Vehicle track prediction method and system oriented to complex traffic scene

Abstract

The invention discloses a vehicle track prediction method and system for a complex traffic scene, and the method comprises the following steps: S1, obtaining the historical track information of a target vehicle and surrounding vehicles in the traffic scene, and extracting the track time sequence characteristics of a prediction target and the historical track information of the surrounding vehicles through a hierarchical graph neural network; s2, acquiring environment information in a traffic scene where the target vehicle is located, and extracting environment multi-scale features in global and local environment information around the target vehicle by using a hierarchical graph neural network and an attention feature fusion module; s3, performing interactive fusion on the vehicle track time sequence features and the environment features by using multiple attention blocks to obtain interactive features between the scene and the vehicle; and S4, decoding the fusion features of the scene and the vehicle, and outputting a multi-modal future trajectory of the prediction target. The method can ensure that the model can adapt to a complex and changeable traffic environment while the prediction accuracy of the model is improved and the complexity is not increased.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a vehicle trajectory prediction method and system for complex traffic scenarios. Background Art

[0002] In recent years, driverless technology has become a hot topic in the scientific research community. In the entire process of driverless driving, trajectory prediction is the core link connecting perception and decision-making planning. Its accuracy is crucial to whether the autonomous driving system can plan a safe and reliable driving route.

[0003] Trajectory prediction is to model the vehicle's historical motion trajectory and surrounding environment information to predict the vehicle's possible future driving trajectory. Traditional trajectory prediction methods mainly rely on physical models and machine learning algorithms, but because the actual traffic environment contains a large number of dynamic factors and complex interactions, traditional methods often find it difficult to fully capture and accurately reflect these complex situations. With the rapid development of deep learning technology, trajectory prediction methods based on deep learning are gradually becoming dominant. Compared with traditional methods, deep learning models have stronger adaptability and higher prediction accuracy, especially when facing complex traffic scenarios.

[0004] However, the real traffic environment is dynamic and when the traffic conditions become more complex, such as when the number of vehicles on the road increases or the traffic environment becomes complicated, the existing prediction models may face the problem of performance degradation. In order for driverless cars to operate in real traffic environments, they not only need to be able to cope with various unknown environmental changes, but also need to respond in a timely manner.

[0005] Therefore, the art needs a new vehicle trajectory prediction solution to solve the above problems. Summary of the invention

[0006] The purpose of the present invention is to provide a vehicle trajectory prediction method and system for complex traffic scenarios to solve the problems raised in the above background technology.

[0007] To achieve the above-mentioned object of the invention, one aspect of the present invention provides a vehicle trajectory prediction method for complex traffic scenarios, comprising the following steps:

[0008] Step S1, obtaining historical trajectory information of the target vehicle and surrounding vehicles in the traffic scene, and using a hierarchical graph neural network to extract trajectory time series features in the historical trajectory information of the predicted target and surrounding vehicles;

[0009] Step S2, obtaining environmental information in the traffic scene where the target vehicle is located, and using a hierarchical graph neural network and an attention feature fusion module to extract multi-scale environmental features in the global and local environmental information around the target vehicle;

[0010] Step S3, using a multi-head attention block to interactively fuse the vehicle trajectory temporal features and environmental features to obtain the interactive features between the scene and the vehicle;

[0011] Step S4, decoding the fusion features of the scene and the vehicle to output a multi-modal future trajectory of the predicted target.

[0012] Furthermore, the historical trajectory information of the vehicle and the environmental information in steps S1 and S2 are obtained based on the Argoverse motion prediction dataset, and the obtained data are preprocessed.

[0013] Furthermore, the specific implementation process of extracting the trajectory time series features from the historical trajectory information of the predicted target and surrounding vehicles using the hierarchical graph neural network in step S1 is as follows:

[0014] The hierarchical graph neural network uses vectorization to create a connected graph for each scene element independently; the polyline is obtained through the vectorization process, and each vector f of the polyline is i is regarded as a node in the graph, i.e., a feature vector with a polyline length of d, f i ∈R d , expressed as the following formula:

[0015]

[0016] where f i (l) Represents the feature vector of the i-th node in the l-th layer; MLP is a multi-layer perceptron, which is used to perform nonlinear transformation on the feature vector. Here L refers to a specific layer configuration. The MLP in the model is configured as 2 linear layers and a ReLU activation function for nonlinearity; pool performs a max-pool operation to obtain information from adjacent nodes; cat performs a concatenation operation.

