A method for assessing the driving risks of traffic vehicles

Abstract

This invention discloses a method for assessing the driving risk of traffic vehicles, comprising the following steps: extracting target scene data from an original publicly available dataset; constructing an exposure quantification Transformer model to predict the driving intentions of surrounding vehicles and then calculating the exposure E of the main vehicle's driving risk; calculating the severity S of the main vehicle's driving risk, considering vehicle type; calculating the controllability C of the driving risk between the main vehicle and any surrounding vehicle based on the magnitude of the main vehicle's acceleration change; multiplying the calculated driving risk exposure E, controllability C, and severity S to obtain the risk assessment result R between the two vehicles; and taking the maximum value of the risk assessment result R between the main vehicle and all surrounding vehicles as the driving risk assessment result of the main vehicle. This invention decouples and quantitatively analyzes the influencing factors of traffic vehicle driving risk, quantifies the three indicators of risk assessment, and finally obtains the total driving risk value.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation, specifically a method for assessing the driving risks of traffic vehicles. Background Technology

[0002] Accurate risk assessment during vehicle operation is a crucial guarantee for driving safety. It can provide drivers with real-time safe driving guidance, help predict potential accident risks, and take corresponding measures to avoid them. In the event of an accident, accurate risk assessment can help minimize personal injury and property damage, which is a key focus that people have always paid attention to.

[0003] Current risk assessment methods primarily rely on ISO 26262, dividing risk assessment into three modules: exposure, severity, and controllability, and using a qualitative analysis approach similar to low, medium, and high risk classifications. Qualitative risk analysis depends heavily on subjective judgment and experience, making it susceptible to personal biases and emotional factors. It cannot accurately quantify the magnitude of driving risks, hindering comparison, evaluation, and optimization, and limiting effective decision support. Furthermore, because qualitative analysis lacks clear data support and quantifiable results, it may limit drivers' ability to make rational and informed decisions when facing risks, potentially leading to underestimation or overestimation of risks, increasing the likelihood of traffic accidents. For complex driving environments and situations, qualitative analysis often fails to provide sufficient detail and depth, making it difficult to fully understand and assess the impact of risk factors. Therefore, quantitative analysis of driving risks and obtaining specific risk values ​​can help improve drivers' safety awareness and driving decision-making abilities, thereby reducing traffic accidents.

[0004] Currently, methods for quantitative assessment of driving risks mainly include risk value calculation models and event tree analysis. Sze et al., in their paper "Development of a risk assessment model for road traffic crashes using Bayesian networks," employed a Bayesian network model, combining multiple factors such as road characteristics, traffic flow, weather conditions, vehicle condition, and driver behavior to quantitatively assess the probability of road traffic accidents. BingLi et al., in their paper "A Dynamic Collision Risk Assessment Model for the Traffic Flow on Expressways in Urban Agglomerations in North China," used decision trees to identify the main influencing factors of dynamic traffic operation risks, and then used these main influencing factors as input variables for an ANFIS model to perform dynamic risk assessment modeling, thereby obtaining the dynamic risk value of traffic operation on urban agglomeration expressway sections. Vuong et al., in their paper "A Safety Evaluation Model of Intersections under Mixed Traffic Conditions Using Traffic Conflicts and Cloud,"... The "Model" incorporates three indicators—collision time, collision speed, and deceleration rate to avoid collision—into the risk assessment model and uses a cloud model to transform qualitative concepts into quantitative ones. Collision time reflects the exposure of risk assessment, while collision speed and deceleration rate reflect severity, thus achieving the quantification of driving risks. Flintsch et al., in their paper "Accident risk quantification in urban road traffic," studied the problem of accident risk quantification in urban road traffic. Through event tree analysis, they analyzed various factors and events that may lead to traffic accidents and quantified the risk of traffic accidents.

