A highway vehicle multi-modal trajectory prediction method based on multi-objective optimization

Abstract

The application particularly relates to a highway vehicle multi-modal trajectory prediction method based on multi-objective optimization, which comprises the following steps: establishing a vehicle trajectory prediction model by building an encoder and a decoder according to a multi-layer perception and an attention mechanism; collecting dynamic data and static data of a target vehicle and its adjacent vehicles whose trajectories are to be predicted, and performing pretreatment; performing model training on the vehicle trajectory prediction model based on a loss function to obtain a trained vehicle trajectory prediction model; collecting static data and dynamic data of the target vehicle and its adjacent vehicles at a plurality of time points before a prediction time point, and inputting the data into the trained vehicle trajectory prediction model to obtain a multi-modal prediction trajectory of the target vehicle. The application can more safely and reasonably predict the future trajectory of a highway vehicle in an intelligent network connection scene, can effectively predict potential collision risks, can reduce traffic accidents, can enhance the safety of road traffic, and can improve the traffic capacity of a road.

Description

Technical Field

[0001] This invention specifically relates to a multi-modal trajectory prediction method for highway vehicles based on multi-objective optimization, belonging to the field of highway vehicle trajectory prediction technology. Background Technology

[0002] With the rapid development of intelligent connected vehicle technology, a large number of intelligent connected vehicles have been deployed on highways. However, unreasonable vehicle trajectory prediction can lead to unsafe driving behaviors, thereby increasing driving risks. Therefore, safe and reasonable multimodal trajectory prediction for highway vehicles is one of the important research topics for intelligent connected vehicles, which can improve road capacity and driving safety.

[0003] Existing technologies have few multimodal design methods for predicting future vehicle trajectories, and even fewer studies that consider the safety and rationality of predicted trajectories. Most prediction methods focus on prediction accuracy and lack consideration for the safety and rationality of predicted trajectories. This makes it difficult for intelligent connected vehicles to make trajectory predictions that balance safety and fairness in real-world scenarios, thus increasing the collision risk that the predicted trajectories may cause. Summary of the Invention

[0004] The technical problem to be solved by this invention is: how to accurately predict the trajectory of vehicles on highways.

[0005] To solve the above-mentioned technical problems, the technical solution proposed in this invention is: a multi-objective optimization-based method for predicting multi-modal trajectories of vehicles on highways, comprising the following steps:

[0006] Step 1: Build an encoder and decoder based on the multilayer perceptron and attention mechanism to establish a vehicle trajectory prediction model;

[0007] Step 2: Collect n sets of dynamic and static data of the target vehicle a and its neighboring vehicles b within the historical time period T. Preprocess the n sets of dynamic and static data of the target vehicle a and its neighboring vehicles b respectively, and collect them as dataset S.

[0008] Step 3: Divide the dataset S into a training set, a validation set, and a test set according to a preset ratio and input them into the vehicle trajectory prediction model to train the model and obtain the trained vehicle trajectory prediction model. The trained vehicle trajectory prediction model takes the dynamic and static data of the target vehicle a and its neighboring vehicles b as inputs and the predicted trajectory of the target vehicle a as output.

[0009] Step 4: Collect static and dynamic data of the target vehicle a and the neighboring vehicle b within the j-th time intervals before the trajectory prediction time t, and input them into the trained vehicle trajectory prediction model to obtain the predicted trajectory of the target vehicle a at the trajectory prediction time t.

[0010] Furthermore, the dynamic data includes the vehicle's running coordinates. The static data includes the vehicle's length. and lane direction .

[0011] Furthermore, the specific content of the preprocessing is as follows:

[0012] If the vehicle is collected a certain number of times within the historical time period T The coordinate data is missing. It is supplemented based on the coordinate values ​​collected before and after it. The calculation formula is as follows:

[0013]

[0014] In the formula, It is the number of times a vehicle collects data within a certain historical time period T. Dynamic data, It is the number of times a vehicle collects data within a certain historical time period T. x-axis and y-axis coordinates; and These are the number of times a vehicle collects data within a specific historical time period T. The x-axis and y-axis coordinates of the previous data collection; and These are the number of times a vehicle collects data within a specific historical time period T. The x-axis and y-axis coordinates of the last data collection;

[0015] All dynamic data of vehicles after missing value imputation are normalized sequentially, where the number of times a vehicle is collected within a certain historical time period T is recorded. Dynamic data below The normalization process is as follows:

[0016]

[0017] In the formula, and These are the mean and variance of the x-coordinates of all vehicles within the historical time period T, respectively. and These are the mean and variance of the y-coordinates of all vehicles within the historical time period T, respectively.

