A vehicle optimal lane-changing timing prompting method and system based on pre-trained LLMs trajectory prediction

Abstract

The present application relates to a kind of vehicle optimal lane changing time prompt method and system based on pre-training LLMs trajectory prediction, belong to intelligent driving and intelligent transportation system field.The present application establishes vehicle motion trajectory data by trajectory prediction module and generates information joint coding feature vector by information joint encoder;The joint coding feature vector is obtained by pre-training LLMs and multilayer vector machine senior interactive representation feature vector;Establish the long short-term memory neural network of multimodal trajectory prediction module, and the trajectory information of target vehicle and its surrounding vehicles in future time domain is predicted;Future trajectory information is input into lane changing intention prediction module, and lane changing scene coding module extracts vehicle lane changing intention feature;The context feature information of lane changing scene is modeled and extracted by week car interaction modeling module;The maximum lane changing probability is output by lane changing intention output module;If target vehicle needs to change lane, the best lane changing time is predicted by lane changing optimal time prediction module.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intelligent driving and intelligent transportation system, and relates to a vehicle optimal lane-changing timing prompting method and system based on pre-trained LLMs trajectory prediction. BACKGROUND

[0002] With the continuous growth of urbanization and the number of cars, the pressure on the traffic network has increased dramatically, and urban traffic has gradually become more dense and complex, with a significantly increased incidence of traffic accidents. At the same time, with the rapid development of autonomous driving technology, accurately predicting the optimal lane-changing timing of vehicles has become one of the key points of autonomous driving safety. For the driving behavior of changing lanes, the lane-changing of the target vehicle not only affects itself but also affects other vehicles, which is easy to cause traffic accidents. Therefore, the lane-changing process needs to consider not only the current positions of surrounding vehicles but also the historical and future trajectories of both and their mutual influence. Lane-changing in urban areas is one of the main causes of scratch accidents, while lane-changing on highways is easy to cause serious accidents. Therefore, accurately finding the optimal lane-changing timing, reducing the impact on other vehicles, and maximizing safety have become one of the keys to intelligent driving.

[0003] Existing trajectory prediction methods mainly fall into two categories: traditional and deep learning-based. Traditional methods such as physical models, kinematic models, and Kalman filters, usually rely on mature mathematical and statistical techniques to make predictions based on historical data and predefined models. However, these trajectory prediction methods have certain limitations when dealing with complex scenarios with intricate interactions and uncertainties. In recent years, deep learning-based methods have shown more promising prospects. Deep learning methods use artificial neural networks to learn complex patterns and relationships from large amounts of data. Generally, convolutional neural networks or recurrent neural networks are used to extract features from the historical trajectories of vehicles and their surrounding road environments, and then predict future driving logic. Although deep learning-based methods have made certain progress, existing methods still have certain limitations: first, there is a lack of understanding and recognition of complex traffic semantics, and existing methods are characterized by time dependence and spatial interaction, focusing more on learning the motion logic of vehicles from visual data, making it difficult to utilize higher-level semantic information such as road environment and traffic signal information, which are crucial for the accuracy of trajectory prediction; second, there is a lack of a clear mechanism to integrate various information and fully understand and utilize their complex interaction relationships.

[0004] In addition, existing lane change intention recognition models such as physical models cannot well capture long-term timing information, and lane change decision is essentially a long-term behavior that needs a certain time interval to complete, and the driving behavior of surrounding vehicles at future time points needs to be considered, and finding the best lane change opportunity is also the same. Therefore, an accurate vehicle future trajectory prediction is needed as the basis for lane change intention recognition and lane change timing prompt, so as to achieve the best lane change timing. SUMMARY

[0005] Therefore, the purpose of the present application is to provide a vehicle optimal lane change timing prompt method and system based on pre-trained LLMs trajectory prediction, which effectively processes different data features using large language models, performs more accurate, realistic, and real traffic distribution multi-modal trajectory prediction, accurately identifies lane change intention for predicted trajectories, and provides the best lane change timing for target vehicles that need to change lanes.