[0005] Analysis of existing methods for quantitative assessment of driving risks reveals the following problems: First, numerous risk value calculation models consider different factors, resulting in varying main factors and thus low universality of the quantification methods. Second, the components of risk values ​​differ. For the three modules of risk assessment—exposure, severity, and controllability—most existing methods only consider exposure (the probability of a dangerous event) and severity (the damage caused by the dangerous event), while rarely considering controllability, i.e., how appropriate actions by the driver can reduce damage or prevent an accident. This leads to a significant discrepancy between the calculated risk value and the actual risk. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention provides a method for assessing traffic vehicle driving risks. This method decouples and quantitatively analyzes the influencing factors of traffic vehicle driving risks, quantifies the three risk assessment indicators, and ultimately obtains the total driving risk value. Specifically, exposure quantification uses a Transformer model to predict the driving intentions of surrounding vehicles and obtains the probability distribution of all intentions; the probability value of the driving intention is used as the exposure degree of the dangerous event caused by that intention. Severity quantification uses the probability that the maximum simplified damage level considering vehicle type is greater than or equal to 2. Controllability quantification uses the change in deceleration to avoid a collision. By quantifying driving risks according to the three modules of risk assessment and decoupling different factors affecting risk levels, the contribution of different factors to the risk value can be clearly obtained. Based on the quantitative analysis results, appropriate driving behaviors can be adopted to reduce driving risks and improve road traffic safety.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A method for assessing the driving risks of traffic vehicles includes the following steps:

[0009] Step 1: Extract target scene data from the original public dataset. The target scene data includes the main vehicle feature sequence, the weekly vehicle feature sequence of each weekly vehicle, the main vehicle's left lane label, the main vehicle's right lane label, and add the main vehicle behavior label.

[0010] Step 2: Construct an exposure quantification Transformer model, predict the driving intention of the surrounding vehicles through the model, and then calculate the exposure degree E of the main vehicle's driving risk;

[0011] Step 3: Calculate the severity S of the main vehicle driving risk, taking into account the vehicle type;

[0012] Step 4: Based on the magnitude of the change in the acceleration of the main vehicle, calculate the controllability C of the driving risk between the main vehicle and any surrounding vehicle;

[0013] Step 5: Multiply the driving risk exposure E, controllability C, and severity S calculated in Steps 2 to 4 to obtain the risk assessment result R between the two vehicles;

[0014] Step 6: Take the maximum value of the risk assessment result R between the main vehicle and all surrounding vehicles as the driving risk assessment result of the main vehicle.

[0015] Furthermore, in step one, the principle for extracting target scene data is as follows:

[0016] First, based on the driving status of the vehicles, they are classified into lane-keeping vehicles or lane-changing vehicles, and the classified vehicles are taken as the main vehicles.

[0017] If the main vehicle is a lane-keeping vehicle, extract its feature scene data for any 2 seconds;

[0018] If the main vehicle is a lane-changing vehicle, then extract the feature scene data 2 seconds before its lane-changing start point. The lane-changing start point is set as the moment when the lateral distance between the lane-changing vehicle and the lane centerline exceeds half the vehicle width. The lane-changing start point is defined as follows:

[0019]

[0020] In the formula, Indicates the lateral distance between the vehicle changing lanes and the center line of the lane. Indicates the width of the vehicle.

[0021] Furthermore, in step one, the target scene data specifically includes:

[0022] The main vehicle feature sequence includes four features: the main vehicle's horizontal coordinate, the main vehicle's vertical coordinate, the main vehicle's longitudinal speed, and the main vehicle's offset distance relative to the center line of its lane.

[0023] The feature sequence of each car includes three features: the horizontal coordinate, the vertical coordinate, and the longitudinal speed of the car. For the main car, there are 6 cars: left front, front, right front, left rear, rear, and right rear, for a total of 18 car feature sequence features.

[0024] The left and right lane labels for the main vehicle are added based on road structure information and the lane information of the main vehicle.

[0025] The main vehicle behavior label is used to characterize whether the main vehicle is classified as a lane-keeping vehicle or a lane-changing vehicle. The main vehicle behavior label is not used as input to the exposure metric Transformer model in step two.

