 3D map of the damage transformation points is fused into the initial 3D structural model, thus obtaining a real-time 3D architecture model that reflects the damage situation of the accident vehicle. Then, the vehicle model information corresponding to the accident vehicle is queried from the vehicle database. Based on the vehicle model information, the corresponding standard vehicle model 3D model is obtained. The volume of the non-overlapping area is determined based on same-scale comparison and non-overlapping area volume calculation. Finally, based on the proportion of the non-overlapping area volume to the standard vehicle model 3D model, the loading level of the accident vehicle under the damaged structure is determined, i.e., the loading risk level (ordinary, minor, high risk).

[0058] S102. If the loading risk level is determined to be high-risk, a deformable convolutional network is used to identify the damage characteristics at multiple scales in the set of vehicle damage images taken from multiple angles of the accident vehicle, and to determine the key damage structure data in each damage area.

[0059] Specifically, if a user determines that the loading risk level is high-risk, the collected set of vehicle damage images is then uploaded to the deformable convolutional network. The set of vehicle damage images is a collection of images from several angles corresponding to each damaged area.

[0060] Furthermore, a deformable convolutional network is needed to resize the images in the damage image set for each damage region to obtain damage images of the same size.

[0061] Furthermore, the pre-acquired vehicle accident damage type images are processed by capturing key points related to the damage type to determine the trend of graphic line changes. Based on the trend of graphic line changes, robustness enhancement calculations are performed on blurred features in damage images of the same size under graphic changes to obtain data-enhanced damage images.

[0062] Furthermore, the data-enhanced damage images in the whole vehicle damage images are subjected to type identification calculations based on the vehicle structural damage deformation characteristics to determine the damage characteristic data of each damage area. The damage characteristic data includes both visual damage characteristic data and hidden damage characteristic data.

[0063] In one embodiment, loss recognition of a whole vehicle damage image set can utilize deformable convolutional networks (DCNs) as the backbone of the computational model, combined with various optimization strategies such as multi-scale training, ALBU data augmentation, Focal Loss, and Soft NMS, ultimately achieving accurate identification of key structures in the accident vehicle damage area. The deformable convolutional network can be trained by inputting pre-processed images. Compared to traditional convolution, deformable convolution can adaptively adjust the size and shape of its kernel and receptive field, enabling it to more flexibly capture complex object shapes, such as irregular vehicle damage. Soft non-maximum suppression (NFS) can handle situations where multiple vehicle damage targets overlap in the same area. NFS is used during model testing to reduce the risk of falsely discarding overlapping targets, thereby improving detection accuracy and recall. Regarding the calculation of the multi-task focus loss function, the model's predicted results need to be compared with the true labels during training. Therefore, a hybrid loss function is used, including classification loss, bounding box loss, and mask loss, to evaluate the accuracy of the model's predictions from multiple perspectives.

[0064] As a feasible implementation method, deformable convolutional kernels in a deformable convolutional network are used to perform feature capture processing on the data-enhanced damage image of the whole vehicle damage image, involving dynamic convolutional kernels and dynamic receptive fields. Based on an irregular vehicle damage shape template, secondary bounding box feature capture is performed on the data-enhanced damage image to obtain intuitive damage characteristic data. This intuitive damage characteristic data consists of single-structure damage deformation data of the vehicle. Then, according to a soft non-maximum suppression bounding box algorithm, the score-high and low score markings of the overlapping bounding boxes in the data-enhanced damage image are performed. Based on the marked score of each overlapping bounding box, hierarchical feature capture is performed on all bounding boxes to obtain hidden damage characteristic data. This hidden damage characteristic data consists of overlapping compression deformation data under the multi-structure association of the vehicle.