[0018] Z-score normalization is applied to the vehicle length in all static data of the vehicle, and one-hot encoding is applied to the lane direction.

[0019] Furthermore, the dataset S is shown in the following formula:

[0020]

[0021] In the formula, It is the first historical target data group formed by the first set of dynamic data and static data collected by the target vehicle a within the historical time period T, wherein, These are the vehicle coordinates of the target vehicle a during the first data collection within the historical time period T. These are the x-axis coordinates and y-axis coordinates of the target vehicle a during the first data collection within the historical time period T; These are the lane length and lane direction of the target vehicle a during the first data collection within the historical time period T; It is the second historical target data group formed by the second set of dynamic data and static data collected by the target vehicle a within the historical time period T, wherein, These are the vehicle coordinates of the target vehicle a during the second data collection within the historical time period T. These are the x-axis coordinates and y-axis coordinates of the target vehicle a during the second data collection within the historical time period T; These are the lane length and lane direction of the target vehicle a during the second data collection within the historical time period T; It is the nth historical target data group formed by the nth set of dynamic data and static data collected by the target vehicle a within the historical time period T, wherein, These are the vehicle coordinates of the target vehicle a during the nth data collection within the historical time period T. These are the x-axis coordinates and y-axis coordinates of the target vehicle a during the nth data collection within the historical time period T, respectively. These are the lane length and lane direction of the target vehicle a during the nth data collection within the historical time period T; It is the first historical neighbor data group formed by the first set of dynamic data and static data collected from the neighboring vehicle b within the historical time period T, wherein These are the coordinates of the neighboring vehicle b during the first data collection within historical time period T. These are the x-axis coordinates and y-axis coordinates of the neighboring vehicle b during the first data collection within the historical time period T. These are the vehicle length and lane direction of the neighboring vehicle b during the first data collection within the historical time period T. It is the second historical proximity data group formed by the second set of dynamic data and static data collected by the neighboring vehicle b within the historical time period T, wherein These are the coordinates of the neighboring vehicle b, which were collected for the second time within historical time period T. These are the x-axis coordinates and y-axis coordinates of the neighboring vehicle b during the second data collection within the historical time period T; These are the vehicle length and lane direction of the neighboring vehicle b during the second data collection within the historical time period T. It is the nth historical neighbor data group formed by the nth set of dynamic data and static data collected by the neighboring vehicle b within the historical time period T, wherein These are the coordinates of the neighboring vehicle b during the nth data collection within the historical time period T. These are the x-axis coordinates and y-axis coordinates of the neighboring vehicle b during the nth data collection within the historical time period T, respectively. These are the vehicle length and lane direction of the neighboring vehicle b during the nth data collection within the historical time period T.

[0022] Furthermore, the model training process in step 3 is as follows:

[0023] Step 3.1: Initialize the model training iterations i, making the model training iterations i equal to 1;

[0024] Step 3.2: Randomly assign values ​​to the modal data of the target vehicle a and the neighboring vehicle b in K modalities to obtain the modal data of the target vehicle a and the neighboring vehicle b in K modalities when the model training number i equals 1, as shown in the following formula.

[0025]

[0026] In the formula, These are the modal data of the target vehicle a and the neighboring vehicle b in K different modes, respectively. These are the random assignments of modal data for the target vehicle a and the neighboring vehicle b under the K modalities when the model training number i equals 1;

[0027] The encoder in the vehicle trajectory prediction model extracts the final fusion features of K modalities under the condition that the model training number i equals 1;

[0028] Step 3.3: Input the final fused features of the K modalities under the model training number i equals 1 into the decoder of the vehicle trajectory prediction model in sequence to obtain the occurrence probability of each modality when the model training number i equals 1, and select the modality with the highest occurrence probability as the optimal modality. The optimal mode The final fused features are substituted into the decoder to generate the corresponding trajectory coordinates as the predicted trajectory of the target vehicle a when the model training number i equals 1.

[0029] Step 3.4: Obtain the true trajectory of the target vehicle a when the model training iteration i equals 1, and calculate the total loss when the model training iteration i equals 1 using the following formula. The vehicle trajectory prediction model is completed when the model training count i equals 1, and the model training count i is incremented by one.