[0017] Furthermore, the implementation process of extracting environmental features using the hierarchical graph neural network and the attention feature fusion module in step S2 includes the following steps:

[0018] Step S201, using a layered graph neural network to create a connected graph using vector representation to obtain a feature vector;

[0019] Step S202, using the Attentional Feature Fusion (AFF) module to perform feature fusion on local and global feature vectors. Before fusing features of different scales, the importance of each feature is evaluated first, and then the features are weighted and fused according to the importance, which is expressed as the following formula:

[0020]

[0021] where f i (l) Represents the feature vector of the i-th node in the l-th layer; MLP is a multi-layer perceptron, which is used to perform nonlinear transformation on the feature vector. Here L refers to a specific layer configuration. The MLP in the model is configured as 2 linear layers and a ReLU activation function for nonlinearity; pool performs a max-pool operation to obtain information from adjacent nodes; cat performs a concatenation operation; AFF is an attention feature fusion module, which is used to fuse feature vectors of different scales.

[0022] Furthermore, the feature fusion includes the following steps:

[0023] Step S221, compressing the features into a vector of a fixed size through global average pooling to reduce computational complexity;

[0024] Step S222, the compressed feature vector is fused with multi-scale information through point-by-point convolution and nonlinear activation function;

[0025] Step S223, through the Sigmoid activation function, the output of the fully connected layer is mapped to the interval (0,1), the importance weight of each feature is obtained, the original features are weighted using the calculated weights, and then the weighted features are multiplied to form the final fusion feature.

[0026] Furthermore, the specific implementation process of step S3 is as follows:

[0027] In step S301, different types of interactions are modeled, including vehicle-to-lane (AL), lane-to-lane (LL), lane-to-vehicle (LA), and vehicle-to-vehicle (AA) interactions, and different attention mechanisms are used to extract different interactions.

[0028] Step S302, using the self-attention encoder and the feed-forward neural network (FFN) to update the self-relation (AA, LL), and using the cross-attention encoder and the feed-forward neural network FNN to update the cross-relation (AL, LA):

[0029] Step S303, model each interaction in sequence and repeat the process L times, updating the intermediate features at each iteration, and then using the updated features to calculate attention in the next iteration.

[0030] Furthermore, in step S3, the multi-head attention mechanism is defined as follows:

[0031]

[0032] where Q,K,V=W q f q ,W k f kv ,W v f kv

[0033] Where Q, K, and V represent query, key, and value matrices, respectively; W q , W k , W v represents the learnable linear transformation of the attention head; K T Represents the transpose of vector K; dim k Represents the dimension size of the K vector; softmax is a nonlinear activation function used to normalize the attention weight to ensure that the output is a probability distribution;

[0034] Then the multi-head attention block definition formula is as follows:

[0035]

[0036] where f q represents the query input feature vector, f kv represents the key and value input feature vector; where is the updated feature vector, which is obtained by adding the result of the multi-head attention operation to the original input feature vector and then normalizing it; norm indicates that a normalization operation is performed to stabilize the training process and accelerate convergence;

[0037] FFN is a feed-forward neural network, defined as follows:

[0038] FFN(a)=ReLU(aW1+b1)W2+b2

[0039] Among them, W1 and W2 represent learnable weight matrices; b1 and b2 represent bias vectors of the feedforward network; a represents the input of the feedforward network; ReLU represents the activation function, which is used to introduce nonlinearity so that the model can fit more complex functions.

[0040] Furthermore, the implementation process of step S4 is as follows:

[0041] The fused features are decoded using a 2-layer MLP (Multilayer Perceptron) with ReLU activation function, layer normalization is used after the first layer, and residual connection is used in the last layer to obtain the multimodal future trajectory of the predicted target.

[0042] Another aspect of the present invention provides a vehicle trajectory prediction system for complex traffic scenarios, including a trajectory encoding module, an environment encoding module, an interaction module and a decoding module, wherein:

[0043] The trajectory encoding module uses vectors to represent the trajectory of the predicted target and surrounding intelligent agents, and uses a hierarchical graph neural network to extract the trajectory time series features from the historical trajectory information of the predicted target and surrounding vehicles;

[0044] The environment encoding module uses vector representation to predict the trajectory of the target and surrounding agents, and uses a hierarchical graph neural network and an attention feature fusion module to extract multi-scale features of the target vehicle's surroundings;

[0045] The interactive module uses different multi-head attention blocks to interactively fuse different trajectory temporal features and multi-scale environmental features to obtain the interactive features between the scene and the vehicle;

[0046] The decoding module uses a two-layer MLP to decode the fused features and output the multimodal future trajectory of the predicted target.

