A multi-vehicle cooperative trajectory prediction method and system based on federated learning

Abstract

The application discloses a kind of multi-vehicle cooperative trajectory prediction method and system based on federal learning, it is related to intelligent transportation and artificial intelligence cross technical field.The method includes: each vehicle collects and pre-processes own historical trajectory data, generates semantic feature vector with driving intention discrimination by comparing learning task trajectory encoding module training;Vehicle exchanges semantic feature vector through communication, constructs multiple relationship graphs based on spatial distance and semantic similarity, and utilizes multi-relationship graph attention mechanism to carry out message transmission and feature aggregation, generates vehicle representation fusing context information;Vehicle representation is input into trajectory decoding module, and future trajectory prediction result is output;Each vehicle uploads local model parameter to central server, and server is aggregated according to vehicle contribution degree, generates global model parameter and issues update.The application realizes high-precision multi-vehicle cooperative trajectory prediction under the premise of protecting data privacy.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and artificial intelligence, and in particular to a multi-vehicle cooperative trajectory prediction method and system based on federated learning. Background Technology

[0002] In the fields of intelligent transportation and autonomous driving, vehicle trajectory prediction is a key technology for achieving safe decision-making and collaborative control. Traditional methods often employ centralized learning frameworks, relying on roadside units or cloud servers to collect raw trajectory data from each vehicle for unified modeling. However, this approach carries serious privacy risks and is highly dependent on communication bandwidth and centralized infrastructure, making it difficult to apply to large-scale, decentralized vehicle-to-everything (V2X) collaborative scenarios.

[0003] In recent years, some studies have attempted to introduce federated learning frameworks for distributed model training to protect data privacy. However, existing federated trajectory prediction methods are typically trained independently on local data only, lacking effective modeling of neighboring vehicle behavior, resulting in insufficient collaborative perception capabilities. Furthermore, most methods construct interaction graphs using simple spatial adjacency relationships, making it difficult to capture complex semantic-level interactions between vehicles, and they fail to introduce self-supervised mechanisms in local training to improve feature representation quality, leading to poor generalization ability of the models in heterogeneous data scenarios.

[0004] Publication number CN116595478A, entitled "An Interactive Vehicle Multimodal Trajectory Prediction Method Based on GRU-GCN," proposes to extract vehicle dynamic features through a GRU encoder, model the spatiotemporal interaction between vehicles using a graph convolutional network (GCN), and output multimodal predicted trajectories by combining a behavior recognition module. However, this method relies on a centralized server for model training, posing a privacy risk, and lacks a federated learning mechanism, failing to achieve multi-vehicle collaborative perception and model updates while protecting data privacy.

[0005] The paper, published under CN118430256A and titled "A Vehicle Trajectory Prediction Method Based on Aerial Computation and Federated Learning," proposes using aerial computing technology to achieve concurrent uploading and aggregation of gradients from multiple vehicle models through signal superposition effects, significantly reducing communication latency in federated learning. However, this method does not consider real-time interactions between vehicles during local training, relying solely on a central base station for model aggregation. It lacks deep perception and collaborative modeling of neighboring vehicle behaviors, resulting in insufficient collaborative reasoning capabilities of the prediction model in complex traffic scenarios.

[0006] Therefore, how to effectively integrate semantic features of multiple vehicles and model diverse interaction relationships without sharing the original data, and improve the accuracy of collaborative trajectory prediction under the federated learning framework, remains a technical challenge that urgently needs to be solved. Summary of the Invention

[0007] To address the aforementioned technical problems, one technical solution adopted by this invention is: providing a multi-vehicle cooperative trajectory prediction method based on federated learning, the method comprising:

[0008] S1: Collect and store historical trajectory data of the target vehicle, including position, speed, acceleration, heading angle, and timestamp information. Use the Z-score normalization method to normalize each dimension of the historical trajectory data, and divide the normalized historical trajectory data into fixed time windows. The trajectory sequence is obtained, and each target vehicle maintains a collaborative prediction model locally, which includes a trajectory encoding module, a graph interaction fusion module, and a trajectory decoding module;

[0009] S2: Design a contrastive learning task, input the trajectory sequence into the trajectory encoding module to obtain the semantic feature vector of the trajectory, and train the trajectory encoding module by minimizing the contrastive learning loss function;

[0010] S3: The target vehicle communicates with neighboring vehicles to obtain the semantic feature vectors of the neighboring vehicles' trajectories, calculates the distance between each vehicle, defines three graphs based on the distance and semantic feature vector similarity, and combines the multi-relationship graph attention mechanism in the graph interaction fusion module to perform message passing and feature aggregation on the three graphs to generate vehicle representations.

