Head collision damage prediction method, device and storage medium for vehicle

Abstract

The application relates to the technical field of automobile pedestrian protection, in particular to a head collision injury prediction method and device of a vehicle and a storage medium, wherein the method comprises the following steps: obtaining original data of head collision and corresponding variable values, then encoding three-dimensional geometric information of the vehicle into a two-dimensional RGB image, generating a training set, a verification set and a test set according to the variable values and the two-dimensional RGB image, constructing a head collision injury prediction model by using the training set, the verification set and the test set, and outputting a head collision injury prediction result by using the head collision injury prediction model. Therefore, the problem that the prediction accuracy of the model is limited due to the fact that the feature parameters extracted by the related technology are limited and it is difficult to comprehensively describe the complex free surface and three-dimensional topological relationship of the cover plate is solved.

Description

Technical Field

[0001] This application relates to the field of vehicle pedestrian protection technology, and in particular to a method, device and storage medium for predicting head collision injuries of a vehicle. Background Technology

[0002] In recent years, data-driven methods have provided a new path for the rapid prediction of pedestrian protection performance. The core challenge lies in how to efficiently and comprehensively characterize vehicle structural information.

[0003] In related technologies, structured data is used for training. For example, feature data of the head shape itself and feature data of the vehicle body structure are extracted from pedestrian head collision simulation. The simulated HIC (Head Injury Criterion) values ​​calculated based on the feature data are saved together to establish a database. The database is used to train a preset neural network until the HIC prediction value output by the preset neural network meets the preset accuracy requirements to obtain a prediction model. The prediction model is then used for prediction and evaluation.

[0004] However, the extracted feature parameters (such as head mass, impact velocity, hair cover length, hair cover width, and thickness of inner and outer plates) in related technologies are limited, making it difficult to fully describe the complex free-form surface and three-dimensional topological relationship of the cover plate, which limits the prediction accuracy of the model and urgently needs to be solved. Summary of the Invention

[0005] This application provides a method, device, and storage medium for predicting head collision injuries in vehicles, in order to solve the problem that the limited feature parameters extracted by related technologies make it difficult to fully describe the complex freeform surface and three-dimensional topological relationship of the cover plate, thus limiting the prediction accuracy of the model.

[0006] The first aspect of this application provides a method for predicting head collision injuries in a vehicle, comprising the following steps: acquiring raw data of a head collision and corresponding variable values; encoding the three-dimensional geometric information of the vehicle into a two-dimensional RGB image to generate a training set, a validation set, and a test set based on the variable values ​​and the two-dimensional RGB image; constructing a head collision injury prediction model using the training set, the validation set, and the test set, and outputting a head collision injury prediction result using the head collision injury prediction model.

[0007] Based on the above technical means, this application embodiment encodes the three-dimensional geometric information of the vehicle into a two-dimensional RGB image, and generates training, validation, and test sets based on the original data, variable values, and two-dimensional RGB images. This facilitates the model's learning of geometric relationships, avoids information loss, reduces computational load, and shortens model training and inference time. Furthermore, since RGB images can be mapped back to three-dimensional surfaces, the model's output prediction results can be traced back to actual engineering parameters, enhancing engineering interpretability. In addition, by using the head collision injury prediction model to output head collision injury prediction results, end-to-end rapid prediction can be achieved, reducing the degree of human intervention and lowering the application threshold.

[0008] Optionally, in one embodiment of this application, encoding the three-dimensional geometric information of the vehicle into a two-dimensional RGB image includes: extracting the node coordinates of the inner panel of the engine hood and the spatial distance field from the inner panel of the engine hood to the target hard point based on the three-dimensional geometric information to generate a multi-dimensional vector point cloud; based on the multi-dimensional vector point cloud, independently normalizing the corresponding distance channel data to the target range to map them into R, G, and B channels respectively to generate a two-dimensional image; and cropping the target area image from the two-dimensional image with the collision point as the center and uniformly inputting the size to generate the two-dimensional RGB image.

[0009] Based on the above technical means, the embodiments of this application generate multi-dimensional vector point clouds by fusing coordinate and distance information, and combine RGB three-channel independent encoding to avoid losing three-dimensional depth and topological relationships, thereby achieving efficient and complete two-dimensionalization of three-dimensional geometric information. In addition, by truncating the target area and unifying the size, focusing on collision features, reducing redundant data, ensuring the uniformity of model input, and improving the efficiency and prediction accuracy of subsequent model training.

[0010] Optionally, in one embodiment of this application, obtaining the original data and corresponding variable values ​​of the head impact includes: performing a head impact simulation task based on a preset whole vehicle finite element model to obtain the acceleration time history curve of the vehicle, thereby generating the original data, and calculating the variable values.

[0011] Based on the above technical means, the embodiments of this application obtain the original data and variable values ​​through the simulation of the whole vehicle finite element model, without the need to carry out actual physical collision tests, which greatly reduces the test cost and shortens the data acquisition cycle. In addition, parameters such as impact speed and impact angle can be strictly controlled during the simulation process to ensure the consistency and repeatability of the data, avoiding errors caused by environmental and operational factors in actual tests. Furthermore, the simulation can accurately obtain the acceleration time history curve of the entire collision process, providing data assurance for the calculation of variable values.

[0012] Optionally, in one embodiment of this application, the step of constructing a head collision injury prediction model using the training set, the validation set, and the test set includes: removing the classification head and adding a custom regression head based on a pre-trained model network to determine the network model; and training the model using a preset objective loss function, an Adam optimizer, and a learning rate scheduler based on the network model, the training set, the validation set, and the test set until the model meets a preset evaluation metric, thereby establishing the head collision injury prediction model.

[0013] Based on the above technical means, the embodiments of this application utilize the transfer learning capability of pre-trained model networks to significantly reduce the difficulty of model training and improve feature extraction efficiency. By replacing it with a custom regression head, the model is made to better meet the needs of continuous prediction of head injury values. With the optimization and learning rate scheduler, the training process can be made more stable and converge faster, effectively avoiding overfitting and gradient oscillation. It can build a head collision injury prediction model with high prediction accuracy, strong generalization and good engineering applicability.

