Method for generating occupant injury prediction model, occupant injury prediction method and device

Abstract

The present application provides a method for generating an occupant injury prediction model, an occupant injury prediction method, and an apparatus. The generation method includes the following steps: generating multiple autonomous vehicle occupant collision simulation models; generating multiple parameterized simulated collision characteristic curves based on real vehicle collision characteristic curves; using the parameterized simulated collision characteristic curves to simulate the collision process of the autonomous vehicle occupant collision simulation model and establish an autonomous vehicle occupant injury database; using an existing traditional vehicle occupant injury database to pre-train the pre-trained model; and using the autonomous vehicle occupant injury database to formally train the pre-trained model to obtain an occupant injury prediction model. The technical solution of the present application integrates the collision characteristics of the autonomous vehicle's unique riding scenario on the basis of the traditional vehicle occupant injury prediction model, and can accurately predict the occupant injury of the autonomous vehicle.

Description

Technical Field

[0001] The present application belongs to the technical field of automobile safety performance prediction and optimization. Specifically, it provides a method for generating an occupant injury prediction model for an autonomous driving vehicle, an occupant injury prediction method, and an apparatus. Background Art

[0002] With the development of artificial intelligence (AI) technology, autonomous driving technology has gradually matured, and self-driving cars have initially achieved the ability to operate on real-world road networks. However, due to the increasing complexity of autonomous driving systems and the expansion of their designed operating areas, even self-driving cars equipped with numerous intelligent devices, including sensors and controllers, still face the risk of collision. To enhance occupant safety, it is particularly important to use occupant injury prediction algorithms to predict occupant injuries in near real time, enabling subsequent vehicle control decisions or the implementation of proactive protective measures.

[0003] At present, in order to improve the safety performance of automobiles, a method for predicting occupant injuries before an accident has emerged. Its technical solution is to collect information required for occupant injury prediction, including vehicle characteristics, occupant characteristics, etc., through on-board sensor devices under dangerous traffic conditions (situations where the vehicle cannot avoid a collision due to excessively high vehicle speed or too small distance to obstacles), and input it into a pre-trained automobile occupant injury model to predict the occupant injury situation and injury level. This can then guide the vehicle decision-making control system to select the optimal emergency path or adaptive restraint system for timely adjustment in the event of an imminent collision to minimize occupant injuries.

[0004] However, most existing methods are based on the vehicle characteristics, occupant characteristics and restraint configuration of traditional cars, and use deep learning or machine learning methods to establish occupant injury prediction models. However, for future self-driving cars, the car seat layout will be more diverse and the occupant posture in the car will have more possibilities. The above-mentioned diverse information makes the injury of occupants in self-driving cars after an accident significantly different from that of traditional cars. Therefore, there is an urgent need for an occupant injury prediction model and prediction method for self-driving cars that includes seat orientation information and seat back tilt information. Summary of the Invention

[0005] To address the problems in the prior art described above, the present application provides a method for generating an occupant injury prediction model for an autonomous vehicle. The occupant injury prediction model is used to predict occupant injuries when the autonomous vehicle is involved in a collision, and includes the following steps:

[0006] S1, construct multiple autonomous vehicle riding scenarios and generate multiple autonomous vehicle occupant collision simulation models;

[0007] S2, generating multiple parameterized simulation collision characteristic curves based on the real vehicle collision characteristic curve;

[0008] S3, using each of the parameterized simulated collision characteristic curves to simulate a collision process for each of the autonomous driving vehicle occupant collision simulation models, and establishing an autonomous driving vehicle occupant injury database based on the simulation results;

[0009] S4, builds a pre-training model and uses the existing traditional automobile occupant injury database for pre-training;

[0010] S5. Formal training is performed on the pre-trained model using the autonomous driving vehicle occupant injury database to obtain the occupant injury prediction model.

[0011] Furthermore, step S1 includes the following steps:

[0012] Based on seat orientation, seat back angle, occupant characteristics, and occupant restraint configuration, multiple autonomous vehicle riding scenarios are constructed through orthogonal design. A corresponding autonomous vehicle occupant collision simulation model is generated for each autonomous vehicle riding scenario.

[0013] Preferably, the seat orientation includes forward orientation, side orientation, and rear orientation; the seat back inclination includes normal posture inclination and relaxed posture inclination; the occupant characteristics include occupant gender; the occupant restraint configuration includes wearing a seat belt and equipped with an airbag, wearing a seat belt and not equipped with an airbag, not wearing a seat belt and equipped with an airbag, and not wearing a seat belt and not equipped with an airbag.

[0014] Preferably, the method for generating the occupant injury prediction model further includes the step of screening the autonomous driving vehicle riding plan based on the impact of seat orientation on the occupant restraint configuration.

[0015] Preferably, the method for generating the occupant injury prediction model further includes the step of screening the autonomous driving vehicle riding plan based on the influence of the seat back inclination angle on the occupant restraint configuration.