[0065] In one embodiment, in deformable convolution, the receptive field can dynamically adjust its size and shape according to the characteristics of the input data, thus adapting more flexibly to different image features. This deformability allows the convolution window to adaptively deform, more accurately and effectively covering the shape of the target, especially irregularly shaped targets. In other words, data augmentation of damaged images involves feature capture processing related to dynamic convolution kernels and dynamic receptive fields, completing the analysis of salient target detection results, and identifying many irregularly shaped targets in vehicle damage. Deformable convolutional networks, by dynamically adjusting the convolution kernel, can more effectively capture and learn the morphological features of these irregular targets, improving the accuracy of damage recognition and segmentation. This allows for feature capture processing of deformable data on single-structure vehicle damage, yielding intuitive damage characteristic data.

[0066] In one embodiment, the soft nonmaximum suppression bounding box selection algorithm is an improved bounding box filtering algorithm. Unlike traditional nonmaximum suppression, soft nonmaximum suppression does not completely remove other boxes that overlap with high-scoring bounding boxes, but rather reduces the scores of these boxes based on the degree of overlap. This is particularly important for vehicle damage recognition tasks, because during data collection and annotation, it can be observed that different vehicle damage categories may have overlapping locations. Soft nonmaximum suppression can reduce the risk of mistakenly discarding overlapping targets, thereby improving detection accuracy and recall. Specifically, by capturing layered features under structural overlap in overlapping extrusion deformation data associated with multiple vehicle structures, hidden damage characteristic data is obtained, thus completing the identification of the main types of core damage characteristic structures in the accident vehicle damage area.

[0067] Furthermore, using a pre-defined hybrid loss function, the focus type under key damage structures is evaluated on the damage characteristic data of each damaged region. Based on the localization, segmentation, and classification of the damaged target, the key damage structure data of each damaged region is identified and determined. These key damage structure data include: dent damage, crack damage, wrinkle damage, perforation damage, and structural loss damage. The hybrid loss function includes: classification loss, bounding box loss, and masking loss. In other words, to alleviate the imbalance in the difficulty of classifying various types of damage in the image dataset, focus loss is introduced into the damage classification subtask. The core idea of ​​focus loss is to reduce attention to easily classified samples and focus on difficult-to-classify samples, i.e., assigning greater weight to samples with misclassification or high uncertainty. Then, the base classification loss task, bounding box loss task, and masking loss task evaluate the focus type under key damage structures on the damage characteristic data of each damaged region, ultimately completing the instance segmentation task and identifying and determining the key damage structure data of each damaged region.

[0068] S103. Perform factor influence contribution correlation processing on the key damaged structural data in each damaged area based on damage characteristics, and evaluate and obtain the structural safety data of each damaged area.

[0069] Specifically, the damage types of key damaged structural data in each damaged region are first labeled and the corresponding numerical information for each damage type is recorded. Then, the contribution value of each damage type and its corresponding numerical information in the current damaged region is calculated to obtain the factor influence contribution value of each damage type in each damaged region.

[0070] It should be noted that SHAP (Shapley Additive Explanation) is an algorithm based on cooperative game theory used to explain the influence of factors on model prediction. It approximates the output of the prediction model as the sum of the contributions of each input element, with each feature associated with a contribution value.

[0071] Furthermore, the SHAP algorithm is used to calculate the marginal contribution weighted average of the contribution value of each factor in each damaged area under feature combination, so as to obtain the marginal contribution weighted average based on the contribution value of each factor.

[0072] Furthermore, it is necessary to perform cross-correlation calculations on the weighted average marginal contribution of each damage type within the same damage region to obtain the SHAP cross-correlation value. Based on the SHAP cross-correlation value between any two damage types, an average segmentation process is performed between each damage type to obtain the SHAP value for each damage type within each damage region. In other words, the SHAP algorithm calculates the weighted average marginal contribution, which is the SHAP value after considering the synergistic effect of feature combinations. The SHAP value is the weighted average marginal contribution of each feature across all possible feature combinations.

[0073] Furthermore, a structural safety assessment and prediction model needs to be established. This model is used to comprehensively evaluate key damaged structural data. This structural safety assessment and prediction model can be constructed using expert scoring and predictive estimation to create a score-based assessment and prediction model.