[0014] Optionally, in one embodiment of this application, the step of generating a training set, a validation set, and a test set based on the variable values ​​and the two-dimensional RGB image includes: removing outlier samples whose variable values ​​are higher than a preset threshold based on the variable values ​​and the two-dimensional RGB image to obtain cleaned data; and performing double stratified sampling based on the cleaned data according to vehicle type and damage score to determine the training set, the validation set, and the test set.

[0015] Based on the above technical means, the embodiments of this application effectively reduce the training disturbance caused by abnormal data by removing outlier samples, thereby improving the overall data quality. In addition, by adopting double stratified sampling, sample distribution bias is avoided, ensuring that each subset covers all working conditions, significantly improving the model's generalization ability and prediction stability, while making the subsequent model training process more convergent and the evaluation results more realistic and reliable.

[0016] A second aspect of this application provides a head collision injury prediction device for a vehicle, comprising: an acquisition module for acquiring raw data of a head collision and corresponding variable values; a generation module for encoding the three-dimensional geometric information of the vehicle into a two-dimensional RGB image, and generating a training set, a validation set, and a test set based on the variable values ​​and the two-dimensional RGB image; and a prediction module for constructing a head collision injury prediction model using the training set, the validation set, and the test set, and outputting a head collision injury prediction result using the head collision injury prediction model.

[0017] Based on the above technical means, this application embodiment encodes the three-dimensional geometric information of the vehicle into a two-dimensional RGB image, and generates training, validation, and test sets based on the original data, variable values, and two-dimensional RGB images. This facilitates the model's learning of geometric relationships, avoids information loss, reduces computational load, and shortens model training and inference time. Furthermore, since RGB images can be mapped back to three-dimensional surfaces, the model's output prediction results can be traced back to actual engineering parameters, enhancing engineering interpretability. In addition, by using the head collision injury prediction model to output head collision injury prediction results, end-to-end rapid prediction can be achieved, reducing the degree of human intervention and lowering the application threshold.

[0018] Optionally, in one embodiment of this application, the generation module includes: a first generation unit, configured to extract the node coordinates of the inner plate of the engine hood and the spatial distance field from the inner plate of the engine hood to the target hard point based on the three-dimensional geometric information, so as to generate a multi-dimensional vector point cloud; a second generation unit, configured to independently normalize the corresponding distance channel data to the target range based on the multi-dimensional vector point cloud, so as to map them to R, G, and B channels respectively, to generate a two-dimensional image; and a third generation unit, configured to extract the target area image from the two-dimensional image with the collision point as the center, and uniformly input the size, so as to generate the two-dimensional RGB image.

[0019] Based on the above technical means, the embodiments of this application generate multi-dimensional vector point clouds by fusing coordinate and distance information, and combine RGB three-channel independent encoding to avoid losing three-dimensional depth and topological relationships, thereby achieving efficient and complete two-dimensionalization of three-dimensional geometric information. In addition, by truncating the target area and unifying the size, focusing on collision features, reducing redundant data, ensuring the uniformity of model input, and improving the efficiency and prediction accuracy of subsequent model training.

[0020] A third aspect of this application provides a vehicle, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the vehicle head collision injury prediction method as described in the above embodiments.

[0021] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting head-on collision injuries in vehicles.

[0022] A fifth aspect of this application provides a computer program product, including a computer program that, when executed, implements the above-described method for predicting head-on collision injuries in vehicles.

[0023] This application's embodiments encode the vehicle's three-dimensional geometric information into a two-dimensional RGB image, and generate training, validation, and test sets based on the original data, variable values, and the two-dimensional RGB image. This facilitates the model's learning of geometric relationships, avoids information loss, reduces computational load, and shortens model training and inference time. Furthermore, since RGB images can be mapped back to three-dimensional surfaces, the model's output prediction results can be traced back to actual engineering parameters, enhancing engineering interpretability. Additionally, by utilizing the head collision injury prediction model to output head collision injury prediction results, end-to-end rapid prediction can be achieved, reducing the degree of human intervention and lowering the application threshold. Thus, it solves the problem that related technologies have limited feature parameters, making it difficult to fully describe the complex free-form surface and three-dimensional topological relationships of the cover plate, thereby limiting the model's prediction accuracy.

[0024] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0025] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for predicting head impact damage in a vehicle according to an embodiment of this application; Figure 2(a) is a schematic diagram of the collision point of a child's head in the construction of a finite element model of head collision according to an embodiment of this application; Figure 2(b) is a schematic diagram of building a collision finite element model according to an embodiment of the present application; Figure 3(a) is a schematic diagram of the inner plate structure in a multi-channel distance field RGB characterization according to an embodiment of this application; Figure 3(b) is a schematic diagram of the R channel sub-graph in a multi-channel distance field RGB representation provided according to an embodiment of this application; Figure 3(c) is a schematic diagram of the G channel sub-map in a multi-channel distance field RGB representation provided according to an embodiment of this application; Figure 3(d) is a schematic diagram of the B channel sub-map in a multi-channel distance field RGB representation provided according to an embodiment of this application; Figure 3(e) is a schematic diagram of RGB image synthesis in a multi-channel distance field RGB representation according to an embodiment of this application; Figure 4 This is a schematic diagram of a partial sample of a collision point provided according to an embodiment of this application; Figure 5 This is a schematic diagram of a three-dimensional structure reconstruction according to an embodiment of this application; Figure 6(a) is a histogram of label distribution before data cleaning according to an embodiment of this application; Figure 6(b) is a histogram of label distribution after data cleaning according to an embodiment of this application; Figure 7 This is a schematic diagram of a deep learning model architecture provided according to an embodiment of this application; Figure 8 This is a schematic diagram comparing the prediction performance of ResNet models with different depths according to an embodiment of this application; Figure 9 This is a schematic diagram comparing the impact of data cleaning on ResNet-18 prediction performance according to an embodiment of this application; Figure 10 A flowchart of a method for predicting head impact injuries in a vehicle according to an embodiment of this application; Figure 11 This is a block diagram of a vehicle head collision injury prediction device provided according to an embodiment of this application; Figure 12 This is a structural schematic diagram of a vehicle provided according to an embodiment of this application.