[0030]

[0031] In the formula, the The predicted trajectory coordinates for the optimal mode ( , ) and the actual trajectory coordinates ( , The regression loss of the bias, L1(·), is the smooth L1 loss; The probability of the optimal mode occurring The classification follows the trend, and log(·) is the natural logarithm; the aforementioned To predict the risk loss of the trajectory, where, To minimize the total risk loss for all vehicles, To average the loss of all vehicles at risk, To minimize the loss of the maximum risk among all vehicles; the risk R is the probability of collision between the trajectories of target vehicle a and neighboring vehicle b. Severity of the collision The product; For the target vehicle With the adjacent vehicle The distance between them; It is the safety distance; exp(·) is an exponential function with the natural constant e as the base; sigmoid(·) is the activation function; It represents the risk coefficient for different road users, and is a constant value. Generally, the risk coefficient for vulnerable road users is lower than that for non-vulnerable road users. For the target vehicle With the adjacent vehicle The average risk between; max(·) is the maximum value.

[0032] Step 3.5: Update the modal data of the target vehicle a and the neighboring vehicle b under the K modalities when the model training number i equals 1 according to the following formula, to obtain the modal data of the target vehicle a and the neighboring vehicle b under the K modalities when the model training number i equals 2.

[0033]

[0034] In the formula, The learning rate;

[0035] Step 3.6: Repeat the principles of steps 3.3 to 3.5 to train the vehicle trajectory prediction model repeatedly until the number of training iterations i equals n-1, thus obtaining the trained vehicle trajectory prediction model.

[0036] Furthermore, the specific content of extracting the final fused features of the K modalities under the condition that the model training number i equals 1 through the encoder in step 3.2 is as follows:

[0037] Step 3.2.1: Calculate the dynamic feature vectors of the target vehicle a and the neighboring vehicle b at model training number i according to the following formula.

[0038]

[0039] In the formula, These are the dynamic feature vectors of the target vehicle a and the neighboring vehicle b at model training number i; The weights are for the first layer of the dynamic perceptron. This is used to bias the first-layer dynamic perceptron. For layer normalization, For activation function, For the weights of the second-layer dynamic perceptron, For the bias of the second-layer dynamic perceptron; The vehicle displacements of the target vehicle a and the neighboring vehicle b at model training number i;

[0040] Step 3.2.2: Calculate the static feature vectors of the target vehicle a and the neighboring vehicle b at model training number i according to the following formula.

[0041]

[0042] In the formula, These are the static feature vectors of the target vehicle a and the neighboring vehicle b at model training number i. The weights are those of the first-layer static perceptron. This is the bias for the first layer static perceptron. The weights are for the second-layer static perceptron. The second-layer static perceptron is biased;

[0043] Step 3.2.3: Calculate the initial fused dynamic features and initial fused static features of the target vehicle a and neighboring vehicle b under K modalities at model training number i according to the following formula.

[0044]

[0045] In the formula, It is the attention mechanism in the encoder; It is the multilayer sensor in the encoder;

[0046] The initial fused dynamic features and initial fused static features of the target vehicle a and the neighboring vehicle b under K modalities at model training number i are concatenated to obtain the initial fused features under K modalities at model training number i, as shown below.

[0047] ;

[0048] Step 3.2.4: Obtain the final fused features of the target vehicle a in K modalities at model training number i according to the following formula.

[0049]

[0050] In the formula, and These are the initial fusion features of the target vehicle a and the neighboring vehicle b in the k-th modality at the i-th training iteration of the model.

[0051] Furthermore, in step 3.3, the final fused features of the K modalities under the model training number i equals 1 are sequentially input into the decoder in the vehicle trajectory prediction model to obtain the occurrence probability of the target vehicle a in each modality when the model training number i equals 1. The specific decoding formula is as follows:

[0052]

[0053] In the formula, It is the probability of the target vehicle a occurring in the k-th mode when the model training number i equals 1; It is an exponential function with the natural constant e as its base.

[0054] Furthermore, in step 3.3, the optimal mode is... The final fused features are then substituted into the decoder to generate the corresponding trajectory coordinates. The specific decoding formula is as follows:

[0055]

[0056] In the formula, and These are the optimal modes The final fused features are then substituted into the decoder to generate the predicted x-axis and y-axis coordinates of the target vehicle a.