[0016] Furthermore, step S2 further includes the following steps:

[0017] S21, obtaining at least one real vehicle collision characteristic curve;

[0018] S22, for each real vehicle collision characteristic curve, perform the following steps in sequence:

[0019] S221, determining a parameterized approximate collision characteristic curve for replacing the actual vehicle collision characteristic curve based on a dynamic response characteristic criterion of a specific body part,

[0020] S222 , performing orthogonal design on the parameterized approximate collision characteristic curve to obtain a plurality of parameterized simulated collision characteristic curves corresponding to the real vehicle collision characteristic curve.

[0021] Furthermore, the criterion for the dynamic response characteristics of the specific body part is specifically: using the real vehicle collision characteristic curve and the alternative parameterized approximate collision characteristic curve to simulate the collision process of the autonomous driving vehicle occupant collision simulation model respectively; if the difference between the dynamic response curves of the specific body part simulated by the two is less than a preset threshold, then the alternative parameterized approximate collision characteristic curve is determined as the parameterized approximate collision characteristic curve used to replace the real vehicle collision characteristic curve.

[0022] Furthermore, step S3 further includes the following steps:

[0023] S31, obtaining any parameterized simulation collision characteristic curve and any autonomous driving vehicle occupant collision simulation model;

[0024] S32, using the parameterized simulated collision characteristic curve to simulate a collision process of the autonomous driving vehicle occupant collision simulation model, and obtaining damage data of various parts of the occupant's body;

[0025] S33, calculating a single damage index of the corresponding part based on the damage data of each part of the occupant's body;

[0026] S34, converting the individual injury indicators of various body parts of the occupant into probabilities corresponding to different injury levels, and determining the injury level of each body part of the occupant based on a preset probability threshold;

[0027] S35, return to execute step S31 until each parameterized simulation collision characteristic curve and each autonomous driving vehicle occupant collision simulation model are traversed, and finally the autonomous driving vehicle occupant injury database is obtained.

[0028] Preferably, step S3 further includes the step of adjusting the injury levels of various parts of the occupant's body based on the influence of the seat orientation and / or seat back inclination on the occupant restraint configuration.

[0029] Preferably, the pre-trained model includes a first pre-trained model and a second pre-trained model constructed using different deep learning network models; the occupant injury prediction model includes a first occupant injury prediction model and a second occupant injury prediction model.

[0030] This application also provides a method for predicting occupant injury when an autonomous vehicle collides, comprising the following steps:

[0031] The first step is to obtain the vehicle characteristic data, current collision condition characteristic data, occupant characteristic data, and constraint configuration characteristic data of the autonomous vehicle;

[0032] The second step is to convert the vehicle characteristic data, the current collision condition characteristic data, the occupant characteristic data, and the restraint configuration characteristic data of the autonomous vehicle into input data for an occupant injury prediction model, wherein the occupant injury prediction model is generated using the aforementioned occupant injury prediction model generation method;

[0033] The third step is to input the input data into the occupant injury prediction model to obtain a prediction result of the injury level of the occupants in the autonomous driving car.

[0034] The present application also provides an occupant injury prediction device for predicting occupant injuries when an autonomous vehicle collides, comprising:

[0035] An acquisition module, configured to acquire vehicle characteristic data, current collision condition characteristic data, occupant characteristic data, and constraint configuration characteristic data of the autonomous vehicle;

[0036] a conversion module, configured to convert vehicle characteristic data of the autonomous vehicle, current collision condition characteristic data, occupant characteristic data, and restraint configuration characteristic data into input data for an occupant injury prediction model, wherein the occupant injury prediction model is generated using the aforementioned occupant injury prediction model generation method;

[0037] The prediction module is used to input the input data into the occupant injury prediction model to obtain a prediction result of the injury level of the occupant in the autonomous driving vehicle.

[0038] The present application also provides a computer-readable storage medium storing an executable program, wherein the executable program can be used to execute the aforementioned method for generating an occupant injury prediction model.

[0039] The present application also provides a computer-readable storage medium storing an executable program, wherein the executable program can be used to execute the aforementioned occupant injury prediction method.

[0040] The technical solution of the present application uses the currently available massive database of occupant injuries of traditional automobiles to pre-train the prediction model, so that the prediction model has basic occupant injury prediction accuracy. At the same time, it uses measured collision data, combined with the unique seat orientation and backrest inclination layout of autonomous vehicles, to expand a small-data-volume automobile occupant injury database for autonomous vehicles, which is used to train the injury prediction model, so that the model obtained based on conventional data training further has the ability to predict occupant injuries under the unique riding conditions of autonomous vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a flowchart of an implementation method of a generation method of an occupant injury prediction model according to an embodiment of the present application;

[0042] Figure 2a is a schematic diagram of an autonomous vehicle occupant collision simulation model established according to multiple forward-facing autonomous vehicle riding scenarios in some preferred embodiments;