[0074] Furthermore, the SHAP values ​​for each damage type in each damage region are summed to obtain the total SHAP value for each damage region. Then, an interaction correlation analysis of factor nodes is performed on the branch SHAP values ​​in the total SHAP value to obtain the factor interaction relationships between each damage type. Here, the factor interaction relationships refer to the interaction relationships under different damage types.

[0075] Furthermore, using a structural safety assessment prediction model, the contribution of relevant model score predictions to the total SAP value in each damaged area is calculated to obtain the percentage of accident damage factors for each damaged area. The percentage of accident damage factors is positively correlated with the total SAP value. Finally, the percentage of accident damage factors is converted into numerical scores to obtain the structural safety status data for each damaged area. A higher structural safety status data corresponds to a more severe degree of damage.

[0076] In one embodiment, a structural safety assessment and prediction model can be trained based on the key structural damage data of each damaged area obtained above, including damage characteristics and related factors, combined with machine learning algorithms (such as random forests and gradient boosting trees). This model can predict structural safety data based on the features. Then, the SHAP algorithm can be applied to the data of each damaged area to calculate the contribution of each factor to the structural safety data. The total SHAP value of each damaged area is then converted into a percentage, representing the contribution of the accident damage factors to that area. The percentage of accident damage factors is then numerically converted to obtain the structural safety data for each damaged area. Finally, based on the scores in the structural safety data, the degree of damage to each damaged area is assessed. This data can also be visualized for towing personnel, further assisting and facilitating their subsequent manual restraint.

[0077] S104. Based on the real-time three-dimensional architecture model, perform fixed force balance calculations on the accident vehicle to obtain initial vehicle fixed position data.

[0078] Specifically, firstly, the vehicle body attitude angle balance calculation is performed on the real-time 3D architecture model to obtain vehicle body attitude angle data. Next, static balance calculations related to single-point load-bearing capacity are performed on the real-time 3D architecture model to obtain single-point load-bearing capacity data. Finally, automatic compensation calculations related to vehicle body center of gravity offset are performed on the real-time 3D architecture model to obtain vehicle body center of gravity balance data.

[0079] Furthermore, the vehicle body attitude angle data, single-point load capacity data, and vehicle center of gravity balance data obtained above are respectively determined as the total constraint conditions. That is, the total constraint conditions of the objective function need to be calculated in advance, which is beneficial to the subsequent nonlinear calculation under the fixed force of the whole vehicle.

[0080] Furthermore, by combining the objective function corresponding to the overall constraint conditions, nonlinear force balance calculations are performed on the fixed points of the vehicle body in the real-time three-dimensional architecture model. Based on the fitness function, the stress concentration factor, displacement compensation amount and number of support points are comprehensively evaluated to generate the initial vehicle fixed position data of the real-time three-dimensional architecture model.

[0081] In one embodiment, the following steps can be taken to obtain the initial vehicle fixed position data:

[0082] (1) First, establish a multi-constraint optimization model: First, determine the total constraints: body posture angle data, single-point load capacity data of the vehicle body, and body center of gravity balance data are determined as the total constraints. Then, construct the objective function.

[0083] (2) An improved particle swarm optimization algorithm is used for solving the problem: First, 200 particles are initialized, each representing a set of fixed point coordinates (x, y, z). Then, the stress concentration factor, displacement compensation, and number of support points are comprehensively evaluated based on the fitness function. Finally, a simulated annealing mechanism is introduced, and when the number of iterations > 500, a suboptimal solution is accepted with a probability of 0.95^k.

[0084] (3) Calculation of initial vehicle fixed position data: Based on the above steps, the initial fixed force balance calculation is performed on the real-time three-dimensional architecture model. That is, the nonlinear force balance analysis of the accident vehicle body in the real-time three-dimensional architecture model is completed by using the multi-constraint optimization model and the improved particle swarm algorithm, so as to obtain a relatively general analysis of the initial vehicle fixed position data, that is, the simplified analysis of the vehicle fixed position data without considering the structural safety data of each damaged area, which includes multiple initial fixed position points.