[0047] Compared with the prior art, the present system and method have the following advantages:

[0048] 1. The present invention uses a hierarchical graph neural network and an attention feature fusion module to encode the historical trajectory and environmental information of the target vehicle and its surrounding vehicles in the scene. It not only has the advantages of high processing efficiency and parallel computing, but also provides rich feature information for subsequent trajectory prediction to improve prediction accuracy. It comprehensively considers environmental information from both global and local aspects, making up for the shortcomings of traditional methods in terms of insufficient flexibility in complex environments.

[0049] 2. The present invention is not limited to a single interaction, but takes into account multiple interactions, and adopts different attention mechanisms to extract and fuse different interactions; the multi-factor fusion consideration of the vehicle trajectory timing characteristics and environmental characteristics, which have a greater impact on the vehicle trajectory, further improves the adaptability to complex traffic conditions.

[0050] 3. The present invention can obtain more accurate probability distribution and multimodal trajectory prediction results for each target vehicle by decoding the interactive features through a double-layer MLP, so that the autonomous driving vehicle itself can assist autonomous driving according to the trajectory prediction results of the target vehicle, achieve more accurate and reasonable autonomous driving control, and effectively improve the user experience.

[0051] 4. Compared with the traditional trajectory prediction method, the technical solution of the present invention can improve the prediction accuracy of the model without increasing the complexity, and at the same time ensure that the model can adapt to complex and changeable traffic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 The present invention is a flow chart of a method for predicting vehicle trajectories in complex traffic scenarios according to an embodiment of the present invention.

[0053] Figure 2 The figure is an overall framework diagram of a vehicle trajectory prediction method for complex traffic scenarios according to an embodiment of the present invention.

[0054] Figure 3 This is a comparative analysis diagram of the experimental results of the present invention and the existing model. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] like Figure 1 The flowchart of the method of the present invention is shown. The embodiment of the present invention provides a vehicle trajectory prediction method for complex traffic scenarios. The specific steps are as follows:

[0057] Step S1, obtaining historical trajectory information of the target vehicle and surrounding vehicles in the traffic scene in the trajectory encoding module, and using the layered graph neural network VectorNet to extract trajectory time series features in the historical trajectory information of the predicted target and surrounding vehicles;

[0058] Before step S1, the vehicle trajectory sequence and high-precision map data are obtained based on the Argoverse motion prediction dataset, and the obtained data are preprocessed to obtain the historical trajectory information and surrounding environment information of the target vehicle and the surrounding vehicles in the traffic scene.

[0059] Vehicle history trajectory: X = {x i}, where x i ∈R t×2 Represents the position information of agent i in the previous T time steps, where each position is a two-dimensional coordinate (x, y). n represents the n-dimensional real number space, that is, a vector set consisting of n real numbers. Therefore x i It can be viewed as a T×2 matrix, where each row corresponds to the position of a time step.

[0060] Surrounding environment: M = {m i}, where m i ∈R li×2 Indicates the i-th lane, consisting of continuous l i points, each of which is also a two-dimensional coordinate (x, y). This means that m i is a l i ×2 matrix, each row represents a point on the lane.

[0061] A hierarchical graph neural network is used to extract the trajectory time series features from the predicted target and the historical trajectory information of the surrounding vehicles. The specific implementation process is as follows:

[0062] VectorNet is used to create a connected graph for each scene element using vectorized representation. Through the vectorization process, a polyline is obtained, and each vector f belonging to the polyline i It can be regarded as a node in the graph, that is, the feature vector of each broken line length d, f i ∈R d .

[0063]

[0064] where f i (l) Represents the feature vector of the i-th node in the l-th layer; MLP is a multi-layer perceptron, which is used to perform nonlinear transformation on the feature vector. Here L refers to a specific layer configuration. The MLP in the model is configured as 2 linear layers and a ReLU activation function for nonlinearity; pool performs a max-pool operation to obtain information from adjacent nodes; cat performs a concatenation operation.

[0065] Step S2, in the environment encoding module, obtains the environmental information of the traffic scene where the target vehicle is located, and uses the layered graph neural network VectorNet and the attention feature fusion module to extract the multi-scale environmental features in the global and local environmental information around the target vehicle.