[0026] Figure label: 110 - Vehicle head collision injury prediction device; 100 - Acquisition module, 200 - Generation module, 300 - Prediction module; 1201 - Memory, 1202 - Processor, 1203 - Communication interface. Detailed Implementation

[0027] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0028] Since collisions between pedestrian heads and engine hood panels are one of the main scenarios leading to fatal injuries, it is crucial to quickly and accurately assess the protective performance of the hood structure for pedestrian heads during the forward development of vehicles.

[0029] While existing finite element simulation methods offer high accuracy, they are computationally expensive and time-consuming, making them unsuitable for rapid iterative design requirements. In recent years, data-driven methods have provided a new path for rapid prediction of pedestrian protection performance, with the core challenge being how to efficiently and comprehensively characterize vehicle structural information.

[0030] Table 1 is a review of recent damage prediction research methods and performance. The relevant structural characterization methods are mainly divided into three categories: image data, structured data, and voxel data.

[0031] (1) Image data is predicted by extracting a small number of two-dimensional cross-sectional images of the collision point and using a convolutional neural network. Although it has advantages in processing efficiency, its two-dimensional projection process loses spatial depth information and cannot reflect the three-dimensional topological relationship, making it difficult to further improve the prediction accuracy.

[0032] (2) Structured data can extract limited design features (such as plate thickness, material, etc.), but it is difficult to fully describe the complex free-form surface and three-dimensional topological relationship of the cover plate.

[0033] (3) Voxel data discretizes the vehicle structure in three dimensions and combines it with an autoencoder for feature learning. However, voxel representation can only mark the existence of the structure in binary form and cannot accurately express the key spatial distance information, which is the core physical factor affecting collision energy absorption. Improving the accuracy of distance representation requires a significant increase in grid density, resulting in high data dimensionality, large computational load, and limited generalization ability.

[0034] Table 1

[0035] Based on the above explanation, it is understandable that related structural representation methods share the same pain point: they cannot accurately represent the complete information of three-dimensional geometry while controlling data complexity to adapt to efficient deep learning models.

[0036] To address the aforementioned problems, this application provides a method, apparatus, and storage medium for predicting head-on collision injuries in vehicles. The following description, with reference to the accompanying drawings, details the method, apparatus, and storage medium for predicting head-on collision injuries in vehicles according to embodiments of this application.

[0037] Figure 1 This is a flowchart of a method for predicting head impact damage in a vehicle according to an embodiment of this application.

[0038] like Figure 1 As shown, the method for predicting head impact injuries in this vehicle includes the following steps: In step S101, the raw data of head collision and the corresponding variable values ​​are obtained.

[0039] Specifically, raw data refers to the collision dynamics response data that is directly output without processing after conducting head impact simulation through a preset whole vehicle finite element model. Specifically, it is the acceleration time history curve that changes with time during the pedestrian head collision, used to record the intensity, duration and variation of the collision impact.

[0040] The corresponding variable value refers to the quantitative index of head injury calculated based on the original acceleration time history curve. Its value directly reflects the degree of head injury caused by the collision. It serves as the label value required for training the deep learning model, and is used to supervise the model learning and output prediction results.

[0041] In some cases, embodiments of this application can simulate a real collision scenario between a pedestrian's head and the engine hood of a vehicle based on a preset finite element model of the whole vehicle, perform a head impact simulation task, output the acceleration time history curve of the head shape during the collision through simulation software (such as LS-DYNA), use this curve as the raw data of the head collision, and then calculate the corresponding variable value based on the acceleration time history curve and the head injury index calculation formula.

[0042] In step S102, the three-dimensional geometric information of the vehicle is encoded into a two-dimensional RGB image to generate a training set, a validation set, and a test set based on the variable values ​​and the two-dimensional RGB image.

[0043] Specifically, three-dimensional geometric information refers to the morphological data of the vehicle's engine hood and surrounding related structures in three-dimensional space, which may include, but is not limited to, the curved shapes of the inner and outer panels of the engine hood.

[0044] Two-dimensional RGB images refer to standardized images obtained by encoding and converting the three-dimensional geometric information of a vehicle. They are generated by mapping the distance fields from the inner panel of the cover to different target hard points to the red, green, and blue color channels, and then normalizing and cropping the regions. These images are used to compactly carry the three-dimensional spatial structural features in the form of two-dimensional images.

[0045] In some cases, embodiments of this application can extract the node coordinates of the inner panel of the engine hood based on the vehicle's three-dimensional geometric information, and calculate the spatial distance field from the node to the target hard point (inner surface of the outer panel, outer surface of the engine compartment, horizontal ground), thereby fusing the coordinate and distance information to form a multi-dimensional vector point cloud. Then, the multiple distance channel data in the multi-dimensional vector point cloud are normalized to map to the target range of the adapted image display to generate an original two-dimensional image. Finally, the target area image is cropped with the collision point as the center, the input size is unified, and a two-dimensional RGB image is generated.

[0046] Furthermore, in this embodiment, based on two-dimensional RGB images and corresponding variable values, outlier samples with variable values ​​higher than a preset threshold are first removed to obtain cleaned data. Then, based on the cleaned data, stratified sampling is performed using vehicle type and damage score as dual criteria to ensure that samples of different vehicle types and different damage levels are evenly distributed in each data subset, so as to construct training set, validation set and test set.

[0047] In step S103, a head collision injury prediction model is constructed using the training set, validation set, and test set, and the head collision injury prediction result is output using the head collision injury prediction model.

[0048] Specifically, the head collision injury prediction model refers to a regression model formed by custom modification based on a pre-trained deep learning network. It learns the mapping relationship between the three-dimensional structural features represented by two-dimensional RGB images in the training set and the corresponding injury variable values. After training is completed when the preset evaluation indicators are met, it can directly use two-dimensional RGB images as input to quickly infer and output the numerical value of head collision injury.

[0049] Head collision injury prediction results refer to the continuous quantitative values ​​output by the trained prediction model. They are usually head injury index prediction values, which are used to directly characterize the severity of head injury when a pedestrian collides with the engine hood of a vehicle.

[0050] In some cases, embodiments of this application use a pre-trained model network as a basis, remove the original network head used for classification tasks, replace it with a custom regression head adapted for continuous numerical prediction, construct a network structure suitable for head injury index regression, and then use a preset target loss function, Adam optimizer and learning rate scheduler to iteratively optimize network parameters on the training set, adjust the model state through the validation set and test the generalization ability through the test set until the model meets the preset evaluation index, and finally establish a head collision injury prediction model.