[0043] Figure 2b is a schematic diagram of an autonomous vehicle occupant collision simulation model established according to multiple side-facing autonomous vehicle riding scenarios in some preferred embodiments;

[0044] Figure 2c is a schematic diagram of an autonomous vehicle occupant collision simulation model established according to multiple rear-facing autonomous vehicle riding scenarios in some preferred embodiments;

[0045] Figure 3 A schematic diagram of determining parameters of a parameterized approximate collision characteristic curve using a dynamic response curve of an occupant's head as a criterion in some embodiments;

[0046] Figure 4 A parameterized approximate collision characteristic curve generated in some embodiments and a comparison thereof with a real vehicle collision characteristic curve;

[0047] Figure 5 is a schematic diagram of multiple parameterized simulated collision characteristic curves generated in some embodiments;

[0048] Figure 6 A flowchart for implementing training of an occupant injury prediction model in some preferred embodiments;

[0049] Figure 7 Schematic diagram of the framework of the occupant injury prediction device according to an embodiment of the present application. DETAILED DESCRIPTION

[0050] Hereinafter, the present application will be further described based on preferred embodiments with reference to the accompanying drawings.

[0051] In addition, for ease of understanding, various components on the drawings are enlarged or reduced, but this practice is not intended to limit the scope of protection of this application. In the description of the embodiments of the present application, it should be noted that if the terms "upper", "lower", "inside", "outside" and the like indicate an orientation or positional relationship based on the orientation or positional relationship shown in the drawings, or are the orientation or positional relationship in which the products of the embodiments of the present application are usually placed when in use, it is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. In addition, in the description of the present application, in order to distinguish different units, words such as first and second are used in this specification, but these are not limited by the order of manufacture, nor can they be understood as indicating or implying relative importance.

[0052] The present application provides a method for generating an occupant injury prediction model through an embodiment. The occupant injury prediction model generated by this method can predict the injury status of occupants in an autonomous vehicle under dangerous traffic conditions. Figure 1 A flow chart of the generation method is shown in some embodiments, such as Figure 1 As shown, the generation method includes the following steps:

[0053] S1, construct multiple autonomous vehicle riding scenarios and generate multiple autonomous vehicle occupant collision simulation models;

[0054] S2, generating multiple parameterized simulation collision characteristic curves based on the real vehicle collision characteristic curve;

[0055] S3, using each of the parameterized simulated collision characteristic curves to simulate a collision process for each of the autonomous driving vehicle occupant collision simulation models, and establishing an autonomous driving vehicle occupant injury database based on the simulation results;

[0056] S4, builds a pre-training model and uses the existing traditional automobile occupant injury database for pre-training;

[0057] S5. Formal training is performed on the pre-trained model using the autonomous driving vehicle occupant injury database to obtain the occupant injury prediction model.

[0058] The following describes steps S1 to S5 in detail with reference to the accompanying drawings and specific embodiments.

[0059] In an embodiment of the present application, step S1 is used to construct a riding scenario adapted to the vehicle characteristics of the autonomous vehicle, and generate a corresponding autonomous vehicle occupant collision simulation model for each riding scenario.

[0060] Compared with traditional cars, self-driving cars do not require drivers to always face the direction of the car's travel, so the arrangement of their car seats is more diverse, and thus the passengers' postures during driving have more possibilities. For example, in some specific embodiments, the car will be equipped with rotatable seats so that passengers can freely control the direction of the seats in the car to communicate with other passengers. In addition, passengers may adopt the same sitting posture as in traditional cars during driving, or they can adopt a relaxed posture close to lying on their backs by adjusting the seat back angle.

[0061] Occupants of different genders use different occupant restraint configurations (in the embodiments of this application, the occupant restraint configuration refers to the configuration for safety restraint of the occupants, including whether the occupants wear seat belts and whether the vehicle is equipped with airbags) under different seat orientations and seat back angles, thus forming different autonomous driving car riding schemes. Then, for each autonomous driving car riding scheme, a finite element-multi-rigid body coupling model is established (where the occupants use a finite element model and the car uses a multi-rigid body model). After the model is established, the material properties of each part and the contact conditions between different parts are defined, and a collision simulation model with each autonomous driving car occupant can be obtained.

[0062] In some preferred embodiments, seat orientations include forward (0°), side (90°), and rearward (180°); seatback tilt includes normal (21° relative to vertical) and relaxed (48° relative to vertical); occupant characteristics include occupant gender (male, female); and occupant restraint configurations include: seatbelt with airbag, seatbelt without airbag, unbelt with airbag, and unbelt without airbag. Based on these factors, multiple autonomous vehicle passenger scenarios can be constructed through orthogonal design.