[0006] To achieve the above purpose, one aspect of the present application provides a vehicle optimal lane change timing prompt system based on pre-trained LLMs trajectory prediction, which comprises a trajectory prediction module, a lane change intention prediction module, and a lane change optimal timing prediction module.

[0007] The trajectory prediction module comprises an information joint encoding submodule, an interaction information modeling submodule, and a multi-modal trajectory prediction submodule. The information joint encoding submodule encodes the vehicle historical trajectory information and motion information of the target vehicle and surrounding vehicles to obtain a joint encoding feature vector; the interaction information modeling submodule receives the joint encoding feature vector and generates an interaction representation feature vector through a pre-trained LLM, and matches the dimension of the interaction representation feature vector with the dimension of the joint encoding feature vector through a multi-layer perception machine to obtain a high-level interaction representation feature vector; and the multi-modal trajectory prediction submodule generates the predicted trajectory of the target vehicle and surrounding vehicles in the future time domain according to the high-level interaction representation feature vector.

[0008] The lane change intention prediction module predicts the behavior intention of the target vehicle according to the predicted trajectory of the target vehicle and surrounding vehicles in the future time domain; and the lane change optimal timing prediction module finds the best lane change time through a Markov decision process when the behavior intention is lane change, and issues a prompt.

[0009] Further, the lane change intention prediction module comprises a lane change scene encoding submodule, a surrounding vehicle interaction modeling submodule, and a lane change intention output submodule.

[0010] The lane changing scene coding sub-module encodes the predicted trajectory and extracts a vehicle lane changing behavior intention feature.

[0011] Further, the information joint coding sub-module encodes the historical trajectory information of the vehicle through a trajectory encoder and encodes the motion information of the vehicle through a motion encoder, and fuses the encoding vectors of the trajectory encoder and the motion encoder to obtain the joint encoding feature vector.

[0012] Another aspect of the application provides a vehicle optimal lane changing timing prompting method based on pre-trained LLMs trajectory prediction, which comprises:

[0013] S1, obtaining the motion trajectory data set and the historical time domain trajectory information of the target vehicle and the surrounding vehicles through the vehicle-mounted sensor;

[0014] S2, jointly encoding the motion trajectory data and the historical time domain trajectory information to obtain a joint encoding feature vector;

[0015] S3, inputting the joint encoding feature vector into the pre-trained LLMs to generate an interactive representation feature vector, and matching the dimension of the interactive representation feature vector with the dimension of the joint encoding feature vector through a multilayer perceptron to generate a high-level interactive representation feature vector;

[0016] S4, decoding the high-level interactive representation feature vector in combination with the historical data, the current state of the target vehicle and the surrounding traffic state of the target vehicle to generate the predicted trajectory of the target vehicle and the surrounding vehicles in the future time domain;

[0017] S5, inputting the predicted trajectory of the target vehicle and the surrounding vehicles in the future time domain into a Transformer encoder to obtain a vehicle lane changing intention feature, and inputting the obtained vehicle lane changing intention feature into a GCN network to represent the spatial relationship between the surrounding vehicles and the target vehicle and further extract features;

[0018] S6, inputting the further extracted features of the GCN network into a lane changing intention output module composed of a linear fully connected layer and an activation function to obtain the behavior intention with the maximum probability among the three behavior intentions of left lane changing, right lane changing and keeping straight driving of the target vehicle;

[0019] S7, if the probability of the target vehicle lane changing intention is the maximum, a Markov decision process is used to obtain the maximum expected long-term return, find the best lane changing time point, and issue a prompt.

[0020] Further, in step S2, the motion trajectory data and the spatial information are jointly encoded, specifically, the historical trajectory information of the vehicle is encoded by a trajectory encoder, and the motion information of the vehicle is encoded by a motion encoder, wherein the trajectory encoder and the motion encoder are both encoded by a gated recurrent unit. The joint encoding feature vector is obtained by concatenating the encoding results of the trajectory encoder and the motion encoder.