[0051] Furthermore, in this embodiment of the application, the two-dimensional RGB image corresponding to the engine hood area of ​​the vehicle to be predicted is input into the established head collision injury prediction model. The model extracts features and performs inference calculations on the input image through the mapping relationship between the three-dimensional geometric features it has learned and the head collision injury, and directly outputs the corresponding continuous numerical result, which is the predicted value characterizing the degree of pedestrian head collision injury.

[0052] Optionally, in one embodiment of this application, obtaining the original data and corresponding variable values ​​of head impact includes: performing a head impact simulation task based on a preset whole vehicle finite element model to obtain the vehicle's acceleration time history curve to generate the original data and calculate the variable values.

[0053] Specifically, the pre-set whole vehicle finite element model refers to a three-dimensional digital simulation model of the whole vehicle, including the body structure and engine hood, which is established in advance and verified by experiments. The model restores the geometry, material properties, connection relationships and boundary conditions of each part of the vehicle body through mesh discretization. It can realistically reflect the mechanical response characteristics of the vehicle under collision impact, and provide a standardized and reusable numerical calculation carrier for subsequent pedestrian head impact simulation.

[0054] Head impact simulation refers to a virtual simulation task that, based on pedestrian protection regulations or evaluation procedures (such as C-NCAP), applies a standard head impact to a pre-set whole vehicle finite element model, sets working condition parameters such as impact speed, impact angle, and impact position, simulates the entire process of a pedestrian's head colliding with the engine hood, and obtains dynamic response data such as head acceleration, deformation, and contact force, in order to acquire raw data and calculate variable values.

[0055] In actual implementation, this application embodiment uses a preset whole vehicle finite element model as a basis to simulate a real collision scenario between a pedestrian's head and the vehicle's engine hood, and performs a head impact simulation task. The acceleration time history curve of the head shape during the collision is output by simulation software (such as LS-DYNA). This curve is used as the raw data of the head collision. Based on the acceleration time history curve, the corresponding variable values ​​are calculated using the head injury index calculation formula.

[0056] For example, the embodiments of this application are based on the whole vehicle finite element models of 11 mass-produced models. First, the material properties of the inner and outer panels of the engine hood are unified, and unnecessary structures such as reinforcing ribs are removed to eliminate the interference of irrelevant variables on the simulation results.

[0057] As shown in Figure 2(a), in this embodiment, test points were arranged in the potential collision area of ​​the cover plate, and a child head impact simulation was performed using LS-DYNA software. The simulation parameters strictly followed the C-NCAP pedestrian protection test standards: impact speed 40 km / h, impact angle 50°, and head mass 3.5 kg. As shown in Figure 2(b), in this embodiment, to simplify the calculation, the portion of the vehicle body in front of the A-pillar was cut off, and a 6-DOF constraint was applied at the cut-off point to ensure consistent simulation boundary conditions.

[0058] The embodiments of this application obtain the head shape acceleration time history curve through simulation, and calculate HIC based on formula (1). 15 A total of 1254 original samples were obtained.

[0059] HIC 15 The expression for calculating the value can be, but is not limited to, the following: (1) in, For HIC 15 value, The beginning of the impact, At the end of the impact, For impact acceleration. In the embodiments of this application, HIC 15 The calculation time window is less than or equal to 15ms, the acceleration integration interval is accurate to 0.1ms, and the simulation model has been experimentally verified to ensure data reliability.

[0060] This application embodiment obtains original data and variable values ​​through whole vehicle finite element model simulation, eliminating the need for actual physical collision tests, significantly reducing test costs and shortening the data acquisition cycle. In addition, parameters such as impact speed and impact angle can be strictly controlled during the simulation process to ensure data consistency and repeatability, avoiding errors caused by environmental and operational factors in actual tests. Furthermore, the simulation can accurately obtain the acceleration time history curve of the entire collision process, providing data assurance for the calculation of variable values.

[0061] Optionally, in one embodiment of this application, the three-dimensional geometric information of the vehicle is encoded into a two-dimensional RGB image, including: extracting the node coordinates of the inner panel of the engine hood and the spatial distance field from the inner panel of the engine hood to the target hard point based on the three-dimensional geometric information to generate a multi-dimensional vector point cloud; based on the multi-dimensional vector point cloud, independently normalizing the corresponding distance channel data to the target range to map them to R, G, and B channels respectively to generate an original two-dimensional image; and cropping the target area image from the original two-dimensional image with the collision point as the center and uniformly inputting the size to generate a two-dimensional RGB image.

[0062] Specifically, the target hard point refers to the reference structural location that plays a key role in constraining the deformation and energy absorption of the engine hood in a pedestrian head collision scenario. It includes the inner surface of the outer engine hood panel, the outer surface of the engine compartment, and the horizontal ground, and is used to quantify the spatial distance relationship between the inner panel of the hood and the surrounding rigid structures.

[0063] Multidimensional vector point cloud refers to a set of points composed of the shell unit nodes of the inner plate of the engine cover. Each node not only contains three-dimensional spatial coordinate information, but also includes the shortest distance data to multiple target hard points, forming a high-dimensional feature vector, which is used to fully characterize the three-dimensional geometric shape and spatial position relationship of the cover.

[0064] Distance channel data refers to multiple sets of distance data separated from the multidimensional vector point cloud, each corresponding to a different target hard point. Each set of data independently represents the spatial distance distribution from the inner plate of the cover to a certain target hard point, and is used as an independent feature channel for subsequent RGB three-channel mapping.

[0065] The target range refers to the numerical range defined after normalizing the distance channel data, usually 0-255, which is consistent with the grayscale value range of image pixels. This allows distance data of different magnitudes to be uniformly converted into image-recognizable values, making it easier to generate standard RGB images.

[0066] The original two-dimensional image refers to a two-dimensional image obtained by normalizing multiple sets of distance channel data, independently mapping them into R, G, and B single-channel images, and then stacking the channels. It fully carries three-dimensional geometric and spatial distance information and has not yet undergone region cropping and size normalization.