[0063] Through the above-mentioned orthogonal design, it is possible to completely generate riding scenarios covering all values ​​of the above-mentioned factors. However, by analyzing the influence of the seat orientation and / or seat back inclination on the occupant restraint configuration in each scenario, it can be seen that among the riding scenarios generated by the above steps, there are riding scenarios that do not require collision process simulation. Therefore, in some preferred embodiments, after the above-mentioned orthogonal design is completed, it also includes the step of screening the riding scenarios of the autonomous driving vehicle based on the influence of the seat orientation on the occupant restraint configuration, and the step of screening the riding scenarios of the autonomous driving vehicle based on the influence of the seat back inclination on the occupant restraint configuration, thereby effectively reducing the workload of subsequent modeling and collision process simulation.

[0064] For example, when the seat is facing backward (i.e., 180°), the orientation of the occupant and the seat is opposite to the direction of the car's movement. When the autonomous vehicle collides head-on, the occupant will not come into contact with the traditional airbag. Whether the airbag is deployed will not affect the simulation results. Therefore, the 8 riding scenarios in which the airbag is deployed when the seat is facing backward should be eliminated, and finally 40 autonomous vehicle riding scenarios are obtained. Figures 2a to 2c An autonomous vehicle occupant collision simulation model established based on the 40 screened autonomous vehicle riding scenarios is shown according to the front, side, and rear orientations, respectively, as well as a schematic diagram of simulating the collision process.

[0065] In an embodiment of the present application, step S2 is expanded based on real vehicle collision data to construct multiple parameterized simulated collision characteristic curves for use in simulating the collision process of the autonomous vehicle occupant collision simulation model established in step S1. In some preferred embodiments, step S2 further includes the following steps:

[0066] S21, obtaining at least one real vehicle collision characteristic curve;

[0067] S22, for each real vehicle collision characteristic curve, perform the following steps in sequence:

[0068] S221, determining a parameterized approximate collision characteristic curve for replacing the actual vehicle collision characteristic curve based on a dynamic response characteristic criterion of a specific body part,

[0069] S222 , performing orthogonal design on the parameterized approximate collision characteristic curve to obtain a plurality of parameterized simulated collision characteristic curves corresponding to the real vehicle collision characteristic curve.

[0070] In step S21 , real car collision test data may be obtained through the Internet or other channels, and a real vehicle collision characteristic curve may be generated based on the test data.

[0071] The real vehicle collision characteristic curve can accurately depict the change of the collision deceleration (g) of a real vehicle over time when it collides. However, on the one hand, the change of this curve often contains violent fluctuations, and the curve shape is relatively complex, resulting in excessive calculation when applying it to the autonomous driving vehicle occupant collision simulation model to simulate the collision process, even exceeding an acceptable level; on the other hand, due to experimental cost considerations, generally only a limited number of collision experiments can be conducted on real vehicles, and it is difficult to obtain data that can cover a wide range of various collision situations. Therefore, in the embodiment of the present application, it is necessary to parameterize and orthogonalize the real vehicle collision characteristic curve through step S22 to simplify the curve shape and cover more vehicle collision situations, thereby effectively increasing the vehicle collision process simulation data to enrich the data set used to train the occupant injury prediction model.

[0072] Specifically, step S221 is used to determine a parameterized approximate collision characteristic curve that replaces the actual vehicle collision characteristic curve. In some preferred embodiments, the parameterized approximate collision characteristic curve can be expressed as follows using two parameters: the collision pulse peak value (i.e., collision deceleration) and the collision pulse duration:

[0073]

[0074] Among them, a veh (t) represents the vehicle collision pulse, A amp Indicates the magnitude of the vehicle collision pulse (in gravitational acceleration g), θ dur Vehicle collision pulse duration (in ms).

[0075] In the embodiment of the present application, the parameter A of the parameterized approximate collision characteristic curve is amp and θ dur The dynamic response characteristic criterion is determined based on the dynamic response characteristic of a specific body part. Specifically, a specific body part, such as the head and neck, can be selected, and the dynamic response curve of the body part can be calculated using the real vehicle collision characteristic curve and the alternative parameterized approximate collision characteristic curve respectively. If the difference between the two is less than a preset threshold, the alternative parameterized approximate collision characteristic curve replaces the real vehicle collision characteristic curve. Otherwise, the parameters are replaced to regenerate the alternative parameterized approximate collision characteristic curve, and the above steps are repeated.

[0076] In a specific embodiment, the actual vehicle collision characteristic curve is generated by experimental data of a Toyota Yaris undergoing a frontal collision at a speed of 56 km / h. Figure 3 A schematic diagram showing the use of the dynamic response curve of the occupant's head as a criterion to determine the parameters is shown. Figure 4The parameterized approximate crash characteristic curve determined in the above manner and its comparison with the actual vehicle crash characteristic curve are shown.

[0077] Furthermore, in step S222 , the parameters of the above-mentioned parameterized approximate collision characteristic curve are expanded according to a certain value interval, thereby obtaining a plurality of parameterized simulated collision characteristic curves. Figure 5 Shown with Figure 4 The parametric approximate collision characteristic curve in is used as a benchmark, with the collision pulse peak ranging from 10g to 60g (with a sampling interval of 10g) and the duration ranging from 60ms to 160ms (with a sampling interval of 20ms). This is a schematic diagram of multiple parametric simulation collision characteristic curves generated by orthogonal design.