[0026] Furthermore, step two includes:

[0027] S21. Construct an exposure metric Transformer model and train the model;

[0028] S22. Apply the trained exposure metric Transformer model to the target scene data, and calculate the exposure degree E of the driving risk of the main vehicle by predicting the driving intention of the surrounding vehicles:

[0029] For the main vehicle, the vehicle being tested for predicting driving intention is any of the surrounding vehicles B. The target scene data of surrounding vehicle B when it is the main vehicle is input into the exposure quantification Transformer model to obtain the probabilities of the three driving intentions of surrounding vehicle B. The probability value of the driving intention of surrounding vehicle B that has the greatest impact on the driving of the main vehicle is taken as the exposure value of the driving risk brought by surrounding vehicle B to the main vehicle.

[0030] S23. By predicting the driving intentions of all surrounding vehicles of the main vehicle, the exposure of driving risks to the main vehicle from any surrounding vehicle is obtained.

[0031] Furthermore, in step S22, the probability of the three driving intentions refers to the probability distribution of the three driving intentions of lane keeping, left lane changing, and right lane changing at the next time step under the influence of surrounding vehicles.

[0032] Further, step S21 includes:

[0033] Building an exposure metric Transformer model includes:

[0034] The input embedding module is used to take the feature sequence as input and transform the feature sequence into a vector representation of fixed dimensions;

[0035] The position encoding module is used to add position encoding to the embedding vector. After embedding, position encoding is added for each time step to obtain the embedding vector for each time step.

[0036] The multi-head attention module compares the embedding vector at each time step with the embedding vectors at all other time steps in the feature sequence and selects the embedding vector most relevant to the current time step for computation.

[0037] The feedforward neural network layer performs a non-linear transformation on the representation at each time step. The feedforward neural network consists of two fully connected layers, each using the ReLU activation function. The output of the feedforward neural network is calculated as follows:

[0038]

[0039] Where H is the output of the multi-head attention module, It is a weight matrix. It is a bias vector. Activation function;

[0040] The prediction output module takes the vector of the last time step, passes it through a fully connected layer, and then through a softmax function to output a vector of dimension 3. :

[0041] E vec =softmax(F)

[0042] In the formula, This is the output of the fully connected layer in a feedforward neural network.

[0043] The target scene fragment data extracted in step one is used as the input of the exposure metric Transformer model to train the model. The trained model is used to predict the driving intention of the tested vehicle, that is, the probability distribution of the three driving intentions of the tested vehicle under the influence of surrounding vehicles at the next time step: lane keeping, left lane change and right lane change. The probability value that has the greatest impact on the driving of the host vehicle in the probability distribution of the tested vehicle's driving intention is taken as the exposure value E of the driving risk brought by the tested vehicle to the host vehicle.

[0044] Furthermore, in step three, the probability that a person involved in the danger will suffer a maximum simple injury rating of 2 or higher when the danger occurs is taken as the severity S of the driving risk between the main vehicle and one of its neighboring vehicles. The calculation formula is as follows:

[0045]

[0046]

[0047] In the formula, The product of the risk coefficients of the two vehicles involved in the accident. and The risk coefficients for the two vehicles are respectively; For hyperparameters; The relative speed at the moment of collision between the two vehicles is calculated using the following formula:

[0048]

[0049]

[0050] In the formula, , , , , , These are the speed of the main vehicle, the acceleration of the main vehicle, the speed and acceleration of the circumferential vehicle, the relative distance between the two vehicles at the initial moment, and the time span from the initial moment to the collision between the two vehicles.

[0051] Furthermore, in step four, the magnitude of the acceleration change of the main vehicle is used to quantify the controllability of the risk, and the calculation formula is as follows:

[0052]

[0053] In the formula, C represents the controllability of the driving risk assessment between the main vehicle and one of its surrounding vehicles; The current acceleration of the main vehicle; The target acceleration for the main vehicle to avoid a collision; The maximum acceleration of the main vehicle; The maximum deceleration of the main vehicle.