[0067] The target region image refers to the local feature region extracted from the original two-dimensional image with the projection of the collision point on the two-dimensional plane as the center. It corresponds to the main impact range of the collision in physical space. After being standardized in size, it is used as the standard input of the deep learning model to focus on key structural features and reduce data redundancy.

[0068] In actual implementation, the embodiments of this application can extract the node coordinates of the inner panel of the engine hood based on the vehicle's three-dimensional geometric information, and calculate the spatial distance field from the node to the target hard point (inner surface of the outer panel, outer surface of the engine compartment, and horizontal ground). The coordinate and distance information are then fused to form a multi-dimensional vector point cloud. The multiple distance channel data in the multi-dimensional vector point cloud are then normalized to map to the target range of the adapted image display to generate the original two-dimensional image. The target area image is then cropped with the collision point projection as the center, and the input size is unified to generate a two-dimensional RGB image.

[0069] For example, as shown in Figure 3, this embodiment of the application uses preprocessing software (such as HyperMesh, Primer, ANSA, etc.) to extract the three-dimensional coordinates (x, y, z) of all shell unit nodes of the inner panel of the engine hood, and calculates the shortest distance from each node to three target hard points: the shortest distance do to the inner surface of the outer panel, the shortest distance dc to the outer surface of the engine compartment, and the shortest distance dg to the horizontal ground. Each node generates a six-dimensional vector (x, y, z, do, dc, dg) to form an enhanced point cloud. The point cloud density depends on the mesh size, and tens of thousands to hundreds of thousands of nodes are generated for each vehicle model, completely covering the collision-related structure of the hood.

[0070] As shown in Figure 3(a), firstly, in this embodiment of the application, the (x, y) coordinates of the point cloud are projected onto a two-dimensional plane to generate a black and white scatter plot as the base map, while preserving the spatial position correspondence of the nodes.

[0071] Then, in this embodiment, the distance data for the three channels do, dc, and dg are extracted respectively, and each channel is independently normalized to a range of 0-255. The expression for the normalization process can be, but is not limited to, the following: Normalized value = (original distance - minimum distance) / (maximum distance - minimum distance) × 255. (2) As shown in Figures 3(b)-(d), after normalization, the embodiments of this application map do to the red channel R, dc to the green channel G, and dg to the blue channel B, generating a single-channel sub-image.

[0072] As shown in Figure 3(e), in this embodiment, the R, G, and B channels are stacked to synthesize a complete RGB image. The image size is consistent with the point cloud distribution area, ensuring that each pixel corresponds to the actual spatial position, thus achieving two-dimensional lossless encoding of three-dimensional geometry and distance information.

[0073] like Figure 4 As shown, in this embodiment, a 300 mm × 300 mm physical region is cropped centered on the collision point projection (xp, yp), corresponding to an image size of 300 × 300 pixels. This is used to focus on the main deformation area of ​​the collision, unify the model input size, and facilitate subsequent deep learning model processing. Figure 5 As shown, this characterization method is reversible, capable of mapping back from an RGB image to a 3D surface, supporting engineering backtracking and design verification.

[0074] This application embodiment generates a multi-dimensional vector point cloud by fusing coordinate and distance information, and combines independent encoding of RGB three channels to avoid losing three-dimensional depth and topological relationships, thereby achieving efficient and complete two-dimensionalization of three-dimensional geometric information. In addition, by truncating the target region and unifying its size, it focuses on collision features, reduces redundant data, ensures the uniformity of model input, and improves the efficiency of subsequent model training and prediction accuracy.

[0075] Optionally, in one embodiment of this application, generating a training set, a validation set, and a test set based on variable values ​​and a two-dimensional RGB image includes: removing outlier samples with variable values ​​higher than a preset threshold based on variable values ​​and a two-dimensional RGB image to obtain cleaned data; and performing double stratified sampling based on the cleaned data according to vehicle type and damage score to determine the training set, validation set, and test set.

[0076] Specifically, the preset threshold refers to a critical value pre-set based on the statistical law of sample distribution. It is calculated using the interquartile range (IQR) method and is used to determine whether the samples in the original data are abnormal. Data exceeding this value will be judged as abnormal data and removed to ensure the rationality of subsequent training data.

[0077] Outliers are a few samples in the original data whose variable values ​​deviate significantly from the overall data distribution and are much higher than the normal range. These samples are often generated by simulation anomalies or extreme working conditions, which can interfere with the model training direction and reduce prediction accuracy. They are invalid or distorted data that need to be removed.

[0078] Cleaned data refers to the effective dataset obtained after the original data has undergone outlier removal. It removes the interference of outlier samples, and the data distribution is more concentrated and the features are more representative. It is used to provide a high-quality, stable and reliable data foundation for subsequent stratified sampling and model training.

[0079] In actual implementation, this application embodiment is based on variable values ​​and two-dimensional RGB images. First, outlier samples with variable values ​​higher than a preset threshold are removed to obtain cleaned data. Then, based on the cleaned data, stratified sampling is performed using vehicle type and damage score as dual criteria to ensure that samples of different vehicle types and different damage levels are evenly distributed in each data subset, so as to construct training set, validation set and test set.

[0080] For example, embodiments of this application employ the IQR method to clean HIC. 15 Outliers in the values: Calculate the first quartile Q1 = 631.30 and the third quartile Q3 = 1152.08, obtaining the interquartile range (IQR) = Q3 - Q1 = 520.78. Set the outlier coefficient k = 2.5, calculate the outlier threshold as Q3 + k × IQR = 2454.04, and remove outliers from the HIC range. 15 Ten outlier samples with values ​​greater than 2454.04 were identified, and 1244 samples were retained after cleaning. Table 2 shows the HIC values ​​before and after outlier cleaning. 15 Comparison of key statistics: Figure 6(a) is the histogram of label distribution before data cleaning, and Figure 6(b) is the histogram of label distribution after data cleaning. As can be seen from Table 2 and Figure 6, the data distribution is more concentrated after cleaning, the standard deviation is reduced by 6.66%, and the skewness is improved from 1.52 to 1.28, effectively improving the data quality.

[0081] Table 2

[0082] Table 3 shows the C-NCAP head injury color scores. In this embodiment, the vehicle model and C-NCAP head injury color scores are used as dual stratification criteria, and the dataset is divided into three sets in a 7:2:1 ratio: 871 samples for training, 249 samples for validation, and 124 samples for testing. This stratified sampling strategy ensures that each subset covers all operating conditions, avoids data distribution bias, and improves the model's generalization ability.