[0057] The beneficial effects of this invention are: it provides a multi-objective optimization-based method for predicting the multimodal trajectory of highway vehicles. By utilizing a multilayer perceptron and attention mechanism, it constructs a multi-objective optimization loss function that considers minimizing overall risk, distributing risk equally, and minimizing the risk of maximum damage. This method can safely and reasonably predict the multimodal trajectory of highway vehicles, providing technical support for highway vehicle trajectory decision-making and planning, and enhancing vehicle driving safety. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating the prediction method in an embodiment of the present invention.

[0059] Figure 2 This is a schematic diagram of the encoder extracting the final fused features according to an embodiment of the present invention.

[0060] Figure 3 This is a schematic diagram of the decoder decoding final fusion feature process according to an embodiment of the present invention. Detailed Implementation

[0061] The following description, in conjunction with the accompanying drawings and specific embodiments, further illustrates the multi-objective optimization-based multimodal trajectory prediction method for highway vehicles according to the present invention.

[0062] Example

[0063] The prediction method in this embodiment, such as Figure 1 As shown, it includes the following steps:

[0064] A method for predicting multimodal trajectories of vehicles on highways based on multi-objective optimization includes the following steps:

[0065] Step 1: Build an encoder and decoder based on multi-layer perception and attention mechanisms to establish a vehicle trajectory prediction model. Details are as follows:

[0066] The preprocessed static and dynamic data are input into the encoder to obtain fused features, which are then decoded by the decoder to obtain the multimodal predicted trajectory and its corresponding probability. The encoder is based on an attention mechanism and a multilayer perceptron, while the decoder is also based on a multilayer perceptron.

[0067] Step 2: Collect n sets of dynamic and static data for the target vehicle a and its neighboring vehicles b within the historical time period T. The dynamic data includes the vehicle's running coordinates. Static data includes the vehicle's length within the vehicle's orbit. and lane direction .

[0068] The n sets of dynamic and static data of target vehicle a and neighboring vehicle b are preprocessed separately. The specific preprocessing content is as follows:

[0069] If the vehicle is collected a certain number of times within the historical time period T. The coordinate data is missing. It is supplemented based on the coordinate values ​​collected before and after it. The calculation formula is as follows:

[0070]

[0071] In the formula, It is the number of times a vehicle collects data within a certain historical time period T. Dynamic data, It is the number of times a vehicle collects data within a certain historical time period T. x-axis and y-axis coordinates; and These are the number of times a vehicle collects data within a specific historical time period T. The x-axis and y-axis coordinates of the previous data collection; and These are the number of times a vehicle collects data within a specific historical time period T. The x-axis and y-axis coordinates of the last data collection;

[0072] All dynamic data of vehicles after missing value imputation are normalized sequentially, where the number of times a vehicle is collected within a certain historical time period T is recorded. Dynamic data below The normalization process is as follows:

[0073]

[0074] In the formula, and These are the mean and variance of the x-coordinates of all vehicles within the historical time period T, respectively. and These are the mean and variance of the y-coordinates of all vehicles within the historical time period T, respectively.

[0075] Z-score normalization is applied to the vehicle length in all static data of the vehicle, and one-hot encoding is applied to the lane direction.

[0076] The preprocessed dynamic and static data of target vehicle a and neighboring vehicle b are then divided into n data groups and collected as dataset S; dataset S is shown in the following formula.

[0077]

[0078] In the formula, This is the first historical target data set, formed by the first set of dynamic and static data collected from target vehicle a within historical time period T. These are the coordinates of target vehicle a during the first data collection within historical time period T. These are the x-axis coordinates and y-axis coordinates of target vehicle a during the first data collection within historical time period T; These are the lane length and lane direction of target vehicle a when data is first collected within historical time period T. This is the second set of historical target data, formed by the second set of dynamic and static data collected from target vehicle a within historical time period T. These are the coordinates of target vehicle a during the second data collection within historical time period T. These are the x-axis coordinates and y-axis coordinates of target vehicle a during the second data collection within historical time period T; These are the lane length and lane direction of target vehicle a during the second data collection within historical time period T; It is the nth historical target data set formed by the nth set of dynamic and static data collected from target vehicle a within historical time period T, where, These are the coordinates of target vehicle a during the nth data collection within historical time period T. These are the x-axis coordinates and y-axis coordinates of target vehicle a during the nth data collection within historical time period T, respectively. These are the lane length and lane direction of target vehicle a during the nth data collection within historical time period T; This is the first historical neighbor data set, formed by the first set of dynamic and static data collected from neighboring vehicle b within the historical time period T. These are the coordinates of the neighboring vehicle b, which was first collected within historical time period T. These are the x-axis coordinates and y-axis coordinates of the neighboring vehicle b during the first data collection within the historical time period T. These are the vehicle length and lane direction of the neighboring vehicle b, which is the first vehicle to have its data collected within the historical time period T. This is the second historical neighbor data set, formed by the second set of dynamic and static data collected from neighboring vehicle b within the historical time period T. These are the coordinates of vehicle b, which was the second vehicle to have its data collected within historical time period T. These are the x-axis coordinates and y-axis coordinates of the neighboring vehicle b during the second data collection within the historical time period T; These are the vehicle length and lane direction of the neighboring vehicle b during the second data collection within the historical time period T; It is the nth historical neighbor data set formed by the nth set of dynamic and static data collected from neighboring vehicle b within the historical time period T, where These are the coordinates of the neighboring vehicle b during the nth data collection within historical time period T. These are the x-axis coordinates and y-axis coordinates of the neighboring vehicle b during the nth data collection within the historical time period T, respectively. These are the vehicle length and lane direction of the neighboring vehicle b during the nth data collection within the historical time period T.