[0054] The beneficial effects of this invention are as follows:

[0055] 1) More accurate risk assessment: It can provide a more objective and accurate risk assessment, rather than just being based on subjective judgment or qualitative analysis, which can help drivers better understand and quantify potential risks;

[0056] 2) More effective risk avoidance: By decoupling risk assessment, the contribution of different factors to the risk value can be obtained, thereby enabling targeted driving behaviors to more effectively reduce driving risks;

[0057] 3) Better decision-making: By using specific risk values, drivers can better compare the risk levels of different driving options, thus making better decisions. For example, when choosing between different routes or driving times, drivers can use the calculated risk values ​​as a reference to select the safest driving option.

[0058] 4) Higher traffic efficiency: Accurate risk assessment can help drivers plan their routes and speeds more effectively, avoiding unnecessary delays and waste. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below.

[0060] Figure 1 This is a schematic diagram of a traffic vehicle driving risk assessment method according to the present invention.

[0061] Figure 2 Example of a scenario for conducting a risk assessment of a vehicle.

[0062] Figure 3 This is a schematic diagram of the Transformer model structure used in the embodiments of the present invention. Detailed Implementation

[0063] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments:

[0064] like Figure 1 As shown, this embodiment is a method for assessing the driving risks of traffic vehicles, including the following steps:

[0065] Step 1: Extract the target scene from the original public dataset:

[0066] The principle for extracting target scene data is as follows:

[0067] First, the vehicles are classified into lane-keeping vehicles or lane-changing vehicles according to their driving status, and the classified vehicles are taken as the main vehicles. If the main vehicle is a lane-keeping vehicle, its feature scene data for any 2 seconds is extracted. If the main vehicle is a lane-changing vehicle, its feature scene data for the 2 seconds before the lane-changing start point is extracted. The lane-changing start point is set when the lateral distance between the lane-changing vehicle and the center line of the lane exceeds half the width of the vehicle. The lane-changing start point is defined as shown in formula (1):

[0068] (1)

[0069] In the formula, Indicates the lateral distance between the vehicle changing lanes and the center line of the lane. Indicates the width of the vehicle.

[0070] Based on the above-mentioned target scene data extraction principles, feature scene data is extracted, including the main vehicle feature sequence, the weekly vehicle feature sequence of each weekly vehicle, the main vehicle's left lane label, the main vehicle's right lane label, and the main vehicle's behavior label, thus forming scene fragment data with 25 feature sequences, which is the target scene data.

[0071] The main vehicle feature sequence includes four features: the main vehicle's horizontal coordinate, the main vehicle's vertical coordinate, the main vehicle's longitudinal speed, and the main vehicle's offset distance relative to the center line of its lane.

[0072] The feature sequence of each car includes three features: the horizontal coordinate, the vertical coordinate, and the longitudinal speed of that car. For the main car, there are 6 cars: left front, front, right front, left rear, rear, and right rear. Therefore, there are a total of 3*6=18 car feature sequence features.

[0073] The lane labels for the main vehicle's left and right lanes are added based on road structure information and the lane information of the main vehicle;

[0074] The main vehicle behavior label is used to characterize whether the main vehicle is classified as a lane-keeping vehicle or a lane-changing vehicle. The main vehicle behavior label needs to be obtained by prediction through the Transformer model.

[0075] The target scene data provides crucial data support for the training of the exposure metric Transformer model and the final driving risk assessment.

[0076] In this embodiment, the original public dataset includes all data from the first 63 subsets of the AD4CHE dataset.