[0066] The specific implementation process of extracting environmental features using the hierarchical graph neural network and attention feature fusion module is as follows:

[0067] Step S201, using a layered graph neural network VectorNet to create a connected graph using vector representation to obtain a feature vector;

[0068] Step S202, use the attention feature fusion module AFF to fuse the local and global feature vectors. Before fusing features of different scales, the importance of each feature is evaluated, and then the features are weighted and fused according to the importance, which is expressed as the following formula:

[0069]

[0070] where f i (l) Represents the feature vector of the ith node in the lth layer; MLP is a multi-layer perceptron, which is used to perform nonlinear transformation on feature vectors. Here, L refers to a specific layer configuration. The MLP in the model is configured with 2 linear layers and a ReLU activation function for nonlinearity; pool performs max-pool operations, and cat performs concatenation operations. AFF is an attention feature fusion module, which is used to fuse feature vectors of different scales.

[0071] The steps of feature fusion are as follows:

[0072] Step S221, compressing the features into a vector of a fixed size through global average pooling GAP to reduce computational complexity.

[0073] In step S222, the compressed feature vector is used to fuse multi-scale information through point-by-point convolution and nonlinear activation function.

[0074] Step S223, through the Sigmoid activation function, the output of the fully connected layer is mapped to the interval (0, 1) to obtain the importance weight of each feature. The original features are weighted using the calculated weights, and then the weighted features are multiplied to form the final fusion feature.

[0075] The AFF module focuses on the scale problem of the channel and mainly fuses multi-scale information through point-by-point convolution rather than convolution kernels of different sizes, which helps to keep our encoding module lightweight.

[0076] In step S3, a multi-head attention block (MHAB) is used in the interaction module to interactively fuse the vehicle trajectory temporal features and environmental features to obtain the interaction features between the scene and the vehicle.

[0077] The specific implementation process of step S3 includes the following steps:

[0078] Step S301, different from the common interaction module, the present invention models different types of interactions: vehicle-to-lane (AL), lane-to-lane (LL), lane-to-vehicle (LA) and vehicle-to-vehicle (AA); and uses different attention mechanisms to extract different interactions.

[0079] Step S302, using the self-attention mechanism and the feed-forward neural network (FFN) to update the self-relation (AA, LL); using the cross-attention mechanism and the feed-forward neural network (FFN) to update the cross-relation (AL, LA);

[0080] The multi-head attention mechanism is defined as follows:

[0081]

[0082] where Q, K, V = W q f q , W k f kv , W v f kv

[0083] Where Q, K, and V represent query, key, and value matrices, respectively; W q , W k , W v represents the learnable linear transformation of the attention head; K T Represents the transpose of vector K; dim k Represents the dimension size of the K vector; softmax is a nonlinear activation function used to normalize the attention weights to ensure that the output is a probability distribution.

[0084] The multi-head attention block of the present invention is mainly defined as follows:

[0085]

[0086] where f q represents the query input feature vector, f kv represents the key and value input feature vector; where is the updated feature vector, obtained by adding the result of the multi-head attention operation to the original input feature vector and then normalizing it;

[0087] Norm means normalization operation, which is used to stabilize the training process and accelerate convergence; FFN is a feedforward neural network, which is defined as follows:

[0088] FFN(a)=ReLU(aW1+b1)W2+b2

[0089] Among them, W1 and W2 represent learnable weight matrices; b1 and b2 represent bias vectors of the feedforward network; a represents the input of the feedforward network; ReLU represents the activation function, which is used to introduce nonlinearity so that the model can fit more complex functions.

[0090] Step S303, model each interaction in sequence, and repeat the process L times. In this way, the intermediate features can be updated at each iteration, and then the updated features can be used to calculate the attention in the next iteration.

[0091] Step S4: In the decoding module, the fusion features of the scene and the vehicle are decoded and output to predict the multi-modal future trajectory of the target. The specific implementation process is as follows:

[0092] The fused features are decoded using a 2-layer MLP with ReLU, layer normalization is used after the first layer, and residual connections are used in the last layer to finally obtain the multimodal future trajectory of the predicted target.