[0011] S4: Input the vehicle representation into the trajectory decoding module to obtain the future trajectory of the target vehicle and its neighboring vehicles. The trajectory prediction sequence at each time step is combined with the prediction loss function and the contrastive learning loss function, and the parameters of the collaborative prediction model of the target vehicle are jointly updated through the backpropagation algorithm.

[0012] S5: Participating vehicles upload the locally trained collaborative prediction model parameters to the central server. The central server performs weighted aggregation based on the contribution of each participating vehicle to generate a global collaborative prediction model. The global collaborative prediction model parameters are then distributed to each participating vehicle to replace its local collaborative prediction model parameters.

[0013] S6: Repeat S1–S5 until the global collaborative prediction model converges to obtain the final collaborative prediction model;

[0014] S7: During the model deployment phase, the target vehicle acquires its real-time motion state through onboard sensors, forming a trajectory sequence. The trajectory encoding module of the collaborative prediction model is used to extract semantic feature vectors, and then the semantic feature vectors of the trajectories of neighboring vehicles are obtained. The vehicle representation is obtained through the graph interaction fusion module, and finally the trajectory decoding module outputs the trajectory prediction sequence of the target vehicle and neighboring vehicles.

[0015] Furthermore, the Z-score normalization method is used to normalize the various dimensions of the historical trajectory data, including:

[0016] The location is divided into two dimensions: longitude and latitude. Velocity, acceleration, and heading angle are each represented as one dimension. These five dimensions are then processed and expressed as follows: Its normalization process is as follows:

[0017] ,

[0018] in, For historical trajectory data feature dimensions Standardized eigenvalues For historical trajectory data feature dimensions The original eigenvalues, For historical trajectory data feature dimensions The mean, For historical trajectory data feature dimensions standard deviation This represents the feature dimension of historical trajectory data. .

[0019] Furthermore, the contrastive learning task includes:

[0020] Using a trajectory sequence of the target vehicle as an anchor sample, positive samples of the anchor sample are generated through time clipping. Multiple trajectory sequences of the target vehicle are combined into a training batch, and the other trajectory sequences in the batch other than the anchor sample are used as negative samples of the anchor sample.

[0021] By minimizing the contrastive learning loss function, the training trajectory encoding module learns semantic feature vectors that discriminate driving intentions.

[0022] Furthermore, the time trimming includes:

[0023] From anchor sample trajectory sequence In the process, a start time step is randomly selected. and end time step ,satisfy and Extracting subsequences As a clipped trajectory segment The length of the clipped trajectory segment;

[0024] For each time step , The length of the clipped trajectory segment is restored to a fixed time window through a filling mechanism. :

[0025] for Preserve the original data.

[0026] for For frames missing at the beginning, a forward expansion strategy is used to copy the starting frame. Fill in,

[0027] for Missing frames at the end, copy the termination frame. Fill in the blanks.

[0028] Furthermore, the contrastive learning loss function includes:

[0029] The cosine similarity between the anchor sample and its positive samples is taken as the positive term, and the cosine similarity between the anchor sample and multiple negative samples is taken as the negative term.

[0030] The positive and negative terms were respectively subjected to a test with the natural constant as the base and temperature parameter. Weighted by the exponent of the denominator;

[0031] By normalizing all weighting terms using a normalized exponential function, the relative probabilities of anchor samples and positive samples are obtained.

[0032] The trajectory encoding module is trained using the negative logarithm that minimizes the relative probability as the contrastive learning loss function;

[0033] Wherein, the temperature parameter satisfy .

[0034] Furthermore, the trajectory encoding module includes:

[0035] The trajectory encoding module includes an encoder consisting of two stacked Long Short-Term Memory (LSTM) networks, which processes the anchor sample trajectory sequences. With positive sample trajectory sequence , For a fixed time window, The number of feature dimensions of the trajectory sequence. ;

[0036] For each input sequence, obtain the hidden state of the Long Short-Term Memory network at each time step, forming a time window of fixed length. The sequence of hidden vectors;

[0037] The hidden state sequence is flattened into a one-dimensional vector along the time dimension, and then mapped to a fixed dimension through a fully connected layer. The semantic feature vector, where, The preset semantic feature vector length.