[0085] S105. Using the structural safety data of each damaged area, the initial vehicle fixed position data is corrected locally based on the entire vehicle body to determine the reference vehicle fixed position data.

[0086] Specifically, the structural safety data is first processed using a dynamic grading strategy to obtain safety level regions. This involves dynamically segmenting the structural safety data without exceeding a threshold, with the safety level regions including: red, yellow, and green levels.

[0087] Furthermore, the real-time 3D architecture model is discretized for the entire vehicle, and a stiffness transfer function between nodes is established. Based on the node positions corresponding to the safety level regions, the corresponding nodes in the stiffness transfer function are labeled to obtain labeled mesh nodes. These labeled mesh nodes have three fixed meanings and correspond to the colors of the safety level regions.

[0088] Further, the initial vehicle fixed position data is first loaded. Then, a neighborhood scan of all grid nodes is performed on each initial point in the initial vehicle fixed position data. Based on the node avoidance attributes of the labeled grid nodes in the local area, the Euclidean distance of the corresponding neighboring nodes under the spherical correction domain is calculated to obtain the position offset vector of each initial point. The node avoidance attributes include: all avoidance attributes of red nodes and displacement compensation avoidance attributes of yellow nodes. That is, certain nodes need to be avoided or have their displacement compensated.

[0089] Furthermore, by using position offset vectors, all initial fixed points in the initial vehicle fixed position data are offset to obtain corrected reference vehicle fixed position data. Based on the reference vehicle fixed position data, a three-dimensional view of the accident vehicle's fixed position is formed. Using this three-dimensional view containing the fixed position, towing personnel can be assisted in securing the accident vehicle body by referring to the calculated fixed position.

[0090] In one embodiment, Figure 3 This application provides a top view of the optimal fixing position of a trailer flatbed truck, as shown in the embodiment of the present application. Figure 3 As shown, the process of determining the fixed position data of the reference vehicle mainly includes:

[0091] Step 1: Multi-dimensional calculation:

[0092] (1) Establish a safety assessment matrix for the damaged area: Define a five-dimensional evaluation vector S = [K_d, ΔK, μ, A_loss, δ_max], corresponding to the quantitative parameters of five damage types (dent damage, crack damage, wrinkle damage, perforation damage, and structural loss damage), where K_d represents dent damage, ΔK represents perforation damage, μ represents crack damage, A_loss represents minimum wrinkle damage, and δ_max represents maximum crack damage. Then, the weight coefficients are determined using the Analytic Hierarchy Process (AHP): ω = [0.25, 0.30, 0.15, 0.20, 0.10] (verified by experts); then, calculate the area safety score: Score_i = Σ(ω_j × S_ij) / Σω_j, normalized to a 0-100 score scale, where i and j represent the damage types present in each area, and S is the number of combined damage types.

[0093] (2) Dynamic grading strategy: Red zone (Score < 40): Direct fixing is prohibited, or cross-zone support frames can be installed. Yellow zone (40 ≤ Score < 70): Fixing is allowed but displacement compensation is required. Green zone (Score ≥ 70): Direct fixing is allowed. In other words, the dynamic grading strategy calculation under the node avoidance attribute has been completed.

[0094] Step 2: Global grid-based correction processing:

[0095] (a) Constructing a global stress transfer model for the vehicle body: Discretize the entire vehicle into a NURBS surface mesh (basic resolution 10mm×10mm). Then establish the stiffness transfer function between nodes.

[0096] (b) Execute the local correction algorithm: First, load the initial fixed point set P_initial = {p_1, p_2, ..., p_n}. Then, perform a neighborhood scan for each point p_k ∈ P_initial: establish a spherical correction domain with radius R = 300 mm centered at p_k. Next, extract the safety score Score_m of all grid nodes within the domain. Then, calculate the position offset vector. Finally, generate the corrected position set P_corrected = {p_k + Δp_k}.