[0083] Table 3

[0084] The embodiments of this application effectively reduce training disturbances caused by abnormal data by removing outlier samples, thereby improving the overall data quality. In addition, by adopting double stratified sampling, sample distribution bias is avoided, ensuring that each subset covers all working conditions, significantly improving the model's generalization ability and prediction stability, while making the subsequent model training process more convergent and the evaluation results more realistic and reliable.

[0085] Optionally, in one embodiment of this application, a head collision injury prediction model is constructed using a training set, a validation set, and a test set, including: removing a classification head and adding a custom regression head based on a pre-trained model network to determine the network model; and training the model using a preset objective loss function, an Adam optimizer, and a learning rate scheduler based on the network model, the training set, the validation set, and the test set until the model meets preset evaluation metrics, thereby establishing a head collision injury prediction model.

[0086] Specifically, the classification head refers to the terminal network structure in a pre-trained model network used for classification tasks. It is usually composed of fully connected layers and activation functions, and its function is to map features to discrete class probability outputs. Since this application is for continuous numerical prediction rather than classification, it needs to be removed.

[0087] A custom regression head refers to the network terminal structure designed for continuous prediction of head injury indicators. It specifically includes an adaptive average pooling layer and a fully connected layer, which are used to map the extracted image features into a single continuous injury prediction value.

[0088] Preset evaluation metrics refer to standards used to quantitatively evaluate the prediction accuracy of a model. For example, if a relevant metric (such as the coefficient of determination R²) reaches a preset threshold during model training, the model is considered to have converged, and the model satisfies the preset evaluation metrics.

[0089] In actual implementation, this application embodiment is based on a pre-trained model network. The original network head used for classification tasks is removed and replaced with a custom regression head adapted for continuous numerical prediction. A network structure suitable for head injury index regression is constructed. Then, the network parameters are iteratively optimized on the training set using a preset target loss function, Adam optimizer and learning rate scheduler. The model state is adjusted through the validation set and the generalization ability is tested through the test set until the model meets the preset evaluation index. Finally, a head collision injury prediction model is established.

[0090] For example, such as Figure 7 As shown, this embodiment uses a pre-trained ResNet network as a base, removes the original network's classification head, and adds a custom regression head. The regression head includes an adaptive average pooling layer and three fully connected layers (the middle layer contains ReLU activation and Dropout, with a dropout rate p=0.3), and outputs HIC. 15 Scalar. By comparing ResNet-18, ResNet-34, and ResNet-50, this embodiment prefers ResNet-18 (11.2 M parameters), which achieves the best balance between accuracy and computational efficiency.

[0091] This application uses the MSE (Mean Squared Error) loss function, the Adam optimizer (initial learning rate 0.001, weight decay 0.0001), and the ReduceLROnPlateau learning rate scheduler (the learning rate is halved if the loss does not decrease after 10 training epochs). It trains for 1000 epochs with a batch size of 32, and performs data augmentation (random flipping and 5° rotation) to improve model robustness. The MSE loss function can be, but is not limited to, the following: (3) in, For batch size, For the first HIC of a sample 15 The true value For the first HIC of a sample 15 Predicted value.

[0092] This application employs four metrics—the coefficient of determination (R²), RMSE (Root Mean Squared Error), MAE (Mean Absolute Error), and MSE—to comprehensively evaluate model performance. The expression for R² can be, but is not limited to, as follows: (4) in, For R², HIC for all samples 15 average value.

[0093] RMSE expressions can be, but are not limited to, the following: (5) in, It is RMSE.

[0094] MAE expressions can be, but are not limited to: (6) in, For MAE.

[0095] Table 4 shows the experimental environment. The embodiments of this application were executed on an Intel Xeon processor, Ubuntu 22.04.5 LTS system, and PPU-ZW810E GPU (96GB of video memory), based on Python 3.10.13 and PyTorch 2.4.0 framework, to ensure that those skilled in the art can reproduce them independently.

[0096] Table 4

[0097] This application embodiment leverages the transfer learning capabilities of pre-trained model networks to significantly reduce model training difficulty and improve feature extraction efficiency. By replacing the model with a custom regression head, the model better meets the requirements for continuous prediction of head injury values. Combined with an optimizer and learning rate scheduler, the training process becomes more stable and converges faster, effectively avoiding overfitting and gradient oscillations. This enables the construction of a head collision injury prediction model with high prediction accuracy, strong generalization, and good engineering applicability.

[0098] It should be noted that in the model architecture comparison experiments, the embodiments of this application maintained consistent hyperparameters on the IQR-cleaned dataset, training ResNet-18, ResNet-34, and ResNet-50 models respectively. Table 5 shows the performance comparison of ResNet models of different depths on the test set. As can be seen from the table, ResNet-18 achieved the best performance, with R²=0.9897, RMSE=44.3683, and MAE=28.9223 on the test set. Additionally, as... Figure 8 As shown, the predicted points are most concentrated along the y=x reference line, proving that the lightweight model has adequately captured the features. Based on the above analysis, it can be seen that deeper networks do not improve accuracy but instead increase computational cost, verifying the rationale for the optimal choice of ResNet-18.

[0099] Table 5

[0100] In the data cleaning effect verification experiment, this application embodiment uses the ResNet-18 model, trained on the original data (1254 samples) and the cleaned data (1244 samples) respectively. Table 6 shows the performance comparison of ResNet-18 on the datasets before and after abnormal cleaning. As can be seen from the table, the model performance is comprehensively improved after cleaning, R 2 As the RMSE increased from 0.9874 to 0.9897, it decreased from 49.1370 to 44.3683. Figure 9 As shown, the model prediction error range is narrower after cleaning. Based on the above analysis, it can be concluded that IQR cleaning significantly improves model stability and accuracy.

[0101] Table 6

[0102] like Figure 10 As shown below, the method for predicting head collision damage of a vehicle proposed in this application is illustrated with a specific embodiment.

[0103] S1001, obtain the raw data and corresponding variable values ​​of head collision.