[0078] After constructing the model library and collision characteristic curve library for simulating the collision process of the autonomous driving vehicle through steps S1 and S2 respectively, the collision process simulation can be performed in step S3. In the embodiment of the present application, step S3 includes the following steps:

[0079] S31, obtaining any parameterized simulation collision characteristic curve and any autonomous driving vehicle occupant collision simulation model;

[0080] S32, using the parameterized simulated collision characteristic curve to simulate a collision process of the autonomous driving vehicle occupant collision simulation model, and obtaining damage data of various parts of the occupant's body;

[0081] S33, calculating a single damage index of the corresponding part based on the damage data of each part of the occupant's body;

[0082] S34, converting the individual injury indicators of various body parts of the occupant into probabilities corresponding to different injury levels, and determining the injury level of each body part of the occupant based on a preset probability threshold;

[0083] S35, return to execute step S31 until each parameterized simulation collision characteristic curve and each autonomous driving vehicle occupant collision simulation model are traversed, and finally the autonomous driving vehicle occupant injury database is obtained.

[0084] The above simulation process is described below in conjunction with specific embodiments.

[0085] In some specific embodiments, the injuries to the occupant's head, neck, and chest during a collision can be simulated. By using any parameterized simulated collision characteristic curve in an implementation manner known to those skilled in the art to simulate the collision process of any autonomous driving vehicle occupant collision simulation model, damage data of the above-mentioned various parts can be obtained, for example: time series data of the occupant's head cardiac acceleration from the start time to the end time of the collision, time series data of the occupant's neck axial force and bending moment, time series data of the occupant's chest compression, time series data of the chest acceleration, etc.

[0086] In some specific embodiments, the individual injury indicators of the above-mentioned various body parts can be determined by the following methods:

[0087] (1) Head Injury Criterion (HIC)

[0088] HIC represents the skull damage by calculating the acceleration at the center of mass of the occupant's head. The HIC value can be calculated using the time series data of the occupant's head center of mass acceleration in the collision simulation data according to the following formula:

[0089]

[0090] Where a(t) represents the resultant acceleration of the head center of mass (g), t1 and t2 are any two time points in the collision acceleration curve, which are used to calculate the start and end times of HIC. The difference between t1 and t2 is less than a preset threshold. The commonly used preset threshold is 15 ms, that is, t1-t2≤15 ms.

[0091] (2) Neck Injury Criterion ij )

[0092] N ij It is calculated by combining the axial force and bending moment of the occupant's neck during the collision. First, the time series data of the axial force and bending moment of the occupant's neck in the collision simulation data and the type of neck load condition corresponding to each collision moment are obtained. The neck load condition is divided into two categories: buckling load and extension load. Then, according to the axial force and bending moment of the occupant's neck corresponding to each collision moment, the maximum value is calculated according to the following formula. ij value:

[0093]

[0094] Among them, F z Indicates the neck axial force, M y Denotes the bending moment, F intIndicates the intercept value corresponding to the axial force of the neck. When the axial force is expressed as a tensile force, F int The value is 6806N. When the axial force is expressed as a compressive force, F int The value is 6160N, M int Indicates the intercept value corresponding to the bending moment. When the neck load condition type is buckling, M int The value is 310N·m. When the neck is in the extended state, M int The value is 135N·m.

[0095] (3) Chest compressions comp )

[0096] C comp The maximum compression degree of the chest of the occupant during the collision is obtained by first obtaining the time series data of the chest thickness in the collision simulation data and determining the minimum value of the chest thickness. Then, the difference between the chest thickness value at the initial moment of the collision and the minimum value of the chest thickness is calculated to obtain the chest compression C. comp .

[0097] In some other embodiments, a single injury indicator may also include injury indicators of other parts of the body in addition to the above indicators, for example: obtaining an indicator of femoral injury based on the axial force and bending moment time series data at the occupant's femur, obtaining an indicator of tibial injury based on the axial force and bending moment time series data at the tibia, obtaining an indicator of abdominal injury based on the abdominal pressure time series data, etc.

[0098] In the embodiment of the present application, the above HIC, N ij and C comp After that, it is further converted into the probabilities corresponding to different damage levels. When the probability corresponding to a damage level is greater than a certain probability threshold (for example, 25%), it is determined that damage of that level has occurred.