[0021] Further, in step S7, in the process of obtaining the maximum expected long-term return by the Markov decision process, for each policy of the agent, the corresponding action state value function Q π (s,a) is the unique solution of the behavior action and state of the agent at that time, so the optimal policy is solved by a greedy optimization algorithm * :

[0022] π * (s,a)=argmaxQ π (s,a)

[0023]

[0024] The optimal policy is solved * The corresponding action state value function The time t in the equation is the best lane changing time.

[0025] Wherein, s represents the state of the agent, a represents the action of the agent, N represents the final time, gamma represents a discount factor, r t+k represents the direct reward obtained at t+k, t represents the current time, s t and a t represent the state and action of the agent at t respectively.

[0026] The beneficial effects of the present application are:

[0027] (1) The present application introduces a large language model for interactive modeling, fully utilizes high-level semantic information, and fuses semantic information and scene visual information. By fusing multiple data streams, various prediction capabilities under various traffic conditions can be integrated.

[0028] (2) The present application uses a Transformer as an encoder, which can not only consider the timing and continuous features of the lane changing process, but also extract deeper lane changing behavior intention features. Meanwhile, the present application uses a GCN network to model and enhance the dependence between vehicles.

[0029] (3) The present application aims at the problem of inaccurate lane changing opportunity selection of vehicles, and makes lane changing intention prediction and optimal lane changing opportunity prompt based on the future trajectory of the vehicle, so that the prediction and optimal lane changing opportunity are more in line with the real situation.

[0030] Other advantages, objects, and features of the present application will be apparent to those skilled in the art from the following specification, in conjunction with the accompanying drawings, in which: BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to make the purposes, technical solutions and advantages of the present application clearer, the preferred embodiments of the present application will be described in detail below with reference to the accompanying drawings, in which:

[0032] Figure 1 The structural block diagram of the optimal lane changing opportunity prompt system of vehicles provided by an embodiment of the present application is shown in the figure.

[0033] Figure 2 The structural schematic diagram of the trajectory prediction module is shown in the figure.

[0034] Figure 3 The schematic diagram of the large language model parameter fine-tuning process is shown in the figure.

[0035] Figure 4 The structural schematic diagram of the lane changing intention prediction module is shown in the figure.

[0036] Figure 5 The GCN modeling schematic diagram of the scene information representation module is shown in the figure.

[0037] Figure 6 The calculation block diagram of the long short-term memory neural network unit is shown in the figure. DETAILED DESCRIPTION

[0038] The embodiments of the present application are described below through specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the disclosure of the present specification. The present application can also be implemented or applied in different specific embodiments, and the details in the present specification can be modified or changed in different ways without departing from the spirit of the present application. It should be noted that the figures provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0039] As Figure 1As shown, the vehicle optimal lane changing time prompt system based on pre-training LLMs trajectory prediction provided by an embodiment of the application mainly comprises a trajectory prediction module, a lane changing intention prediction module and a lane changing optimal time prediction module.

[0040] The trajectory prediction module comprises an information joint encoding submodule, an interaction information modeling submodule and a multi-modal trajectory prediction submodule, and the specific implementation process is as shown in Figure 2

[0041] ①The information joint encoding submodule encodes the vehicle historical trajectory information and motion information through a trajectory encoder and a motion encoder respectively, and fuses to generate a joint encoding feature vector g i . The trajectory encoder focuses on the spatial collinear coordinates (x, y) of the vehicle, and the motion encoder is used to process the dynamic state of the vehicle, including speed and acceleration.

[0042] ②The interaction information modeling submodule uses the super strong understanding and reasoning ability of the large language model LLM, uses the NLP pre-training Transformer architecture to calculate and interpret the potential interaction and mutual relationship between the traffic signal and the vehicle trajectory, performs high-level interaction modeling to generate an interaction representation feature vector z i , and finally generates a high-level interaction representation feature vector s i through a multi-layer perception MLP layer.