[0104] In this embodiment, a head impact simulation task is performed based on a preset finite element model of a vehicle to simulate a real collision scenario between a pedestrian's head and the vehicle's engine hood. The acceleration time-history curve of the head shape during the collision is output using LS-DYNA software. This curve is used as the raw data for the head impact. Then, based on the acceleration time-history curve, the corresponding variable value (i.e., HIC) is calculated using the head injury index calculation formula. 15 value).

[0105] S1002 encodes the vehicle's three-dimensional geometric information into a two-dimensional RGB image.

[0106] In this embodiment, based on the vehicle's three-dimensional geometric information, the node coordinates of the inner panel of the engine hood are extracted, and the spatial distance field from the node to the target hard point (inner surface of the outer panel, outer surface of the engine compartment, and horizontal ground) is calculated. The coordinate and distance information are then fused to form a multi-dimensional vector point cloud (x, y, z, do, dc, dg). The multiple distance channel data do, dc, and dg in the multi-dimensional vector point cloud are then normalized to map to the target range of 0-255 for the adapted image display, thereby generating an original two-dimensional image. Then, with the collision point projection (xp, yp) as the center, a target area image of size 300 mm × 300 mm is cropped, and the input size is unified to generate a two-dimensional RGB image of size 300 × 300 × 3.

[0107] S1003 generates training, validation, and test sets based on variable values ​​and two-dimensional RGB images.

[0108] In this embodiment, based on variable values ​​and two-dimensional RGB images, the interquartile range (IQR) method is used to remove outlier samples (i.e., HIC samples) whose variable values ​​are higher than a preset threshold. 15 Ten outlier samples with values ​​greater than 2454.04 were identified. After cleaning, the data was then stratified based on the cleaned data (1244 samples were retained after cleaning) and the vehicle model and damage score were used as dual criteria. The dataset was divided into three sets in a 7:2:1 ratio: 871 samples for training, 249 samples for validation, and 124 samples for testing.

[0109] S1004 uses training, validation and test sets to build a head collision injury prediction model.

[0110] In this embodiment, a pre-trained ResNet network is used as the basis. The classification head of the original network is removed and a custom regression head is added to construct a network structure suitable for head injury index regression. Then, the MSE loss function, Adam optimizer and ReduceLROnPlateau learning rate scheduler are used to iteratively optimize the network parameters on the training set. The model state is adjusted through the validation set and the generalization ability is tested through the test set until the model meets the preset evaluation index (including four types of indexes: R², RMSE, MAE and MSE), thereby establishing a head collision injury prediction model.

[0111] S1005, uses the head collision injury prediction model to output the head collision injury prediction results.

[0112] In this embodiment, the two-dimensional RGB image corresponding to the engine hood area of ​​the vehicle to be predicted is input into the established head collision injury prediction model. The model extracts features and performs inference calculations on the input image through the mapping relationship between the three-dimensional geometric features it has learned and the head collision injury, and directly outputs the corresponding continuous numerical result, which is the predicted value characterizing the degree of pedestrian head collision injury.

[0113] The vehicle head-on collision injury prediction method proposed in this application encodes the vehicle's three-dimensional geometric information into a two-dimensional RGB image. Training, validation, and test sets are generated based on the original data, variable values, and the two-dimensional RGB image. This facilitates the model's learning of geometric relationships, avoids information loss, reduces computational load, and shortens model training and inference time. Furthermore, since the RGB image can be mapped back to a three-dimensional surface, the model's prediction results can be traced back to actual engineering parameters, enhancing engineering interpretability. Additionally, by utilizing the head-on collision injury prediction model to output head-on collision injury prediction results, end-to-end rapid prediction can be achieved, reducing the degree of human intervention and lowering the application threshold. Thus, this solves the problem that related technologies have limited feature parameters, making it difficult to fully describe the complex free-form surface and three-dimensional topological relationships of the cover plate, thereby limiting the model's prediction accuracy.

[0114] Next, referring to the accompanying drawings, a vehicle head collision injury prediction device according to an embodiment of this application is described.

[0115] Figure 11 This is a block diagram of a vehicle head collision injury prediction device provided according to an embodiment of this application.

[0116] like Figure 11 As shown, the head collision injury prediction device 110 for the vehicle includes: an acquisition module 100, a generation module 200, and a prediction module 300.

[0117] The acquisition module 100 is used to acquire the raw data of head collision and the corresponding variable values.

[0118] The generation module 200 is used to encode the three-dimensional geometric information of the vehicle into a two-dimensional RGB image to generate a training set, a validation set, and a test set based on the variable values ​​and the two-dimensional RGB image.

[0119] The prediction module 300 is used to construct a head collision injury prediction model using the training set, validation set, and test set, and to output the head collision injury prediction result using the head collision injury prediction model.

[0120] Optionally, in one embodiment of this application, the acquisition module 100 includes: an acquisition unit, used to perform a head impact simulation task based on a preset whole vehicle finite element model to obtain the vehicle's acceleration time history curve, generate raw data, and calculate variable values.

[0121] Optionally, in one embodiment of this application, the generation module 200 includes: a first generation unit, a second generation unit, and a third generation unit.

[0122] The first generation unit is used to extract the node coordinates of the inner plate of the engine cover and the spatial distance field from the inner plate of the engine cover to the target hard point based on three-dimensional geometric information, so as to generate a multi-dimensional vector point cloud.

[0123] The second generation unit is used to independently normalize the corresponding distance channel data to the target range based on the multidimensional vector point cloud, so as to map them into R, G, and B channels respectively, and generate a two-dimensional image.

[0124] The third generation unit is used to extract the target region image from the two-dimensional image with the collision point as the center, and uniformly input the size to generate a two-dimensional RGB image.

[0125] Optionally, in one embodiment of this application, the generation module 200 includes a cleaning unit and a sampling unit.

[0126] The cleaning unit is used to remove outlier samples whose variable values ​​are higher than a preset threshold based on the variable values ​​and the two-dimensional RGB image, so as to obtain cleaned data.

[0127] The sampling unit is used to perform stratified sampling based on the cleaned data, according to vehicle type and damage score, to determine the training set, validation set, and test set.

[0128] Optionally, in one embodiment of this application, the prediction module 300 includes: a determination unit and an establishment unit.