[0038] Furthermore, the target vehicle communicates with neighboring vehicles, including:

[0039] The target vehicle uses its position at the end of its anchor sample trajectory sequence as a reference point to calculate the Euclidean distance between it and surrounding vehicles at the same time. Based on this Euclidean distance, within its set maximum sensing and communication range... Internal filtering selects those with a distance less than The candidate neighbor vehicles are selected, and the K nearest vehicles are selected as neighbor vehicles. If the number of candidate neighbor vehicles is less than K, the zero vector with the same dimension as the semantic feature vector is used to replace the feature input of the missing vehicle.

[0040] Furthermore, the three graphs defined by distance and semantic feature vector similarity include:

[0041] The first graph is a connectivity graph based on spatial proximity. Calculate the reciprocal of the Euclidean distance between vehicles as the initial weights, and normalize the initial weights for each vehicle. If the normalized weights exceed a preset threshold... If , then a connection is established between the two vehicles, indicating a spatial connectivity relationship;

[0042] The second graph is a connection graph based on semantic behavior similarity. Calculate the cosine similarity of semantic feature vectors between vehicles as the initial weights, and then normalize them. If the normalized weights exceed a preset threshold... If , then a connection is established, representing a semantic connectivity relationship;

[0043] The third figure is a comprehensive connectivity graph based on spatial-semantic fusion. The normalized weights corresponding to the first two graphs are weighted and summed according to a preset ratio to obtain the fusion weight, and then the fusion weight is normalized again; if the fusion weight exceeds a preset threshold... If a connection is established, it represents a comprehensive interactive connectivity relationship.

[0044] Furthermore, the multi-relationship graph attention mechanism in the graph interaction fusion module includes:

[0045] For the connection graph based on spatial proximity Connection graph based on semantic behavior similarity and a comprehensive connectivity graph based on spatial-semantic fusion Each of them constructs an independent graph attention layer;

[0046] In each graph attention layer, for any vehicle node in the graph Its semantic feature vector After a learnable linear transformation, it becomes the query vector, and all neighboring nodes connected to it in the relational graph are included. semantic feature vector After undergoing learnable linear transformations, the vectors are used as key and value vectors, respectively. The attention weights are then calculated as follows:

[0047] ,

[0048] in, Indicates the type of relationship diagram. In order to be in Vehicle nodes in the relationship diagram Its neighboring nodes Attention weights Vehicle nodes and neighboring nodes semantic feature vectors, for The learnable projection matrix corresponding to the relationship graph. for vector dimension, Along the neighbor node dimension Normalize, In order to be in Vehicle nodes in the relationship diagram The set of connected neighboring nodes;

[0049] Based on the calculated attention weights, the semantic feature vectors of neighboring nodes are weighted and aggregated to obtain the th... Vehicle nodes under each relationship Contextual representation:

[0050] ,

[0051] in, In order to be in Vehicle nodes in the relationship diagram Contextual representation, In relation Vehicle nodes in the diagram Its neighboring nodes Attention weights In order to be in Vehicle nodes in the relationship diagram The set of connected neighbor nodes, The projection matrix of the value vector. Neighboring nodes semantic feature vectors;

[0052] Context representation obtained from the three relation diagrams The data is stitched together, and vehicle nodes are generated through an output projection layer. Vehicle characterization :

[0053] ,

[0054] in, For vehicle nodes Vehicle characterization, This represents vector concatenation. To output the projection matrix, The context representations are obtained under the three relational graphs.

[0055] Furthermore, the trajectory decoding module includes:

[0056] The trajectory decoding module includes a feedforward neural network composed of multiple fully connected layers. It takes the vehicle representation of each vehicle node output in step S3 as input, processes it layer by layer through linear transformation and the ReLU activation function, and finally outputs a sequence containing... A prediction vector of 1 element;

[0057] In the prediction vector, each pair of consecutive elements represents the future nth... The predicted latitude and longitude coordinates at each time step, arranged in chronological order, constitute the future coordinates of the target vehicle. A trajectory sequence at each time step.