[0093] The present invention has carried out a large number of experiments to evaluate the effectiveness and advancement of the present invention. The experiment is trained based on the Argoverse dataset, which has 323,557 scenes and is a commonly used dataset for motion prediction. All experiments were completed under Intel Core i7-10700 CPU, NVIDIA Quadro RTX 4000 8G GPU, Ubuntu20.04.6LTS, python 3.8, Pytorch 1.12, and CUDA 11.3; the training used the Adam optimizer and the strategy of dynamically adjusting the learning rate. The initial learning rate was set to 0.001, and the learning rate was halved every 20 training cycles to achieve a gradual attenuation of the learning rate.

[0094] The evaluation index of the present invention uses the minimum average displacement error mADE k , minimum final displacement error mFDE k , miss rate MR k and minimum final displacement error brier-mFDE k Evaluate the prediction effect. mADE K is the minimum average Euclidean distance between the predicted trajectory and the true trajectory among the K predicted trajectories, minFDE K is the minimum displacement between the final positions of the K predicted trajectories and the true trajectory. K Refers to the ratio of the distance between the ground truth endpoint and the best predicted endpoint that is greater than 2.0 meters.

[0095] The following table shows the experimental results of comparing the prediction results of this embodiment with the prediction results of the existing model.

[0096]

[0097] Figure 3 This is a comparative analysis chart of the prediction results of this embodiment and the prediction results of the existing model; the smaller the value in the table, the better the model performance, and the bold represents the optimal value under each evaluation indicator. Figure 3It can be seen that the minimum final displacement error mFDE and MR of this method are better than classic models such as LaneGCN and mmTransformer. The minimum average displacement error mADE is lower than other models and only slightly higher than HiVT, and Brier-mFDE is lower than other models and only slightly higher than PAGA model.

[0098] The present invention also provides a vehicle trajectory prediction system for complex traffic scenarios, comprising:

[0099] The trajectory encoding module uses vectors to represent the trajectories of the predicted targets and surrounding intelligent agents, and uses a hierarchical graph neural network VectorNet to extract trajectory temporal features from the historical trajectory information of the predicted targets and surrounding vehicles.

[0100] The environment encoding module uses vector representation to predict the trajectory of the target and surrounding intelligent agents, and uses the layered graph neural network VectorNet and the attention feature fusion module to extract multi-scale features of the environment around the target vehicle.

[0101] The interaction module uses different multi-head attention blocks to interactively fuse different trajectory temporal features and multi-scale environmental features to obtain the interaction features between the scene and the vehicle.

[0102] The decoding module uses a two-layer MLP to decode the fused features and output the multimodal future trajectory of the predicted target.

[0103] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A vehicle trajectory prediction method for complex traffic scenarios, characterized in that: The following steps are involved: Step S1, obtaining historical trajectory information of the target vehicle and surrounding vehicles in the traffic scene, and using a hierarchical graph neural network to extract trajectory time series features in the historical trajectory information of the predicted target and surrounding vehicles; Step S2, obtaining environmental information in the traffic scene where the target vehicle is located, and using a hierarchical graph neural network and an attention feature fusion module to extract multi-scale environmental features in the global and local environmental information around the target vehicle; Step S3, using a multi-head attention block to interactively fuse the vehicle trajectory temporal features and environmental features to obtain the interactive features between the environment and the vehicle; Step S4, decoding the fusion features of the scene and the vehicle to output a multimodal future trajectory of the predicted target.

2. The vehicle trajectory prediction method for complex traffic scenarios according to claim 1 is characterized in that: The historical trajectory information of the vehicle and the environmental information in steps S1 and S2 are obtained based on the Argoverse motion prediction dataset, and the obtained data are preprocessed.

3. The vehicle trajectory prediction method for complex traffic scenarios according to claim 1 is characterized in that: The specific implementation process of extracting the trajectory time series features from the historical trajectory information of the predicted target and surrounding vehicles using the hierarchical graph neural network in step S1 is as follows: The hierarchical graph neural network uses vectorization to create a connected graph for each scene element independently; the polyline is obtained through the vectorization process, and each vector f of the polyline is i is regarded as a node in the graph, i.e., a feature vector with a polyline length of d, f i ∈R d , expressed as the following formula: where f i (l) Represents the feature vector of the i-th node in the l-th layer; MLP is a multi-layer perceptron, which is used to perform nonlinear transformation on the feature vector. Here L refers to a specific layer configuration. The MLP in the model is configured as 2 linear layers and a ReLU activation function for nonlinearity; pool performs a max-pool operation to obtain information from adjacent nodes; cat performs a concatenation operation.