[0079] Step 3: Divide the dataset S into training, validation, and test sets according to a preset ratio, and input them into the vehicle trajectory prediction model for model training.

[0080] The model training process is as follows:

[0081] Step 3.1: Initialize the number of training iterations i of the model so that the number of training iterations i equals 1;

[0082] Step 3.2: Randomly assign values ​​to the modal data of target vehicle a and neighboring vehicle b in K modalities to obtain the modal data of target vehicle a and neighboring vehicle b in K modalities when the model training number i equals 1, as shown in the following formula.

[0083]

[0084] In the formula, These are the modal data of target vehicle a and neighboring vehicle b in K different modes; These are the random assignments of modal data for target vehicle a and neighboring vehicle b under K modalities when the model training number i equals 1;

[0085] The final fused features for K modalities under model training number i equal to 1 are extracted from the encoder in the vehicle trajectory prediction model, such as... Figure 2 As shown, the specific content is as follows:

[0086] Step 3.2.1: Calculate the dynamic feature vectors of target vehicle a and neighboring vehicle b at model training number i according to the following formula.

[0087]

[0088] In the formula, It is the dynamic feature vector of target vehicle a and neighboring vehicle b under model training number i; The weights are for the first layer of the dynamic perceptron. This is used to bias the first-layer dynamic perceptron. For layer normalization, For activation function, For the weights of the second-layer dynamic perceptron, For the bias of the second-layer dynamic perceptron; The vehicle displacements of target vehicle a and neighboring vehicle b at model training iteration i;

[0089] Step 3.2.2: Calculate the static feature vectors of target vehicle a and neighboring vehicle b at model training number i according to the following formula.

[0090]

[0091] In the formula, These are the static feature vectors of target vehicle a and neighboring vehicle b at model training number i. The weights are those of the first-layer static perceptron. This is the bias for the first layer static perceptron. The weights are for the second-layer static perceptron. The second-layer static perceptron is biased;

[0092] Step 3.2.3: Calculate the initial fused dynamic features and initial fused static features of target vehicle a and neighboring vehicle b under K modalities at model training number i using the following formula.

[0093]

[0094] In the formula, It is the attention mechanism in the encoder; It is a multilayer perceptron in the encoder;

[0095] The initial fused dynamic features and initial fused static features of target vehicle a and neighboring vehicle b under K modalities at model training number i are concatenated to obtain the initial fused features under K modalities at model training number i, as shown below.

[0096] ;

[0097] Step 3.2.4: Obtain the final fused features of target vehicle a across K modalities in model training iteration i using the following formula.

[0098]

[0099] In the formula, and These are the initial fused features of target vehicle a and neighboring vehicle b in the k-th modality at model training number i.

[0100] Step 3.3: Input the final fused features of the K modalities under model training number i equals 1 into the decoder of the vehicle trajectory prediction model in sequence to obtain the occurrence probability of each modality when model training number i equals 1. The specific decoding formula is as follows:

[0101]

[0102] In the formula, It is the probability of the target vehicle a occurring in the k-th mode when the model training number i equals 1; It is an exponential function with the natural constant e as its base.

[0103] The mode with the highest probability of occurrence is selected as the optimal mode. , the best mode The final fused features are then fed into the decoder to generate the corresponding trajectory coordinates, which serve as the predicted trajectory of target vehicle a when the model training iteration i equals 1. Figure 3 As shown, the specific decoding formula is as follows:

[0104]

[0105] In the formula, and These are the optimal modes The final fused features are then fed into the decoder to generate the predicted x-axis and y-axis coordinates of the target vehicle a.