[0043] ③The multi-modal trajectory prediction submodule uses a long short-term memory neural network to fuse multiple aspects of data flow, integrates various prediction capabilities under various traffic conditions, can comprehensively understand vehicle dynamics, and reliably and accurately predicts vehicle trajectories. The module takes the high-level interaction representation feature vector s i as the input of the long short-term memory neural network, and finally generates a sequence with time consistency and social acceptability, predicting the trajectory information of the target vehicle and its surrounding vehicles in the future time domain.

[0044] The lane changing intention prediction module comprises a lane changing scene encoding submodule, a surrounding vehicle interaction modeling submodule and a lane changing intention output submodule, and the specific implementation process is as shown in Figure 4

[0045] ①The lane changing scene encoding submodule uses a Transformer encoder to encode the predicted trajectory obtained by the multi-modal trajectory prediction module, considers the timing and continuous features of the lane changing process, not only extracts the hidden features of the vehicle lane changing intention, but also extracts deeper lane changing behavior intention features. The lane changing scene encoding submodule uses the self-attention mechanism of the Transformer to enable each vehicle to pay attention to the influence of the position of the surrounding vehicles on its current lane changing intention at each time step. The self-attention mechanism is represented as follows: ​​

[0046]

[0047] where Q represents a query vector, K and V are a set of key-value pairs, d k is the input feature dimension.

[0048] The multi-head attention mechanism in the Transformer encoder is represented as:

[0049] MultiHead(Q,K,V)=concat(head1,head2,...,head h )W O

[0050] head i =Attention(QW i Q ,KW i K ,VW i V )

[0051] where W is a parameter matrix.

[0052] ②The surrounding vehicle interaction modeling sub-module considers the influence of surrounding vehicles on the driving trajectory of the target vehicle, represents the spatial relationship between surrounding vehicles and the target vehicle through a graph convolution network (GCN), and extracts context feature information of the lane changing scene, wherein the GCN modeling process of the surrounding vehicle interaction modeling sub-module is as shown in Figure 5 .

[0053] The surrounding vehicle interaction modeling sub-module uses a graph convolution network to enhance the mutual dependence between vehicles. When constructing the GCN model, spectral domain graph convolution is used to represent the spatial position information of the vehicles, and the calculation formula is as follows:

[0054]

[0055] where θ is the convolution sum, *g is the spectral domain graph convolution, x is the state information of the vehicle, σ is the activation function, k is the size of the graph convolution kernel, δ j is the convolution kernel parameter, and L j is a normalized Laplacian matrix used to construct the graph.

[0056] ③The lane changing intention output sub-module includes a linear fully connected layer and an activation function, wherein the fully connected layer outputs the predicted lane changing behavior intention, and the Softmax activation function outputs the probabilities of left lane changing, right lane changing and keeping the current road straight. This module finally takes the behavior operation with the maximum probability as the predicted driving intention.

[0057]

[0058] wherein U=3, indicating that there are only 3 categories of output, i.e. changing lanes to the left, changing lanes to the right, and keeping the current lane straight, W j , W u represents the corresponding weight matrix, u=0, 1, 2 represents the probability of judging once for each of the three categories of changing lanes to the left, changing lanes to the right, and keeping the current lane straight, x T is the transpose of the GCN output vector.

[0059] The lane changing optimal timing prediction module uses Markov decision process (MDP) to solve the sequential decision problem in a dynamic environment, which is used to simulate the randomness strategy and return that can be achieved by the agent in an environment with Markov property of system state. The module finds the best lane changing timing by maximizing the expected long-term return.