[0129] The determination unit is used to remove the classification head and add a custom regression head based on the pre-trained model network to determine the network model.

[0130] A unit is established to train a head collision injury prediction model based on a network model, training set, validation set, and test set, using a preset objective loss function, Adam optimizer, and learning rate scheduler until the model meets preset evaluation metrics.

[0131] It should be noted that the foregoing explanation of the vehicle head collision injury prediction method embodiment also applies to the vehicle head collision injury prediction device of this embodiment, and will not be repeated here.

[0132] The vehicle head collision injury prediction device proposed in this application encodes the vehicle's three-dimensional geometric information into a two-dimensional RGB image, and generates training, validation, and test sets based on the original data, variable values, and the two-dimensional RGB image. This facilitates the model's learning of geometric relationships, avoids information loss, reduces computational load, and shortens model training and inference time. Furthermore, since the RGB image can be mapped back to a three-dimensional surface, the model's prediction results can be traced back to actual engineering parameters, enhancing engineering interpretability. In addition, by utilizing the head collision injury prediction model to output head collision injury prediction results, end-to-end rapid prediction can be achieved, reducing the degree of human intervention and lowering the application threshold. Thus, it solves the problem that the limited feature parameters extracted by related technologies make it difficult to fully describe the complex free-form surface and three-dimensional topological relationships of the cover plate, thus limiting the model's prediction accuracy.

[0133] Figure 12 This is a schematic diagram of the structure of a vehicle according to an embodiment of this application. The vehicle may include: The memory 1201, the processor 1202, and the computer program stored on the memory 1201 and executable on the processor 1202.

[0134] When the processor 1202 executes the program, it implements the vehicle head collision injury prediction method provided in the above embodiments.

[0135] Furthermore, the vehicle also includes: Communication interface 1203 is used for communication between memory 1201 and processor 1202.

[0136] The memory 1201 is used to store computer programs that can run on the processor 1202.

[0137] The memory 1201 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage.

[0138] If the memory 1201, processor 1202, and communication interface 1203 are implemented independently, then the communication interface 1203, memory 1201, and processor 1202 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 12 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0139] Optionally, in a specific implementation, if the memory 1201, processor 1202, and communication interface 1203 are integrated on a single chip, then the memory 1201, processor 1202, and communication interface 1203 can communicate with each other through an internal interface.

[0140] The processor 1202 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0141] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting head collision injuries in vehicles.

[0142] This application also provides a computer program product, including a computer program that, when executed, implements the above-described method for predicting head collision injuries in vehicles.

[0143] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0144] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0145] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0146] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0147] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0148] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0149] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0150] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for predicting head impact injuries in vehicles, characterized in that, Includes the following steps: Obtain the raw data and corresponding variable values ​​for head collision; The vehicle's three-dimensional geometric information is encoded into a two-dimensional RGB image to generate a training set, a validation set, and a test set based on the variable values ​​and the two-dimensional RGB image. A head collision injury prediction model is constructed using the training set, the validation set, and the test set, and the head collision injury prediction result is output using the head collision injury prediction model.

2. The method according to claim 1, characterized in that, Encoding the vehicle's three-dimensional geometric information into a two-dimensional RGB image includes: Based on the three-dimensional geometric information, the node coordinates of the inner plate of the engine cover and the spatial distance field from the inner plate of the engine cover to the target hard point are extracted to generate a multi-dimensional vector point cloud. Based on the multidimensional vector point cloud, the corresponding distance channel data is independently normalized to the target range and mapped to R, G, and B channels respectively to generate the original two-dimensional image. Centered on the collision point, the target region image is extracted from the original two-dimensional image, and the size is uniformly input to generate the two-dimensional RGB image.

3. The method according to claim 1, characterized in that, The process of obtaining the raw data and corresponding variable values ​​for head collision includes: Based on a preset whole vehicle finite element model, a head impact simulation task is performed to obtain the vehicle's acceleration time history curve, thereby generating the raw data and calculating the variable values.

4. The method according to claim 1, characterized in that, The step of constructing a head impact injury prediction model using the training set, the validation set, and the test set includes: Based on the pre-trained model network, the classification head is removed and a custom regression head is added to determine the network model; Based on the network model, the training set, the validation set, and the test set, the model is trained using a preset objective loss function, the Adam optimizer, and a learning rate scheduler until it meets the preset evaluation criteria, thus establishing the head collision injury prediction model.

5. The method according to claim 1, characterized in that, The step of generating training, validation, and test sets based on the variable values ​​and the two-dimensional RGB image includes: Based on the variable values ​​and the two-dimensional RGB image, outlier samples with variable values ​​higher than a preset threshold are removed to obtain cleaned data. Based on the cleaned data, stratified sampling is performed according to vehicle type and damage score to determine the training set, the validation set, and the test set.

6. A device for predicting head impact injuries in vehicles, characterized in that, include: The acquisition module is used to acquire the raw data and corresponding variable values ​​of head collision. The generation module is used to encode the three-dimensional geometric information of the vehicle into a two-dimensional RGB image, so as to generate a training set, a validation set and a test set based on the variable values ​​and the two-dimensional RGB image; The prediction module is used to construct a head collision injury prediction model using the training set, the validation set, and the test set, and to output a head collision injury prediction result using the head collision injury prediction model.

7. The apparatus according to claim 6, characterized in that, The generation module includes: The first generation unit is used to extract the node coordinates of the inner plate of the engine cover and the spatial distance field from the inner plate of the engine cover to the target hard point based on the three-dimensional geometric information, so as to generate a multi-dimensional vector point cloud. The second generation unit is used to independently normalize the corresponding distance channel data to the target range based on the multidimensional vector point cloud, so as to map them into R, G, and B channels respectively, and generate a two-dimensional image. The third generation unit is used to extract the target region image from the two-dimensional image with the collision point as the center, and uniformly input the size to generate the two-dimensional RGB image.

8. A vehicle, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor, the processor executing the program to implement the method for predicting head impact injuries of a vehicle as described in any one of claims 1-5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the method for predicting head impact injuries in a vehicle as described in any one of claims 1-5.

10. A computer program product, comprising a computer program, characterized in that, The computer program is executed to implement the method for predicting head impact injuries in a vehicle as described in any one of claims 1-5.

Images