[0097] (c) Perform coordinate position data parsing on the corrected position set and mark it in the corresponding position in the real-time three-dimensional architecture model to form the reference vehicle fixed position data of the current accident vehicle.

[0098] S106. Furthermore, the feedback data from the trailer fixing structure corresponding to the reference vehicle's fixed position data can be periodically monitored to generate loading and driving safety data. This means continuously monitoring the safety and stability of the fixed accident vehicle during transportation.

[0099] Specifically, based on reference vehicle positioning data, the accident vehicle is first loaded and secured to a trailer. The trailer securing structure includes at least: a binding and securing structure, auxiliary securing clips, and vehicle limiters.

[0100] Furthermore, pressure sensors are placed at each fixed point in the reference vehicle's fixed position data. The pressure sensors periodically monitor the pressure data of the trailer's fixed structure, and feedback data from the trailer's fixed structure is obtained.

[0101] Furthermore, the feedback data from the trailer's securing structure is tabulated and statistically processed to generate loading and driving safety data for monitoring the transport of accident vehicles. This loading and driving safety data can be periodically fed back to the mobile devices of the trailer workers, enabling safety monitoring of accident vehicles during transport and preventing problems such as stress fatigue or loosening of the straps or securing structures during transport.

[0102] In addition, this application also provides an intelligent vehicle securing device for accident vehicles based on towing services, such as... Figure 4 As shown, the intelligent vehicle securing device 400 based on towing services specifically includes:

[0103] At least one processor 401. And a memory 402 communicatively connected to the at least one processor 401. The memory 402 stores instructions executable by the at least one processor 401, enabling the at least one processor 401 to execute:

[0104] By comparing the real-time 3D architecture model of the accident vehicle with the 3D model of the standard vehicle, the loading level under the relevant damage structure is determined, and the loading risk level of the accident vehicle is obtained.

[0105] If the loading risk level is high-risk, then a deformable convolutional network is used to identify the damage characteristics of the whole vehicle damage image set taken from multiple angles of the accident vehicle at multiple scales, and to determine the key damage structure data in each damage area.

[0106] The key damaged structural data in each damaged area are processed based on the factor influence contribution correlation under the damage characteristics to evaluate and obtain the structural safety data of each damaged area.

[0107] Based on a real-time 3D architecture model, a fixed force balance calculation is performed on the accident vehicle to obtain initial fixed position data of the vehicle.

[0108] By using the structural safety data of each damaged area, the initial vehicle fixed position data is corrected locally based on the entire vehicle body to determine the reference vehicle fixed position data.

[0109] This application's embodiments, through multi-angle imaging and damage characteristic identification, can more accurately determine vehicle damage, thereby improving the assessment of vehicle structural safety during the fixation process. Furthermore, it can determine the loading risk level of the accident vehicle, especially for vehicles with high-risk loading levels, allowing for more stringent fixation measures to reduce the risk of accidents. It can also utilize deformable convolutional networks for multi-scale damage characteristic identification, more accurately identifying key damaged structures in each damaged area, providing a basis for subsequent fixation measures. Simultaneously, the SHAP algorithm is used to evaluate the factor contribution of key damaged structure data in each damaged area, enabling a more comprehensive assessment of the vehicle's structural safety data. Based on real-time 3D architecture models and structural safety data, the optimal vehicle fixation position can be calculated, ensuring optimal fixation effectiveness and stability. It can also periodically monitor the trailer fixation structure corresponding to the reference vehicle fixation position data, continuously generating loading and driving safety data, facilitating the monitoring and maintenance of vehicle safety status. Moreover, it can automate damage identification, risk level assessment, and optimal fixation position determination, reducing manual intervention and improving work efficiency.

[0110] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0111] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0112] The above description is merely an embodiment of this application and is not intended to limit this application. For those skilled in the art, various modifications and variations can be made to the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the embodiments of this application should be included within the scope of the claims of this application.