[0058] Furthermore, the combination of the prediction loss function and the contrastive learning loss function includes:

[0059] Define the prediction loss function Used to measure the model's predicted trajectory With the actual future trajectory The difference between them is used as the average of the sum of squared Euclidean distances at each time step as the loss:

[0060] ,

[0061] in, For predicting the loss function, The time step length for predicting the trajectory, For the true trajectory in the 1st The position coordinates of each time step. To predict the trajectory in the 1st The position coordinates of each time step. The square of the Euclidean distance between the predicted point and the true point;

[0062] The prediction loss function and the contrastive learning loss function are weighted and summed to construct a comprehensive loss function:

[0063] ,

[0064] in, These are adjustable hyperparameters used to balance trajectory prediction accuracy and feature representation quality. For the comprehensive loss function, For predicting the loss function, To learn the loss function by comparison;

[0065] Based on the aforementioned comprehensive loss function The gradient is calculated using the backpropagation algorithm, and the Adam optimizer is used to jointly update all trainable parameters of the trajectory encoding module, graph interaction fusion module, and trajectory decoding module.

[0066] Furthermore, the weighted aggregation based on the contribution of each participating vehicle includes:

[0067] The central server uses the amount of local historical trajectory data used by each participating vehicle in this round of training as a measure of its contribution and calculates the aggregation weight.

[0068] For the vehicle ensemble participating in this round of training ,vehicle Aggregate weights Based on the amount of its local historical trajectory data The ratio to the sum of the total data for all participating vehicles is determined as follows:

[0069] ,

[0070] in, Indicates vehicle The number of trajectory sequences used for updating the global collaborative prediction model in this round of training. Indicates vehicle The number of trajectory sequences used for updating the global collaborative prediction model in this round of training. For vehicles The aggregate weights used for updating the global collaborative prediction model in this round of training. For the assembly of vehicles participating in this round of training;

[0071] The central server is based on the calculated weights For all participating vehicles, upload local collaborative prediction model parameters Perform a weighted average to generate new global collaborative prediction model parameters. :

[0072] ,

[0073] in, To generate new global collaborative prediction model parameters, For vehicles The aggregate weights used for updating the global collaborative prediction model in this round of training. For vehicles Uploaded local collaborative prediction model parameters, The vehicles participating in this round of training are assembled.

[0074] Furthermore, a multi-vehicle cooperative trajectory prediction system based on federated learning includes:

[0075] Data acquisition and preprocessing module: This module collects historical trajectory data of the vehicle itself and normalizes it using the Z-score standardization method, dividing it into fixed time windows. The trajectory sequence;

[0076] Trajectory encoding calculation module: used to perform a contrastive learning task locally, encoding the input trajectory sequence into a semantic feature vector with driving intent discriminative capabilities;

[0077] The graph interaction fusion computing module is used to obtain semantic feature vectors of surrounding vehicles through vehicle-to-vehicle communication, construct a multi-relationship graph based on spatial distance and semantic similarity, and use the multi-relationship graph attention mechanism for message passing and feature aggregation to generate vehicle representations.

[0078] Trajectory decoding calculation module: used to decode the fused vehicle representation into the future values ​​of the target vehicle and its neighboring vehicles. The position sequence of each time step;

[0079] Model aggregation and update module: used to upload the parameters of the locally trained collaborative prediction model to the central server, and receive the global collaborative prediction model parameters issued by the central server to replace the local model for the next round of training;

[0080] The central server is independently configured to receive collaborative prediction model parameters uploaded by each participating vehicle, calculate the aggregation weight based on the local data volume of each vehicle, perform a weighted average of the collaborative prediction model parameters, generate updated global collaborative prediction model parameters, and distribute them to each participating vehicle.

[0081] The beneficial effects of this invention are:

[0082] This invention achieves high-precision prediction of the future trajectory of a target vehicle by performing self-supervised training based on contrastive learning on a local vehicle and combining it with a multi-relationship graph attention mechanism for contextual feature fusion, without sharing the original trajectory data. It balances data privacy protection and collaborative perception performance and is suitable for federated learning scenarios in multi-vehicle collaborative environments. Attached Figure Description

[0083] Figure 1 The flowchart illustrates a multi-vehicle cooperative trajectory prediction method based on federated learning provided by an embodiment of the present invention.

[0084] Figure 2 The diagram illustrates the architecture of a multi-vehicle cooperative trajectory prediction method based on federated learning, as provided in an embodiment of the present invention.