[0077] Step 2: Based on the exposure quantification Transformer model, predict the driving intention of the surrounding vehicle, and then calculate the exposure degree E of the main vehicle's driving risk:

[0078] S21. Construct an exposure metric Transformer model and train the model:

[0079] See Figure 3 The Transformer model used in this embodiment includes:

[0080] 1. Input embedding module, which is used to take the feature sequence as input and transform the feature sequence into a vector representation of fixed dimension;

[0081] The Transformer model first uses formula (2) for input embedding. The input sequence in this embodiment is the 24 feature sequences of the test vehicle extracted in step one, excluding the main vehicle behavior label. The feature vector of each time step is mapped to the corresponding embedding vector.

[0082] (2)

[0083] In the formula, As the input sequence, in this embodiment , each sequence Mapped to a dimensional embedding vector Stack vectors into a matrix .

[0084] 2. A position encoding module is used to add position encoding to the embedding vector. After embedding, position encoding is added for each time step to obtain an embedding vector for each time step, which helps the model learn the order information in the sequence.

[0085] The location code calculation formulas are shown in (3) and (4):

[0086] (3)

[0087] (4)

[0088] In the formula, , These represent the encoding vectors for even-numbered and odd-numbered positions, respectively. Indicates the position of the time step. Indicates the index of the embedding vector. Indicates the dimension of the embedding vector.

[0089] The positional encoding is then added to the embedding vector to obtain the final embedding vector for each time step, as shown in formula (5):

[0090] (5)

[0091] In the formula, It is a positional encoding vector. It is an embedding vector that takes location information into account.

[0092] 3. Multi-head attention module: The multi-head attention module compares the embedding vector at each time step with the embedding vectors at all other time steps in the feature sequence, and selects the embedding vector most relevant to the current time step for calculation. The calculation process is as follows:

[0093] Use size is weight matrix The query vector is obtained by performing linear transformations on the embedding vectors respectively. Key vector Sum value vector As shown in formulas (6), (7), and (8):

[0094] (6)

[0095] (7)

[0096] (8)

[0097] In the formula, They are respectively the size of The weight matrix, It is the dimension of the embedded vector. These are the dimensions of each head;

[0098] (9)

[0099] (10)

[0100] In the formula, They represent the first The weight matrix of each attention head. In this embodiment, the number of attention heads is indicated. , This indicates a single-head attention mechanism. This means concatenating the outputs of all the headers. It is the output weight matrix of the multi-head attention mechanism;

[0101] It is a single-head attention mechanism, and the calculation formula is shown in formula (10):

[0102] (11)

[0103] In the formula, softmax represents the softmax function.

[0104] 4. Feedforward Neural Network Layer: The feedforward neural network performs a nonlinear transformation on the representation of each time step. The feedforward neural network consists of two fully connected layers, each using the ReLU activation function. The output of the feedforward neural network is calculated by formula (12):

[0105] (12)

[0106] Where H is the output of the multi-head attention module, It is a weight matrix. It is a bias vector. Activation function.

[0107] 5. Prediction Result Output Module:

[0108] In this embodiment, the input sequence length is 60 frames, with 24 features, 8 heads in the multi-head attention unit, and 4 stacked Transformer coding layers. The vector of the last time step is passed through a fully connected layer and then through a softmax function to output a vector of dimension 3, representing the probability distribution of the predicted vehicle driving intention in this embodiment, as shown in formula (13):

[0109] E vec =softmax(F) (13)

[0110] In the formula, This is the output of the fully connected layer.

[0111] The target scene fragment data extracted in step one is used as the input of the exposure metric Transformer model to train the model. After training, the model can predict the driving intention of the tested vehicle, that is, the probability distribution of the three driving intentions of the tested vehicle under the influence of surrounding vehicles at the next time step: lane keeping, left lane change and right lane change. Then, the exposure degree E of the driving risk brought by the tested vehicle to the host vehicle is calculated.