[0099] Specifically, HIC can be converted into the probability corresponding to different damage levels using the following formula:

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106] Specifically, N can be calculated by the following formula ij Converted into the probability corresponding to different damage levels:

[0107]

[0108]

[0109]

[0110]

[0111] Specifically, C can be calculated by the following formula comp Converted into the probability corresponding to different damage levels:

[0112]

[0113]

[0114]

[0115]

[0116] The above-described method for determining the injury level of various body parts of an occupant based on individual injury indicators has been widely used in collision damage assessments for conventional vehicles. In the embodiments of the present application, because the seating arrangement of occupants in some riding scenarios differs from that in conventional vehicles, in some preferred embodiments, step S3 further includes adjusting the injury level of various body parts of the occupant based on the impact of seat orientation and / or seatback recline on the occupant restraint configuration. For example, when the seat orientation is rearward and the seatback recline is at a relaxed position, the injuries to the head and neck are less severe than when the seat orientation is forward and the seatback recline is at a normal position under the same acceleration. Therefore, the injury level assessment for the head and neck in this riding scenario can be reduced, making the damage assessment results more consistent with the vehicle characteristics of autonomous vehicles.

[0117] After performing the above steps in real time for each parameterized simulation collision characteristic curve and each autonomous driving vehicle occupant collision simulation model, the autonomous driving vehicle occupant injury database can be obtained.

[0118] Although the above-mentioned autonomous vehicle occupant injury database can accurately reflect the occupant injuries when the autonomous vehicle collides, since it is only expanded through limited measured collision data, there is a problem of small data volume when it is used to train the injury prediction model. For this reason, in the embodiment of the present application, the prediction model is first pre-trained with the help of the massive traditional vehicle occupant injury database that can be obtained at present, so that the prediction model has basic occupant injury prediction accuracy, and then formal training is continued using the data reflecting the occupant injury characteristics of the autonomous vehicle obtained in step S3, so as to further enable the prediction model to have the ability to predict occupant injuries under the riding conditions unique to the autonomous vehicle.

[0119] In some preferred embodiments, the pre-training model includes a pre-training model 1 and a pre-training model 2, wherein the pre-training model 1 adopts an RNN model architecture and the pre-training model 2 adopts a CNN model architecture. The model architecture and training process are as follows: Figure 6 The input of the training samples of pre-trained model 1 and pre-trained model 2 is the vehicle collision characteristics, occupant characteristics and restraint configuration, and the output is the occupant body injury level.

[0120] A large number of existing traditional automobile occupant injury databases can be obtained through the Internet or professional databases and used to pre-train the pre-trained models. In the embodiment of the present application, the ratio of the training set to the test set is 8:2. Pre-trained model 1 adopts an RNN model architecture based on Bi-LSTM units, and introduces an attention mechanism to improve the performance of pre-trained model 1. Pre-trained model 2 adopts a CNN model architecture based on TCN. During training, pre-trained models 1 and 2 use the ADAM optimizer and the cross-entropy loss function, introduce L2 regularization to prevent model overfitting, and use a dropout layer and an early termination callback function.

[0121] In the specific implementation process, the vehicle collision pulse curve can be discretized into 1×100 vector data, and the occupant gender and restraint configuration are replaced by scalars (for example, 0 for male and 1 for female). All input variables are formed into a new tensor through the Embedding layer, that is, the original input information is converted into a new high-dimensional matrix according to the mapping relationship, effectively amplifying the input features.

[0122] The output occupant injury level is converted into a binary vector using character-level one-hot encoding, associating different AIS levels with a unique integer index. Pre-trained model 1 uses an RNN model architecture based on Bi-LSTM units, enabling it to extract all contextual information from time series data. The introduction of an attention mechanism significantly improves model performance by retaining the intermediate output results of the Bi-LSTM encoder for the input sequence, selectively learning these inputs during training, and associating the input sequence with the intermediate results during output. Pre-trained model 2 uses a TCN-based CNN model architecture, using three TCN blocks, each containing two dilated causal convolutional layers. The dilation factor d is the same for all three TCN blocks, and within each TCN block, the dilation factor d increases exponentially with the depth of the network. A batch normalization layer is added after each convolutional layer to improve model training efficiency and prevent overfitting.

[0123] Furthermore, a parameter-based transfer learning method was used to formally train pre-trained Model 1 and Model 2 using the autonomous vehicle occupant injury database obtained in step S3, ultimately resulting in occupant injury prediction Model 1 and occupant injury prediction Model 2 for autonomous vehicles. The training set and test set ratio was 8:2. The inputs to Occupant Injury Prediction Model 1 and Occupant Injury Prediction Model 2 were the vehicle collision pulse curve, occupant characteristics, restraint configuration, seat orientation, and seatback angle, and the output was the occupant body injury level.

[0124] Compared with the pre-trained model, the final occupant injury prediction model for autonomous vehicles has two additional influencing factors: seat back tilt and seat orientation. The seat back tilt and seat orientation are also replaced by scalars (for example, a 0° seat orientation is represented as 0, a 90° seat orientation is represented as 1, and a 180° seat orientation is represented as 2).