[0085] Figure 3 The diagram shows a block diagram of a multi-vehicle cooperative trajectory prediction system based on federated learning provided by an embodiment of the present invention. Detailed Implementation

[0086] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0087] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments.

[0088] Figure 1 This is a flowchart of a multi-vehicle cooperative trajectory prediction method based on federated learning provided by an embodiment of the present invention. The method includes:

[0089] S1: Collect and store historical trajectory data of the target vehicle, including position, speed, acceleration, heading angle, and timestamp information. Use the Z-score normalization method to normalize each dimension of the historical trajectory data, and divide the normalized historical trajectory data into fixed time windows. The trajectory sequence is obtained, and each target vehicle maintains a collaborative prediction model locally, which includes a trajectory encoding module, a graph interaction fusion module, and a trajectory decoding module;

[0090] S2: Design a contrastive learning task, input the trajectory sequence into the trajectory encoding module to obtain the semantic feature vector of the trajectory, and train the trajectory encoding module by minimizing the contrastive learning loss function;

[0091] S3: The target vehicle communicates with neighboring vehicles to obtain the semantic feature vectors of the neighboring vehicles' trajectories, calculates the distance between each vehicle, defines three graphs based on the distance and semantic feature vector similarity, and combines the multi-relationship graph attention mechanism in the graph interaction fusion module to perform message passing and feature aggregation on the three graphs to generate vehicle representations.

[0092] S4: Input the vehicle representation into the trajectory decoding module to obtain the future trajectory of the target vehicle and its neighboring vehicles. The trajectory prediction sequence at each time step is combined with the prediction loss function and the contrastive learning loss function, and the parameters of the collaborative prediction model of the target vehicle are jointly updated through the backpropagation algorithm.

[0093] S5: Participating vehicles upload the locally trained collaborative prediction model parameters to the central server. The central server performs weighted aggregation based on the contribution of each participating vehicle to generate a global collaborative prediction model. The global collaborative prediction model parameters are then distributed to each participating vehicle to replace its local collaborative prediction model parameters.

[0094] S6: Repeat S1–S5 until the global collaborative prediction model converges to obtain the final collaborative prediction model;

[0095] S7: During the model deployment phase, the target vehicle acquires its real-time motion state through onboard sensors, forming a trajectory sequence. The trajectory encoding module of the collaborative prediction model is used to extract semantic feature vectors, and then the semantic feature vectors of the trajectories of neighboring vehicles are obtained. The vehicle representation is obtained through the graph interaction fusion module, and finally the trajectory decoding module outputs the trajectory prediction sequence of the target vehicle and neighboring vehicles.

[0096] In this embodiment, firstly, historical trajectory data of the target vehicle is collected and stored, including position, speed, acceleration, heading angle, and timestamp information. The Z-score normalization method is used to normalize each dimension of the historical trajectory data, and the normalized historical trajectory data is divided into fixed time windows. The trajectory sequence is obtained, and each target vehicle maintains a collaborative prediction model locally, which includes a trajectory encoding module, a graph interaction fusion module, and a trajectory decoding module. A contrastive learning task is designed, where the trajectory sequence is input into the trajectory encoding module to obtain the semantic feature vector of the trajectory, and the trajectory encoding module is trained by minimizing the contrastive learning loss function. Subsequently, the target vehicle communicates with neighboring vehicles to obtain the semantic feature vectors of the neighboring vehicles' trajectories, calculates the distance between each vehicle, defines three graphs based on the distance and semantic feature vector similarity, and combines the multi-relationship graph attention mechanism in the graph interaction fusion module to perform message passing and feature aggregation on the three graphs to generate vehicle representations. Then, the vehicle representations are input into the trajectory decoding module to obtain the future trajectory of the target vehicle and its neighboring vehicles. The trajectory prediction sequence at each time step is combined with the prediction loss function and the contrastive learning loss function, and the parameters of the collaborative prediction model of the target vehicle are jointly updated through the backpropagation algorithm. Then, the participating vehicles upload the locally trained collaborative prediction model parameters to the central server. The central server performs weighted aggregation based on the contribution of each participating vehicle to generate a global collaborative prediction model, and distributes the global collaborative prediction model parameters to each participating vehicle to replace their local collaborative prediction model parameters. S1–S5 are repeated until the global collaborative prediction model converges to obtain the final collaborative prediction model. Finally, in the model deployment phase, the target vehicle obtains its real-time motion state through onboard sensors to form a trajectory sequence. The trajectory encoding module of the collaborative prediction model is used to extract semantic feature vectors, and then the semantic feature vectors of the trajectories of neighboring vehicles are obtained. The vehicle representation is obtained through the graph interaction fusion module, and finally the trajectory decoding module outputs the trajectory prediction sequence of the target vehicle and neighboring vehicles.