[0112] S22. Apply the trained exposure metric Transformer model to the target scene. Calculate the exposure degree E of the main vehicle's driving risk based on the predicted driving intentions of surrounding vehicles. The specific steps are as follows:

[0113] For the primary vehicle, its exposure to driving risks is primarily determined by the driving intentions of its surrounding vehicles. Therefore, the vehicles tested for predicting driving intentions should be the primary vehicle's surrounding vehicles. Figure 2 As shown, in this embodiment, car A is taken as the primary vehicle. To obtain the exposure value of car B's threat index to car A, the driving intention of car B should be predicted instead of car A's. Therefore, the target scene segment data when car B is the primary vehicle needs to be input into the exposure metric Transformer model to obtain the probabilities of car B's three driving intentions. The driving intention of car B that has the greatest impact on car A's driving is taken as the exposure value for risk assessment between A and B. It should be noted that the driving intentions of surrounding vehicles located at different positions from the primary vehicle are of different focuses. For example, for surrounding vehicles in front of the primary vehicle in the same lane, the probability of lane keeping should be considered; for surrounding vehicles in the lanes to the left and right of the primary vehicle, the probability of changing lanes to the primary vehicle's lane should be considered.

[0114] S23. By predicting the driving intentions of all surrounding vehicles of the main vehicle, the exposure of driving risks to the main vehicle from any surrounding vehicle can be obtained.

[0115] Step 3: Calculate the severity S of the primary vehicle driving risk, taking into account the vehicle type:

[0116] Different vehicle types can significantly impact the severity of accidents. This invention uses risk coefficients for different vehicle types to represent the impact on severity, as shown in the table below.

[0117] Types of traffic participants Risk coefficient car 1.000 the bus 1.363 truck 1.742 Motorcycle / Bicycle 1.641

[0118] The Simplified Injury Rating System is an anatomically based, universally accepted, and globally applicable injury severity rating system. It classifies each injury into six levels according to the degree of injury and by body region, with 0 indicating no injury and 6 indicating fatal injury.

[0119] The maximum simplified injury rating represents the highest simplified injury rating among multiple injuries suffered by the human body. A rating of 2 or higher represents the probability of suffering moderate to fatal injury.

[0120] The probability that a person involved in a hazard will suffer a maximum simple injury rating of 2 or higher when the hazard occurs is calculated by multiplying the original probability of a maximum simple injury rating of 2 or higher by the risk coefficients of the two vehicle types involved in the hazard, as shown in formulas (14) and (15):

[0121] (14)

[0122] (15)

[0123] (16)

[0124] (17)

[0125] In the formula, S is the probability that a person involved in the danger will suffer a maximum simple injury rating of 2 or higher when the danger occurs, considering the type of vehicle; that is, the severity of the driving risk between the main vehicle and one of its neighboring vehicles. The product of the risk coefficients of the two vehicles involved in the accident. and The risk coefficients for the two vehicles are respectively; These are hyperparameters, with values ​​of -6.068, 0.1, and -0.6234.

[0126] When the two vehicles do not collide That is, the severity is 0. The relative speed at the moment of collision between the two vehicles is calculated as shown in formulas (16) and (17), where, , , , , , These are the speed of the main vehicle, the acceleration of the main vehicle, the speed and acceleration of the circumferential vehicle, the relative distance between the two vehicles at the initial moment, and the time span from the initial moment to the collision between the two vehicles.

[0127] Step 4: Based on the magnitude of the change in the main vehicle's acceleration, calculate the controllability C of the driving risk between the main vehicle and any surrounding vehicles.

[0128] like Figure 2 As shown, when assessing the controllability of car A and car B, car B changing lanes to the right is most likely to cause a collision with car A. If car A accelerates appropriately so that it will not collide with car B even if car B changes lanes to the right, the controllability of the risk is quantified by the magnitude of the acceleration change of car A. The calculation method is shown in formula (18):

[0129] (18)

[0130] In the formula, C represents the controllability of the driving risk assessment between the main vehicle and one of its surrounding vehicles; This represents the current acceleration of car A. The target acceleration for car A to avoid a collision; This represents the maximum acceleration of car A. This represents the maximum deceleration of car A.