[0125] In the specific implementation process, several factors that have a significant impact on occupant injuries during a frontal collision of an autonomous vehicle are mainly considered, including vehicle collision characteristics (relative collision speed and relative angle), vehicle characteristics (including seat orientation and seat back tilt), occupant characteristics (occupant gender), and restraint configuration usage (use of seat belts and airbags). As for some detailed factors such as occupant height and weight, it is difficult to uniformly represent them in a certain way and their impact on occupant injuries is relatively small, so they can be ignored. In a specific embodiment, using this method, a pre-trained model is pre-trained based on a traditional occupant injury database, and then the pre-trained model is formally trained based on a transfer learning method using a self-built small autonomous vehicle occupant injury database. With only 1,440 training data, a model accuracy of about 70% is obtained, and the training time only requires 12 minutes of training time (CPU: Ryzen Threadripper 3990x 64-core processor × 128, GPU: NVIDIA Corporation). Experimental results show that this method can be used to train an occupant injury prediction model for autonomous vehicles, and the model training accuracy and efficiency are relatively ideal.

[0126] The present application also provides an occupant injury prediction method through an embodiment, for predicting occupant injuries when an autonomous vehicle collides, the prediction method comprising the following steps:

[0127] The first step is to obtain the vehicle characteristic data, current collision condition characteristic data, occupant characteristic data, and constraint configuration characteristic data of the autonomous vehicle;

[0128] The second step is to convert the vehicle characteristic data, the current collision condition characteristic data, the occupant characteristic data, and the restraint configuration characteristic data of the autonomous vehicle into input data for an occupant injury prediction model, wherein the occupant injury prediction model is generated using the aforementioned occupant injury prediction model generation method;

[0129] The third step is to input the input data into the occupant injury prediction model to obtain a prediction result of the injury level of the occupants in the autonomous driving car.

[0130] The present application also provides an occupant injury prediction device through an embodiment, such as Figure 7 As shown, the prediction device includes:

[0131] An acquisition module, configured to acquire vehicle characteristic data, current collision condition characteristic data, occupant characteristic data, and constraint configuration characteristic data of the autonomous vehicle;

[0132] a conversion module, configured to convert vehicle characteristic data of the autonomous vehicle, current collision condition characteristic data, occupant characteristic data, and restraint configuration characteristic data into input data for an occupant injury prediction model, wherein the occupant injury prediction model is generated using the aforementioned occupant injury prediction model generation method;

[0133] The prediction module is used to input the input data into the occupant injury prediction model to obtain a prediction result of the injury level of the occupant in the autonomous driving vehicle.

[0134] Specifically, each module in the occupant injury prediction device can be implemented in whole or in part through software, hardware, or a combination thereof. Each of these modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0135] The present application also provides, through embodiments, a computer-readable storage medium storing an executable program capable of being used to execute the aforementioned method for generating an occupant injury prediction model. Specifically, the computer-readable storage medium may be a non-transitory computer-readable storage medium, such as a high-speed random access memory, storing a computer-executable program capable of being executed by a processor or controller to implement all or part of the steps of the aforementioned method for generating an occupant injury prediction model.

[0136] The present application also provides, through embodiments, a computer-readable storage medium storing an executable program capable of executing the aforementioned occupant injury prediction method. Specifically, the computer-readable storage medium may be a non-transitory computer-readable storage medium, such as a high-speed random access memory, storing a computer-executable program capable of being executed by a processor or controller to implement all or part of the steps of the aforementioned occupant injury prediction method.

[0137] The above is a detailed introduction to the specific implementation methods of the present application. For those skilled in the art, several improvements and modifications can be made to the present application without departing from the principles of the present application. These improvements and modifications also fall within the scope of protection of the claims of the present application.

Claims

1. A method for generating an occupant injury prediction model, wherein the occupant injury prediction model is used to predict occupant injuries when an autonomous vehicle is involved in a collision, wherein: The following steps are involved: S1, construct multiple autonomous vehicle riding scenarios and generate multiple autonomous vehicle occupant collision simulation models; S2, generating multiple parameterized simulation collision characteristic curves based on the real vehicle collision characteristic curve; S3, using each of the parameterized simulated collision characteristic curves to simulate a collision process for each of the autonomous driving vehicle occupant collision simulation models, and establishing an autonomous driving vehicle occupant injury database based on the simulation results; S4, builds a pre-training model and uses the existing traditional automobile occupant injury database for pre-training; S5, formally training the pre-trained model using the autonomous driving vehicle occupant injury database to obtain the occupant injury prediction model; Step S2 further includes the following steps: S21, obtaining at least one real vehicle collision characteristic curve; S22, for each real vehicle collision characteristic curve, perform the following steps in sequence: S221, determining a parameterized approximate collision characteristic curve for replacing the actual vehicle collision characteristic curve based on a dynamic response characteristic criterion of a specific body part, wherein parameters of the parameterized approximate collision characteristic curve include a collision pulse peak value and a collision duration. S222, performing orthogonal design on the parameterized approximate collision characteristic curve to obtain multiple parameterized simulated collision characteristic curves corresponding to the real vehicle collision characteristic curve. Specifically, based on the parameterized approximate collision characteristic curve, the parameters are expanded according to a certain value interval, and the multiple parameterized simulated collision characteristic curves are generated by orthogonal design.