[0060] Another embodiment of the application provides a vehicle optimal lane changing timing prompting method based on pre-trained LLMs trajectory prediction, which realizes the following steps:

[0061] 1. A roadside sensor device is used to obtain and establish a set of intersection vehicle motion trajectory data. The vehicle trajectory in the roadside area is composed of a series of discrete points, and the vehicle trajectory information includes lane, historical trajectory and other information.

[0062] 2. Based on the vehicle sensor information, the historical time domain trajectory information of the target vehicle and surrounding vehicles is obtained, including longitudinal and lateral position, lateral position, speed and acceleration.

[0063] Among them, the vehicle motion trajectory data set and the historical time domain trajectory information can be collected by the vehicle sensor and processed at the vehicle end.

[0064] 3. The vehicle motion trajectory data and historical time domain trajectory information in step 1 and step 2 are jointly encoded by the information joint encoding module to obtain the information joint encoding feature vector g i .

[0065] Specifically, the trajectory encoder uses a gated recurrent unit (GRU) for encoding to capture the time variation in the vehicle motion trajectory data, thereby deeply understanding the motion trajectory of the vehicle:

[0066] E traj =GRU traj (x,y)

[0067] wherein x and y are vectors composed of trajectory coordinate points, x represents the lateral position, and y represents the lateral position.

[0068] The calculation steps of GRU are as follows:

[0069]

[0070] Among them, x t Input information for the current moment, h t-1 is the hidden state of the previous moment, h t is the hidden state passed to the next moment, To select the hidden state, r t and z t are reset gate and update gate respectively, σ is sigmoid function, tanh(·) is tanh function, W z 、W r , W is the weight matrix.

[0071] The motion encoder uses a gated recurrent unit (GRU) to encode the vehicle's motion state, providing insight into the real-time effects of vehicle motion:

[0072] E motion =GRU state (V,A)

[0073] Where V and A are the vectors composed of vehicle velocity and acceleration, respectively. In the motion encoder, the calculation process of GRU is consistent with that of the trajectory encoder.

[0074] The feature E traj and E motion Connect them in series to get the joint coding feature vector g i :

[0075] g i =Concat(E traj ,E motion )

[0076] Where Concat(·) represents the consolidation function.

[0077] 4. The information joint encoding feature vector in step 3 is input into the advanced interaction modeling layer (i.e., large language model) for further modeling calculation to explain the potential interactions and relationships between vehicle mutual motion and vehicle historical trajectories. The advanced interaction representation feature vector s is obtained through the large language model LLMs and multi-layer perceptron MLP. i .

[0078] 1) If Figure 3 As shown, fine-tune the parameters Q and V of the large language model LLMs. First, freeze all pre-trained parameters of LLMs;

[0079] 2) Injecting the required new trainable parameters through the real-time low-rank adaptation technology LoRA, which is applied to the query vector Q and key-value vector V of the attention layer; among them, for the pre-trained weight matrix W0+ΔW=W0+BA, the parameter W0 is frozen during training, and only the parameters in A and B are trained;

[0080] 3) Pass the required trainable parameters into the network for pre-setting;

[0081] 4) Update the LoRA approximation in step 2) to: W f =W+BA,W f represents the updated weight matrix, W represents the pre-trained weight matrix, and B and A represent the trainable parameters;

[0082] 5) During the training process, W∈R d×k Keeping it frozen, B and A are treated as trainable parameters, B is initialized to zero and A is initialized to a Gaussian distribution; therefore, the forward transfer function of LoRA can be concisely expressed as: a j represents the pre-trained forward transfer function, represents the updated forward transfer function;

[0083] 6) The joint encoding feature vector g i Passed to LLMs after the new trainable parameters have been injected via LoRA;

[0084] 7) Based on the pre-trained Transformer module and the new trainable parameters, the LLMs are trained first, and then the interaction representation feature vector z is generated by the trained LLMs i =LLMs(g i );

[0085] 8)z i After passing through the multi-layer perceptron MLP and the input g i The dimensions of the vector are matched, and finally a high-level interactive representation feature vector s is generated. i The multi-layer perceptron (MLP) is a generalization of the perceptron. Its main feature is that it has multiple neuron layers, usually including an input layer, one or more hidden layers, and an output layer. The implementation process of MLP is as follows:

[0086]

[0087] Where w is the weight matrix, b is the bias vector, x is the input of the network, y is the output of the network, u is the intermediate vector, and sign is the activation function.