[0131] The risk controllability between the two vehicles is calculated based on formula (1). and The smaller the difference, the lower the risk, the higher the controllability, the lower the C value, and the lower the risk value in the final risk assessment. When the target acceleration is less than the maximum deceleration, a collision is inevitable, and the controllability is 1. However, when the current acceleration is less than the target acceleration, car A will not collide even if it does not change its driving state, and the controllability here is 0. As a result, the final risk value between the two cars is 0, meaning there will be no danger.

[0132] Step 5: Multiply the driving risk exposure E, controllability C, and severity S calculated in steps 2 to 4 to obtain the risk assessment result R between the two vehicles, as shown in formula (19):

[0133] (19)

[0134] In the formula, R represents the risk assessment result between the two vehicles.

[0135] Step Six: Take the maximum value of the risk assessment result R between the main vehicle and all surrounding vehicles as the driving risk assessment result of the main vehicle.

[0136] After calculating the risk assessment result R between the main vehicle and all its surrounding vehicles, the maximum value is taken to obtain the assessment result of the driving risk of the main vehicle at the current moment.

[0137] In summary, this invention addresses the issues of low universality and inaccurate quantification of vehicle driving risks by proposing an innovative method. This method quantifies the three risk assessment indicators using a novel or improved approach, aiming to obtain accurate and realistic risk levels among vehicles and ultimately achieve vehicle driving risk assessment results. This lays the foundation for proactive intervention and prevention of traffic accident risks.

Claims

1. A method for assessing the driving risks of traffic vehicles, characterized in that, Includes the following steps: Step 1: Extract target scene data from the original public dataset. The target scene data includes the main vehicle feature sequence, the weekly vehicle feature sequence of each weekly vehicle, the main vehicle's left lane label, the main vehicle's right lane label, and add the main vehicle behavior label. Step 2: Construct an exposure quantification Transformer model to predict the driving intentions of surrounding vehicles, and then calculate the exposure degree E of the main vehicle's driving risk; Step 2 includes: S21. Construct an exposure metric Transformer model and train the model; Building an exposure metric Transformer model includes: The input embedding module is used to take the feature sequence as input and transform the feature sequence into a vector representation of fixed dimensions; The position encoding module is used to add position encoding to the embedding vector. After embedding, position encoding is added for each time step to obtain the embedding vector for each time step. The multi-head attention module compares the embedding vector at each time step with the embedding vectors at all other time steps in the feature sequence and selects the embedding vector most relevant to the current time step for computation. The feedforward neural network layer performs a non-linear transformation on the representation at each time step. The feedforward neural network consists of two fully connected layers, each using the ReLU activation function. The output of the feedforward neural network is calculated as follows: ； Where H is the output of the multi-head attention module, It is a weight matrix. It is a bias vector. Activation function; The prediction output module takes the vector of the last time step, passes it through a fully connected layer, and then through a softmax function to output a vector of dimension 3. : E vec =softmax(F); In the formula, This is the output of the fully connected layer in a feedforward neural network. The target scene fragment data extracted in step one is used as the input of the exposure metric Transformer model to train the model. The trained model is used to predict the driving intention of the tested vehicle, that is, the probability distribution of the three driving intentions of the tested vehicle under the influence of surrounding vehicles at the next time step: lane keeping, left lane change and right lane change. The probability value that has the greatest impact on the driving of the host vehicle in the probability distribution of the tested vehicle's driving intention is taken as the exposure value E of the driving risk brought by the tested vehicle to the host vehicle. S22. Apply the trained exposure metric Transformer model to the target scene data, and calculate the exposure degree E of the driving risk of the main vehicle by predicting the driving intention of the surrounding vehicles: For the master vehicle, the vehicle being predicted is any of its neighboring vehicles B. The target scene data of the neighboring vehicle B when it is the master vehicle is input into the exposure quantification Transformer model to obtain the probabilities of the three driving intentions of the neighboring vehicle B. The driving intention probability value that has the greatest impact on the driving of the master vehicle B is taken as the exposure value of the driving risk brought by the neighboring vehicle B to the master vehicle. The probabilities of the three driving intentions refer to the probability distribution of the three driving intentions of lane keeping, left lane change and right lane change of the predicted vehicle at the next time step under the influence of its neighboring vehicles. S23. By predicting the driving intentions of all surrounding vehicles of the main vehicle, the exposure of driving risks brought to the main vehicle by any surrounding vehicle is obtained; Step 3: Calculate the severity S of the main vehicle driving risk, taking into account the vehicle type; Step 4: Based on the magnitude of the change in the acceleration of the main vehicle, calculate the controllability C of the driving risk between the main vehicle and any surrounding vehicle; Step 5: Multiply the driving risk exposure E, controllability C, and severity S calculated in Steps 2 to 4 to obtain the risk assessment result R between the two vehicles; Step 6: Take the maximum value of the risk assessment result R between the main vehicle and all surrounding vehicles as the driving risk assessment result of the main vehicle.