[0106] Step 3.4: Obtain the true trajectory of target vehicle a when the model training iteration i equals 1, and calculate the total loss when the model training iteration i equals 1 using the following formula. Complete the vehicle trajectory prediction model training process when the model training count i equals 1, and increment the model training count i by one.

[0107]

[0108] In the formula, the The predicted trajectory coordinates for the optimal mode ( , ) and the actual trajectory coordinates ( , The regression loss of the bias, L1(·), is the smooth L1 loss; The probability of the optimal mode occurring The classification follows the trend, and log(·) is the natural logarithm; the aforementioned To predict the risk loss of the trajectory, where, To minimize the total risk loss for all vehicles, To average the loss of all vehicles at risk, To minimize the loss of the maximum risk among all vehicles; the risk R is the probability of collision between the trajectories of target vehicle a and neighboring vehicle b. Severity of the collision The product; For the target vehicle With the adjacent vehicle The distance between them; It is the safety distance; exp(·) is an exponential function with the natural constant e as the base; sigmoid(·) is the activation function; It represents the risk coefficient for different road users, and is a constant value. Generally, the risk coefficient for vulnerable road users is lower than that for non-vulnerable road users. For the target vehicle With the adjacent vehicle The average risk between; max(·) is the maximum value.

[0109]

[0110] In the formula, This is the learning rate.

[0111] Step 3.5: Update the modal data of target vehicle a and neighboring vehicle b under the K modalities when the model training number i equals 1 according to the following formula, to obtain the modal data of target vehicle a and neighboring vehicle b under the K modalities when the model training number i equals 2.

[0112] Step 3.6: Repeat the principles of steps 3.3 to 3.5 to train the vehicle trajectory prediction model repeatedly until the number of training iterations i equals n-1, thus obtaining the trained vehicle trajectory prediction model.

[0113] The trained vehicle trajectory prediction model takes the dynamic and static data of the target vehicle a and its neighboring vehicles b as inputs and the predicted trajectory of the target vehicle a as output.

[0114] Step 4: Collect static and dynamic data of the target vehicle a and the neighboring vehicle b within the j-th time before the trajectory prediction time t, and input them into the trained vehicle trajectory prediction model to obtain the predicted trajectory of the target vehicle a at the trajectory prediction time t.

[0115] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for predicting multimodal trajectories of vehicles on highways based on multi-objective optimization, characterized in that: Includes the following steps: Step 1: Build an encoder and decoder based on the multilayer perceptron and attention mechanism to establish a vehicle trajectory prediction model; Step 2: Collect n sets of dynamic and static data of the target vehicle a and its neighboring vehicles b within the historical time period T. Preprocess the n sets of dynamic and static data of the target vehicle a and its neighboring vehicles b respectively, and collect them as dataset S. Step 3: Divide the dataset S into a training set, a validation set, and a test set according to a preset ratio and input them into the vehicle trajectory prediction model to train the model and obtain the trained vehicle trajectory prediction model. The trained vehicle trajectory prediction model takes the dynamic and static data of the target vehicle a and its neighboring vehicles b as inputs and the predicted trajectory of the target vehicle a as output. Step 4: Collect static and dynamic data of the target vehicle a and the neighboring vehicle b within the j moments before the trajectory prediction time t, and input them into the trained vehicle trajectory prediction model to obtain the predicted trajectory of the target vehicle a at the trajectory prediction time t. Furthermore, the dynamic data includes the vehicle's running coordinates. The static data includes the vehicle's length. and lane direction .

2. The method for predicting multimodal trajectories of highway vehicles based on multi-objective optimization according to claim 1, characterized in that: The specific details of the preprocessing are as follows: If the vehicle is collected a certain number of times within the historical time period T The coordinate data is missing. It is supplemented based on the coordinate values ​​collected before and after it. The calculation formula is as follows: ； In the formula, It is the number of times a vehicle collects data within a certain historical time period T. Dynamic data, It is the number of times a vehicle collects data within a certain historical time period T. x-axis and y-axis coordinates; and These are the number of times a vehicle collects data within a specific historical time period T. The x-axis and y-axis coordinates of the previous data collection; and These are the number of times a vehicle collects data within a specific historical time period T. The last collected x-axis and y-axis coordinate values; All dynamic data of vehicles after missing value imputation are normalized sequentially, where the number of times a vehicle is collected within a certain historical time period T is recorded. Dynamic data below The normalization process is as follows: ； In the formula, and These are the mean and variance of the x-coordinates of all vehicles within the historical time period T, respectively. and These are the mean and variance of the y-coordinates of all vehicles within the historical time period T. The vehicle length in all static data is also normalized using Z-score, and the lane direction is processed using one-hot encoding.