[0088] 5. Represent the high-level interaction feature vector s in step 4 iThe input is decoded by a multi-modal trajectory prediction module, which is implemented by an LSTM framework and introduces a certain degree of randomness r, so that the model can generate a series of possible future trajectories. The multi-modal trajectory prediction module is constructed in the following way:

[0089] The LSTM generates a sequence with temporal consistency and social acceptability according to the historical data, the current self-vehicle state and the surrounding traffic conditions. This dynamic process can be summarized as:

[0090] Trajectory future = LSTM ([s i ; r])

[0091] Where Trajectory future is the generated prediction trajectory sequence with temporal consistency and social acceptability, and r is the introduced vector. The role of r is to introduce a certain degree of randomness, so that the model can generate a series of possible future trajectories, thereby improving its robustness.

[0092] As shown in Figure 6 , the calculation process of LSTM is as follows:

[0093]

[0094] Where i t is the input gate, f t is the forget gate, o t is the output gate, c t is the memory state, h t is the output state of the network, h t-1 and c t-1 are the network output state and memory unit at the previous time, respectively, W xi , W hi , W ci , W xf , W hf , W cf , W xo , W ho , W co , W xc , W hc are weight matrices, and b i , b f , b o , b c are bias vectors.

[0095] 6, input the obtained future trajectory of the target vehicle and its surrounding vehicles into the Transformer encoder to obtain the vehicle lane changing intention feature.

[0096] 7. The obtained vehicle lane change intention features are used through a graph convolutional network to obtain the close relationship between vehicles. At the same time, the impact of surrounding vehicles on the current target vehicle is modeled to further extract features.

[0097] 8. The feature vector extracted by the graph convolutional network is input into the lane change intention output submodule. This submodule outputs the probabilities of changing lanes to the left, changing lanes to the right, and staying straight on the current road through a fully connected layer and activation function. The action with the highest output probability will be regarded as the driving intention predicted by the model.

[0098] 9. If the current probability of the vehicle changing lanes is high, the Markov decision process (MDP) is used to maximize the expected long-term return based on the lane change optimal timing prediction module, find the optimal lane change time point, and finally issue a prompt.

[0099] MDP can be expressed as a five-tuple (S, A, T, γ, R), where S is the state space, A is the behavior space, T is the state transition function, γ is the discount factor, and R is the feedback value of the benefit evaluation. π To evaluate the quality of the agent's strategy, as follows:

[0100]

[0101] Among them, t represents the current time, N represents the final time, r t+k Represents the direct reward obtained at time t+k.

[0102] For each strategy π of the agent, its corresponding action-state value function Q π are the only solutions to the agent’s behavior and state at that time. A greedy optimization algorithm is used to solve the optimal strategy π * :

[0103] π * (s,a)=argmaxQ π (s,a)

[0104] The optimal strategy π is * The corresponding action state value function The time t in the figure is the best time to change lanes.

[0105] In the above embodiment, the specific training parameters of the trajectory prediction module and the lane change intention prediction module are described as follows:

[0106] 1. Trajectory prediction module

[0107] ① Hyperparameters setting: The batch size and initial learning rate are set to 132 and 0.001, respectively, and the model is trained for 50 epochs. First, the information joint encoding module is trained, then the interaction information modeling module, and finally the multi-modal trajectory prediction module. The hidden dimension of all feature vectors is uniformly configured to 128.