2. The method for generating an occupant injury prediction model according to claim 1, wherein: Step S1 further includes the following steps: Create multiple autonomous vehicle occupant scenarios through orthogonal design based on seat orientation, seatback angle, occupant characteristics, and occupant restraint configurations; A corresponding autonomous vehicle occupant collision simulation model is generated based on each autonomous vehicle riding scenario.

3. The method for generating an occupant injury prediction model according to claim 2, wherein: The seat orientation includes front orientation, side orientation, and rear orientation; The seat back inclination angle includes a normal posture inclination angle and a relaxed posture inclination angle; The occupant characteristics include the occupant's gender; The occupant restraint configurations include seatbelts with airbags, seatbelts without airbags, unbelts with airbags, and unbelts without airbags.

4. The method for generating an occupant injury prediction model according to claim 3, wherein: Also included is the step of screening the autonomous vehicle ride options based on the effect of seat orientation on occupant restraint configuration.

5. The method for generating an occupant injury prediction model according to claim 3, wherein: Also included is the step of screening the autonomous vehicle ride options based on the effect of seat back recline on occupant restraint configuration.

6. The method for generating an occupant injury prediction model according to claim 1, wherein: The dynamic response characteristic criterion of the specific body part is specifically: The real vehicle collision characteristic curve and the alternative parameterized approximate collision characteristic curve are used to simulate the collision process of the autonomous driving vehicle occupant collision simulation model respectively. If the difference between the dynamic response curves of the specific body part simulated by the two is less than a preset threshold, the alternative parameterized approximate collision characteristic curve is determined as the parameterized approximate collision characteristic curve used to replace the real vehicle collision characteristic curve.

7. The method for generating an occupant injury prediction model according to claim 1, wherein: Step S3 further includes the following steps: S31, obtaining any parameterized simulation collision characteristic curve and any autonomous driving vehicle occupant collision simulation model; S32, using the parameterized simulated collision characteristic curve to simulate a collision process of the autonomous driving vehicle occupant collision simulation model, and obtaining damage data of various parts of the occupant's body; S33, calculating a single damage index of the corresponding part based on the damage data of each part of the occupant's body; S34, converting the individual injury indicators of various body parts of the occupant into probabilities corresponding to different injury levels, and determining the injury level of each body part of the occupant based on a preset probability threshold; S35, return to execute step S31 until each parameterized simulation collision characteristic curve and each autonomous driving vehicle occupant collision simulation model are traversed, and finally the autonomous driving vehicle occupant injury database is obtained.

8. The method for generating an occupant injury prediction model according to claim 7, wherein: Step S3 also includes the step of adjusting the injury level of various parts of the occupant's body based on the influence of the seat orientation and / or the seat back inclination on the occupant restraint configuration.

9. The method for generating an occupant injury prediction model according to claim 1, wherein: The pre-training model includes a first pre-training model and a second pre-training model constructed using different deep learning network models; The occupant injury prediction model includes a first occupant injury prediction model and a second occupant injury prediction model.

10. A passenger injury prediction method for predicting passenger injuries when an autonomous vehicle collides, characterized in that: The following steps are involved: The first step is to obtain the vehicle characteristic data, current collision condition characteristic data, occupant characteristic data, and constraint configuration characteristic data of the autonomous vehicle; The second step is to convert the vehicle characteristic data, the current collision condition characteristic data, the occupant characteristic data, and the restraint configuration characteristic data of the autonomous vehicle into input data of an occupant injury prediction model, wherein the occupant injury prediction model is generated using the occupant injury prediction model generation method according to claim 1; The third step is to input the input data into the occupant injury prediction model to obtain a prediction result of the injury level of the occupants in the autonomous driving car.

11. A passenger injury prediction device for predicting passenger injuries when an autonomous vehicle collides, characterized in that: include: An acquisition module, configured to acquire vehicle characteristic data, current collision condition characteristic data, occupant characteristic data, and constraint configuration characteristic data of the autonomous vehicle; a conversion module, configured to convert vehicle characteristic data of the autonomous vehicle, current collision condition characteristic data, occupant characteristic data, and restraint configuration characteristic data into input data for an occupant injury prediction model, wherein the occupant injury prediction model is generated using the method for generating an occupant injury prediction model according to claim 1; The prediction module is used to input the input data into the occupant injury prediction model to obtain a prediction result of the injury level of the occupant in the autonomous driving vehicle.

12. A computer-readable storage medium storing an executable program, characterized in that: The executable program can be used to execute the method for generating an occupant injury prediction model according to claim 1 .

13. A computer-readable storage medium storing an executable program, characterized in that: The executable program can be used to execute the occupant injury prediction method according to claim 10 .

Images