3. The method for predicting multimodal trajectories of highway vehicles based on multi-objective optimization according to claim 1, characterized in that: The dataset S is shown in the following formula. ； In the formula, It is the first historical target data group formed by the first set of dynamic data and static data collected by the target vehicle a within the historical time period T, wherein, These are the vehicle coordinates of the target vehicle a during the first data collection within the historical time period T. These are the x-axis coordinates and y-axis coordinates of the target vehicle a during the first data collection within the historical time period T; These are the lane length and lane direction of the target vehicle a during the first data collection within the historical time period T; It is the second historical target data group formed by the second set of dynamic data and static data collected by the target vehicle a within the historical time period T, wherein, These are the vehicle coordinates of the target vehicle a during the second data collection within the historical time period T. These are the x-axis coordinates and y-axis coordinates of the target vehicle a during the second data collection within the historical time period T; These are the lane length and lane direction of the target vehicle a during the second data collection within the historical time period T; It is the nth historical target data group formed by the nth set of dynamic data and static data collected by the target vehicle a within the historical time period T, wherein, These are the vehicle coordinates of the target vehicle a during the nth data collection within the historical time period T. These are the x-axis coordinates and y-axis coordinates of the target vehicle a during the nth data collection within the historical time period T, respectively. These are the lane length and lane direction of the target vehicle a during the nth data collection within the historical time period T; It is the first historical neighbor data group formed by the first set of dynamic data and static data collected from the neighboring vehicle b within the historical time period T, wherein These are the coordinates of the neighboring vehicle b during the first data collection within historical time period T. These are the x-axis coordinates and y-axis coordinates of the neighboring vehicle b during the first data collection within the historical time period T. These are the vehicle length and lane direction of the neighboring vehicle b during the first data collection within the historical time period T. It is the second historical proximity data group formed by the second set of dynamic data and static data collected by the neighboring vehicle b within the historical time period T, wherein These are the coordinates of the neighboring vehicle b, which were collected for the second time within historical time period T. These are the x-axis coordinates and y-axis coordinates of the neighboring vehicle b during the second data collection within the historical time period T; These are the vehicle length and lane direction of the neighboring vehicle b during the second data collection within the historical time period T. It is the nth historical neighbor data group formed by the nth set of dynamic data and static data collected by the neighboring vehicle b within the historical time period T, wherein These are the coordinates of the neighboring vehicle b during the nth data collection within the historical time period T. These are the x-axis coordinates and y-axis coordinates of the neighboring vehicle b during the nth data collection within the historical time period T, respectively. These are the vehicle length and lane direction of the neighboring vehicle b during the nth data collection within the historical time period T.