[0108] ② Evaluation metrics: For the trajectory prediction module, the average displacement error (ADE) and final displacement error (FDE) are used as evaluation metrics for trajectory accuracy. The definitions of the two are as follows:

[0109]

[0110] where N is the total number of trajectories, T is the prediction range, is the ground truth coordinate, is the predicted coordinate of the i-th trajectory at time t.

[0111] ③ Loss function setting: The smooth-L1 loss function is used:

[0112]

[0113] The smooth-L1 loss function is a widely used loss function in computer vision tasks such as object detection, object recognition, and face recognition. Compared with L1 and L2 loss functions, the smooth-L1 loss function has better robustness, faster convergence rate, and smoother gradient.

[0114] 2. Lane change intention prediction module

[0115] ① Hyperparameters setting: The batch size and initial learning rate are set to 64 and 0.001, respectively, and the model is trained for 50 epochs. First, the Transformer encoding module is trained, then the GCN modeling module. The number of GCN network layers is set to 2, and the number of attention heads of the Transformer multi-head attention layer is set to 8.

[0116] ② Loss function setting: The smooth-L1 loss function is also used.

[0117] ③ Model training: The framework of this module is implemented using Pytorch, and the model seed is set to 1. Mixed precision training is used to reduce memory consumption and speed up training.

[0118] Finally, it is to be explained that the above embodiments are only used to illustrate the technical solutions of the present application but not to limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the purpose and scope of the technical solutions, and all should be covered in the scope of the claims of the present application.

Claims

1. A vehicle optimal lane-changing timing prompting method based on pre-trained LLMs trajectory prediction, characterized in that, The method comprises: acquiring a motion trajectory data set and historical time domain trajectory information of a target vehicle and surrounding vehicles through a vehicle-mounted sensor; jointly encoding the motion trajectory data and the historical time domain trajectory information to obtain a joint encoding feature vector; inputting the joint encoding feature vector into a pre-trained LLM to generate an interaction representation feature vector, and matching the dimension of the interaction representation feature vector with the dimension of the joint encoding feature vector through a multilayer perceptron to generate a high-level interaction representation feature vector; decoding the high-level interaction representation feature vector in combination with historical data, a current state of the target vehicle and a surrounding traffic state of the target vehicle to generate a predicted trajectory of the target vehicle and the surrounding vehicles in a future time domain; inputting the predicted trajectory of the target vehicle and the surrounding vehicles in the future time domain into a Transformer encoder to obtain a vehicle lane-changing intention feature, inputting the obtained vehicle lane-changing intention feature into a GCN network, representing a spatial relationship between the surrounding vehicles and the target vehicle through the GCN network, and further extracting features; inputting the further extracted features of the GCN network into a lane-changing intention output module composed of a linear fully connected layer and an activation function to obtain a behavior intention with the maximum probability among three behavior intentions of left lane-changing, right lane-changing and keeping straight on the current road of the target vehicle; if the probability of the lane-changing intention of the target vehicle is the maximum, obtaining a maximum expected long-term reward through a Markov decision process, finding an optimal lane-changing time point, and issuing a prompt.

2. The method of claim 1, wherein, The jointly encoding the motion trajectory data and the historical time domain trajectory information comprises: encoding the historical trajectory information of the vehicle through a trajectory encoder and encoding the motion information of the vehicle through a motion encoder; the trajectory encoder and the motion encoder both adopt a gated recurrent unit for encoding; concatenating the encoding results of the trajectory encoder and the motion encoder to obtain the joint encoding feature vector.

3. The method of claim 1, wherein, In a process of obtaining a maximized expected long-term return through the Markov decision process, for each policy of the agent , a corresponding action state value function , and a unique solution of the agent's current behavior action and state, so as to solve the optimal policy through a greedy optimization algorithm : The optimal strategy sought The corresponding action-state value function The time point in The optimal lane-changing time; wherein, represents the state of the agent, represents the action of the agent, represents the final time point, represents the discount factor, represents the direct reward obtained at The time point, represents the current time point, and respectively represent the state and action of the agent at The time point.

Images