2. The method for assessing traffic vehicle driving risk as described in claim 1, characterized in that, In step one, the principle for extracting target scene data is as follows: First, based on the driving status of the vehicles, they are classified into lane-keeping vehicles or lane-changing vehicles, and the classified vehicles are taken as the main vehicles. If the main vehicle is a lane-keeping vehicle, extract its feature scene data for any 2 seconds; If the main vehicle is a lane-changing vehicle, then extract the feature scene data 2 seconds before its lane-changing start point. The lane-changing start point is set as the moment when the lateral distance between the lane-changing vehicle and the lane centerline exceeds half the vehicle width. The lane-changing start point is defined as follows: ； In the formula, Indicates the lateral distance between the vehicle changing lanes and the center line of the lane. Indicates the width of the vehicle.

3. The method for assessing traffic vehicle driving risk as described in claim 1, characterized in that, In step one, the target scene data specifically includes: The main vehicle feature sequence includes four features: the main vehicle's horizontal coordinate, the main vehicle's vertical coordinate, the main vehicle's longitudinal speed, and the main vehicle's offset distance relative to the center line of its lane. The feature sequence of each car includes three features: the horizontal coordinate, the vertical coordinate, and the longitudinal speed of the car. For the main car, there are 6 cars: left front, front, right front, left rear, rear, and right rear, for a total of 18 car feature sequence features. The left and right lane labels for the main vehicle are added based on road structure information and the lane information of the main vehicle. The main vehicle behavior label is used to characterize whether the main vehicle is classified as a lane-keeping vehicle or a lane-changing vehicle. The main vehicle behavior label is not used as input to the exposure metric Transformer model in step two.

4. The method for assessing traffic vehicle driving risk as described in claim 1, characterized in that, In step three, the probability that a person involved in the danger will suffer a maximum simple injury rating of 2 or higher when the danger occurs is taken as the severity S of the driving risk between the main vehicle and one of its neighboring vehicles. The calculation formula is as follows: ； ； In the formula, The product of the risk coefficients of the two vehicles involved in the accident. and The risk coefficients for the two vehicles are respectively; For hyperparameters; The relative speed at the moment of collision between the two vehicles is calculated using the following formula: ； ； In the formula, , , , , , These are the speed of the main vehicle, the acceleration of the main vehicle, the speed and acceleration of the circumferential vehicle, the relative distance between the two vehicles at the initial moment, and the time span from the initial moment to the collision between the two vehicles.

5. The method for assessing traffic vehicle driving risk as described in claim 1, characterized in that, In step four, the controllability of risk is quantified by the magnitude of the acceleration change of the main vehicle, and the calculation formula is as follows: ； In the formula, C represents the controllability of the driving risk assessment between the main vehicle and one of its surrounding vehicles; The current acceleration of the main vehicle; The target acceleration for the main vehicle to avoid a collision; The maximum acceleration of the main vehicle; The maximum deceleration of the main vehicle.

Images