4. The method for predicting multimodal trajectories of highway vehicles based on multi-objective optimization according to claim 1, characterized in that: The model training process in step 3 is as follows: Step 3.1: Initialize the model training iterations i, making the model training iterations i equal to 1; Step 3.2: Randomly assign values ​​to the modal data of the target vehicle a and the neighboring vehicle b in K modalities to obtain the modal data of the target vehicle a and the neighboring vehicle b in K modalities when the model training number i equals 1, as shown in the following formula. ； In the formula, These are the modal data of the target vehicle a and the neighboring vehicle b in K different modes, respectively. These are the random assignments of modal data for the target vehicle a and the neighboring vehicle b under the K modalities when the model training number i equals 1; The encoder in the vehicle trajectory prediction model extracts the final fusion features of K modalities under the condition that the model training number i equals 1; Step 3.3: Input the final fused features of the K modalities under the model training number i equals 1 into the decoder of the vehicle trajectory prediction model in sequence to obtain the occurrence probability of each modality when the model training number i equals 1, and select the modality with the highest occurrence probability as the optimal modality. The optimal mode The final fused features are substituted into the decoder to generate the corresponding trajectory coordinates as the predicted trajectory of the target vehicle a when the model training number i equals 1. Step 3.4: Obtain the true trajectory of the target vehicle a when the model training iteration i equals 1, and calculate the total loss when the model training iteration i equals 1 using the following formula. The vehicle trajectory prediction model is completed when the model training count i equals 1, and the model training count i is incremented by one. ； In the formula, the The predicted trajectory coordinates for the optimal mode ( , ) and the actual trajectory coordinates ( , The regression loss of the bias, L1(·), is the smooth L1 loss; The probability of the optimal mode occurring The classification follows the trend, and log(·) is the natural logarithm; the aforementioned To predict the risk loss of the trajectory, where, To minimize the total risk loss for all vehicles, To average the loss of all vehicles at risk, To minimize the loss of the maximum risk among all vehicles; the risk R is the probability of collision between the trajectories of target vehicle a and neighboring vehicle b. Severity of the collision The product; For the target vehicle With the adjacent vehicle The distance between them; It is the safety distance; exp(·) is an exponential function with the natural constant e as the base; sigmoid(·) is the activation function; It represents the risk coefficient for different road users, and is a constant value. Generally, the risk coefficient for vulnerable road users is lower than that for non-vulnerable road users. For the target vehicle With the adjacent vehicle The average risk between; max(·) is the maximum value; Step 3.5: Update the modal data of the target vehicle a and the neighboring vehicle b under the K modalities when the model training number i equals 1 according to the following formula, to obtain the modal data of the target vehicle a and the neighboring vehicle b under the K modalities when the model training number i equals 2. ； In the formula, The learning rate; Step 3.6: Repeat the principles of steps 3.3 to 3.5 to train the vehicle trajectory prediction model repeatedly until the number of training iterations i equals n-1, thus obtaining the trained vehicle trajectory prediction model.

5. The method for predicting multimodal trajectories of highway vehicles based on multi-objective optimization according to claim 4, characterized in that: The specific content of extracting the final fused features of K modalities under the condition that the model training number i is equal to 1 through the encoder in the vehicle trajectory prediction model is as follows: Step 3.2.1: Calculate the dynamic feature vectors of the target vehicle a and the neighboring vehicle b at model training number i according to the following formula. ； In the formula, These are the dynamic feature vectors of the target vehicle a and the neighboring vehicle b at model training number i; The weights are for the first layer of the dynamic perceptron. This is used to bias the first-layer dynamic perceptron. For layer normalization, For activation function, For the weights of the second-layer dynamic perceptron, For the bias of the second-layer dynamic perceptron; The vehicle displacements of the target vehicle a and the neighboring vehicle b at model training number i; Step 3.2.2: Calculate the static feature vectors of the target vehicle a and the neighboring vehicle b at model training number i according to the following formula. ； In the formula, These are the static feature vectors of the target vehicle a and the neighboring vehicle b at model training number i. The weights are those of the first-layer static perceptron. This is used to bias the first-layer static perceptron. For the weights of the second-layer static perceptron, The bias is applied to the second-layer static perceptron; Step 3.2.3: Calculate the initial fused dynamic features and initial fused static features of the target vehicle a and the neighboring vehicle b under K modalities at model training number i according to the following formula. ； In the formula, It is the attention mechanism in the encoder; It is the multilayer sensor in the encoder; The initial fused dynamic features and initial fused static features of the target vehicle a and the neighboring vehicle b under K modalities at model training number i are concatenated to obtain the initial fused features under K modalities at model training number i, as shown below. ；； Step 3.2.4: Obtain the final fused features of the target vehicle a in K modalities at model training number i according to the following formula. ； In the formula, and These are the initial fusion features of the target vehicle a and the neighboring vehicle b in the k-th modality at the i-th training iteration of the model.

6. The method for predicting multimodal trajectories of highway vehicles based on multi-objective optimization according to claim 5, characterized in that: In step 3.3, the final fused features of the K modalities under the model training number i equals 1 are sequentially input into the decoder of the vehicle trajectory prediction model to obtain the occurrence probability of the target vehicle a in each modality when the model training number i equals 1. The specific decoding formula is as follows: ； In the formula, It is the probability of the target vehicle a occurring in the k-th mode when the model training number i equals 1; It is an exponential function with the natural constant e as its base.

7. The method for predicting multimodal trajectories of highway vehicles based on multi-objective optimization according to claim 6, characterized in that: In step 3.3, the optimal mode will be... The final fused features are then substituted into the decoder to generate the corresponding trajectory coordinates. The specific decoding formula is as follows: ； In the formula, and These are the optimal modes The final fused features are then substituted into the decoder to generate the predicted x-axis and y-axis coordinates of the target vehicle a.

Images