Claims

1. A method for intelligently securing accident vehicles based on towing services, characterized in that, The method includes: From a dataset of vehicle damage images taken from multiple angles, damage characteristics at multiple scales are identified to determine key damage structure data for each damage region, specifically including: The collected images of the damage to the accident vehicle are uploaded to a deformable convolutional network; wherein, the set of images of the damage to the vehicle is a collection of several angle images corresponding to each damage area. By using a deformable convolutional network, the images in the damage image set under each damage region are adjusted to a uniform size to obtain damage images of the same size. The pre-acquired vehicle accident damage type image is processed by capturing key points related to the damage type to determine the trend of graphic line changes; and based on the trend of graphic line changes, robustness enhancement calculations are performed on the blurred features in the damage image of the same size under graphic changes to obtain a data-enhanced damage image. The data-enhanced damage image is used to perform type identification calculations based on the vehicle structure damage deformation characteristics to determine the damage characteristic data for each damaged area; wherein, the damage characteristic data includes: visual damage characteristic data and hidden damage characteristic data, specifically including: The data-enhanced damage image in the whole vehicle damage image is subjected to feature capture processing related to dynamic convolution kernel and dynamic receptive field, and based on the irregular vehicle damage shape template, the data-enhanced damage image is subjected to secondary bounding box feature capture to obtain the intuitive damage characteristic data. The score-high and low marking processing is performed on the score boundary overlapping boxes in the data augmentation damage image of the whole vehicle damage image, and the hierarchical feature capture is performed on all boundary boxes based on the marked score of each score boundary overlapping box to obtain the hidden damage characteristic data. Using a preset hybrid loss function, the focus type of the damage characteristic data of each damage region is evaluated under the relevant key damage structures. Based on the localization, segmentation, and classification of the damage target, the key damage structure data of each damage region is identified and determined. The key damage structure data includes: dent damage, crack damage, wrinkle damage, perforation damage, and structural loss damage. The hybrid loss function includes: classification loss, bounding box loss, and mask loss. The key damaged structural data in each damaged area are subjected to factor influence contribution correlation processing based on damage characteristics to evaluate and obtain the structural safety data of each damaged area. Based on a real-time 3D architecture model, a fixed force balance calculation is performed on the accident vehicle to obtain initial fixed position data of the vehicle, specifically including: The vehicle body attitude angle balance calculation is performed on the real-time three-dimensional architecture model to obtain vehicle body attitude angle data; Static balance calculations of single-point load-bearing capacity are performed on the real-time three-dimensional architecture model to obtain single-point load-bearing capacity data of the vehicle body. Automatic compensation calculations are performed on the real-time three-dimensional architecture model to compensate for the offset of the vehicle's center of gravity, thereby obtaining the vehicle's center of gravity balance data. The vehicle body attitude angle data, the vehicle body single-point load-bearing capacity data, and the vehicle body center of gravity balance data are respectively determined as the overall constraint conditions; By using the objective function corresponding to the overall constraint conditions, nonlinear force balance calculations are performed on the real-time three-dimensional architecture model for the fixed points of the vehicle body, and the initial vehicle fixed position data of the real-time three-dimensional architecture model is obtained. Using the structural safety data of each damaged area, the initial vehicle fixed position data is corrected for local areas based on the entire vehicle body to determine the reference vehicle fixed position data, specifically including: The structural safety data is processed using a dynamic classification strategy to obtain safety level regions; wherein, the safety level regions include: red level region, yellow level region, and green level region; The real-time 3D architecture model is discretized for the whole vehicle, and a stiffness transfer function between nodes is established. Based on the node positions corresponding to the safety level regions, the corresponding nodes in the inter-node stiffness transfer function are labeled to obtain labeled mesh nodes; wherein, the labeled mesh nodes have three fixed meanings and correspond to the colors of the safety level regions; Load the initial vehicle fixed position data; For each initial point in the initial vehicle fixed position data, a neighborhood scan of the global grid nodes is performed. Based on the node avoidance attributes of the labeled grid nodes in the local area, the Euclidean distance of the corresponding neighboring nodes under the spherical correction domain is calculated to obtain the position offset vector of each initial point. The node avoidance attributes include: all avoidance attributes of red nodes and displacement compensation avoidance attributes of yellow nodes. The position offset vector is used to offset all the initial fixed points in the initial vehicle fixed position data to obtain the corrected reference vehicle fixed position data. The reference vehicle's fixed position data will be displayed in the form of a 3D diagram to provide a reference for the method of fixing the accident vehicle.