4. The vehicle trajectory prediction method for complex traffic scenarios according to claim 1 is characterized in that: The implementation process of extracting environmental features using the hierarchical graph neural network and the attention feature fusion module in step S2 includes the following steps: Step S201, using a layered graph neural network to create a connected graph using vector representation to obtain a feature vector; Step S202, using the attention feature fusion module AFF to fuse local and global feature vectors. Before fusing features of different scales, the importance of each feature is evaluated first, and then the features are weighted and fused according to the importance, which is expressed as the following formula: where f i (l) Represents the feature vector of the i-th node in the l-th layer; MLP is a multi-layer perceptron, which is used to perform nonlinear transformation on the feature vector. Here L refers to a specific layer configuration. The MLP in the model is configured with 2 linear layers and a nonlinear activation function ReLU; pool performs max-pool operation, and cat performs concatenation operation; AFF is the attention feature fusion module, which is used to fuse feature vectors of different scales.

5. The vehicle trajectory prediction method for complex traffic scenarios according to claim 4 is characterized in that: The feature fusion comprises the following steps: Step S221, compressing the features into a vector of a fixed size through global average pooling to reduce computational complexity; Step S222, the compressed feature vector is fused with multi-scale information through point-by-point convolution and nonlinear activation function; Step S223, through the Sigmoid activation function, the output of the fully connected layer is mapped to the interval (0,1), the importance weight of each feature is obtained, the original features are weighted using the calculated weights, and then the weighted features are multiplied to form the final fusion feature.

6. The vehicle trajectory prediction method for complex traffic scenarios according to claim 1, characterized in that: The step S3 comprises the following steps: In step S301, different types of interactions are modeled, including vehicle-to-lane, lane-to-lane, lane-to-vehicle, and vehicle-to-vehicle interactions, and different attention mechanisms are used to extract different interactions. Step S302, using the self-attention encoder and the feedforward neural network to update the self-relation, and using the cross-attention encoder and the feedforward neural network FNN to update the cross-relation: Step S303, model each interaction in sequence and repeat the process L times, updating the intermediate features at each iteration, and then using the updated features to calculate attention in the next iteration.

7. The vehicle trajectory prediction method for complex traffic scenarios according to claim 1 is characterized in that: In step S3, the multi-head attention mechanism is defined as follows: where Q,K,V=W q f q ,W k f kv ,W v f kv Where Q, K, and V represent query, key, and value matrices, respectively; W q , W k , W v represents the learnable linear transformation of the attention head; K T Represents the transpose of vector K; dim k Represents the dimension size of the K vector; softmax is a nonlinear activation function used to normalize the attention weight to ensure that the output is a probability distribution; Then the multi-head attention block definition formula is as follows: where f q represents the query input feature vector, f kv represents the key and value input feature vector; where is the updated feature vector, obtained by adding the result of the multi-head attention operation to the original input feature vector and then normalizing it; Norm means normalization operation, which is used to stabilize the training process and accelerate convergence; FFN is a feedforward neural network, which is defined as follows: FFN(a)=ReLU(aW1+b1)W2+b2 Among them, W1 and W2 represent learnable weight matrices; b1 and b2 represent bias vectors of the feedforward network; a represents the input of the feedforward network; ReLU represents the activation function, which is used to introduce nonlinearity so that the model can fit more complex functions.

8. The vehicle trajectory prediction method for complex traffic scenarios according to claim 1, characterized in that: The implementation process of step S4 is as follows: The fused features are decoded using a 2-layer MLP with ReLU activation function, layer normalization is used after the first layer, and residual connection is used in the last layer to obtain the multimodal future trajectory of the predicted target.

9. A vehicle trajectory prediction system for complex traffic scenarios, characterized in that: It includes trajectory encoding module, environment encoding module, interaction module and decoding module, among which: The trajectory encoding module uses vectors to represent the trajectory of the predicted target and surrounding intelligent agents, and uses a hierarchical graph neural network to extract the trajectory time series features from the historical trajectory information of the predicted target and surrounding vehicles; The environment encoding module uses vector representation to predict the trajectory of the target and surrounding agents, and uses a hierarchical graph neural network and an attention feature fusion module to extract multi-scale features of the target vehicle's surroundings; The interactive module uses different multi-head attention blocks to interactively fuse different trajectory temporal features and multi-scale environmental features to obtain the interactive features between the scene and the vehicle; The decoding module uses a two-layer MLP to decode the fused features and output the multimodal future trajectory of the predicted target.

Images