2. The intelligent fixation method for accident vehicles based on towing services according to claim 1, characterized in that, The method further includes the step of obtaining the real-time 3D architecture model, the step specifically including: Using a handheld visual sensor device, staff members continuously acquire and process image frames of the accident vehicle from all angles to obtain an image set of the accident vehicle and a corresponding visual odometry; wherein, the visual odometry is the cumulative pose estimation position and motion trajectory during the image frame acquisition process. Based on the visual odometer and the vehicle shape features in the accident vehicle image set, an initial three-dimensional structural model of the accident vehicle is constructed. The stereo matching cost aggregation calculation under the left and right visual differences is performed on the whole vehicle damage image set of the accident vehicle to obtain the fitting matching cost contour curve of the whole vehicle damage image set. Global loop detection is performed on the fitted matching cost contour curve under the relevant inter-frame pixel positioning changes, and a three-dimensional map of damage transformation points corresponding to the whole vehicle damage image set is constructed based on the constraint node parameters. Based on the coordinates of the same damage area in the vehicle damage image set, the three-dimensional map of the damage transformation point is fused with the initial three-dimensional structural model to obtain the real-time three-dimensional architecture model.

3. The intelligent fixation method for accident vehicles based on towing services according to claim 1, characterized in that, Before performing factor contribution correlation processing based on damage characteristics on the key damaged structural data in each damaged region, and evaluating and deriving the structural safety status data for each damaged region, the method further includes: For each factor in each damage region of the vehicle damage image, the marginal contribution weighted average is calculated under feature combination to obtain the marginal contribution weighted average. The SHAP interaction value is obtained by cross-correlation calculation of the weighted average marginal contribution of each damage type in the same damage region; and based on the SHAP interaction value between any two damage types, the damage types are averaged to obtain the SHAP value of each damage type in each damage region.

4. The intelligent fixation method for accident vehicles based on towing services according to claim 3, characterized in that, For each damaged region, the key damaged structural data are subjected to factor influence contribution correlation processing based on damage characteristics to evaluate and derive the structural safety status data for each damaged region, specifically including: A structural safety assessment and prediction model is established; wherein, the structural safety assessment and prediction model is used to perform comprehensive factor assessment processing on the key damaged structural data; Using the structural safety assessment and prediction model, the contribution of the model score prediction value to the total SAP value in each damaged area is calculated to obtain the percentage of accident damage factors for each damaged area; wherein, the percentage of accident damage factors is positively correlated with the total SAP value; The percentage of accident damage factors is converted into a numerical score to obtain structural safety data for each damaged area; wherein, the higher the score in the structural safety data, the more severe the damage.

5. The intelligent fixation method for accident vehicles based on towing services according to claim 1, characterized in that, After determining the reference vehicle fixed position data by performing local area correction processing based on the entire vehicle body area using the structural safety status data of each damaged area, the method further includes: Based on the reference vehicle fixed position data, the accident vehicle is loaded and secured to the trailer; wherein the trailer fixing structure includes at least: a binding fixing structure, auxiliary fixing buckles, and a vehicle limiter; The pressure sensor is placed at each fixed point in the reference vehicle fixed position data; The pressure sensor periodically monitors the pressure data of the trailer fixing structure, and the feedback data of the trailer fixing structure is obtained. The feedback data from the trailer fixing structure is tabulated and statistically processed to generate loading and driving safety data for monitoring the transportation of accident vehicles.

6. An intelligent vehicle securing device for accident vehicles based on towing services, characterized in that, The device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to execute a smart fixation method for accident vehicles based on a trailer service, as described in any one of claims 1-5.

Images