[0097] Furthermore, the Z-score normalization method is used to normalize the various dimensions of the historical trajectory data, including:

[0098] The location is divided into two dimensions: longitude and latitude. Velocity, acceleration, and heading angle are each represented as one dimension. These five dimensions are then processed and expressed as follows: Its normalization process is as follows:

[0099] ,

[0100] in, For historical trajectory data feature dimensions Standardized eigenvalues For historical trajectory data feature dimensions The original eigenvalues, For historical trajectory data feature dimensions The mean, For historical trajectory data feature dimensions standard deviation This represents the feature dimension of historical trajectory data. .

[0101] Furthermore, the contrastive learning task includes:

[0102] Using a trajectory sequence of the target vehicle as an anchor sample, positive samples of the anchor sample are generated through time clipping. Multiple trajectory sequences of the target vehicle are combined into a training batch, and the other trajectory sequences in the batch other than the anchor sample are used as negative samples of the anchor sample.

[0103] By minimizing the contrastive learning loss function, the training trajectory encoding module learns semantic feature vectors that discriminate driving intentions.

[0104] Furthermore, the time trimming includes:

[0105] From anchor sample trajectory sequence In the process, a start time step is randomly selected. and end time step ,satisfy and Extracting subsequences As a clipped trajectory segment The length of the clipped trajectory segment;

[0106] For each time step , The length of the clipped trajectory segment is restored to a fixed time window through a filling mechanism. :

[0107] for Preserve the original data.

[0108] for For frames missing at the beginning, a forward expansion strategy is used to copy the starting frame. Fill in,

[0109] for Missing frames at the end, copy the termination frame. Fill in the blanks.

[0110] Furthermore, the contrastive learning loss function includes:

[0111] The cosine similarity between the anchor sample and its positive samples is taken as the positive term, and the cosine similarity between the anchor sample and multiple negative samples is taken as the negative term.

[0112] The positive and negative terms were respectively subjected to a test with the natural constant as the base and temperature parameter. Weighted by the exponent of the denominator;

[0113] By normalizing all weighting terms using a normalized exponential function, the relative probabilities of anchor samples and positive samples are obtained.

[0114] The trajectory encoding module is trained using the negative logarithm that minimizes the relative probability as the contrastive learning loss function;

[0115] Wherein, the temperature parameter satisfy ;

[0116] The specific calculation of the contrastive learning loss function is as follows:

[0117] ,

[0118] in, For the loss function of contrastive learning, For anchor sample semantic feature vectors, For positive sample semantic feature vectors, For negative sample semantic feature vectors, Let cosine similarity function be defined as follows: , For temperature parameters, .

[0119] Furthermore, the trajectory encoding module includes:

[0120] The trajectory encoding module includes an encoder consisting of two stacked Long Short-Term Memory (LSTM) networks, which processes the anchor sample trajectory sequences. With positive sample trajectory sequence , For a fixed time window, The number of feature dimensions of the trajectory sequence. ;

[0121] For each input sequence, obtain the hidden state of the Long Short-Term Memory network at each time step, forming a time window of fixed length. The sequence of hidden vectors;

[0122] The hidden state sequence is flattened into a one-dimensional vector along the time dimension, and then mapped to a fixed dimension through a fully connected layer. The semantic feature vector, where, The preset semantic feature vector length.

[0123] Furthermore, the target vehicle communicates with neighboring vehicles, including:

[0124] The target vehicle uses its position at the end of its anchor sample trajectory sequence as a reference point to calculate the Euclidean distance between it and surrounding vehicles at the same time. Based on this Euclidean distance, within its set maximum sensing and communication range... Internal filtering selects those with a distance less than The candidate neighbor vehicles are selected, and the K nearest vehicles are selected as neighbor vehicles. If the number of candidate neighbor vehicles is less than K, the zero vector with the same dimension as the semantic feature vector is used to replace the feature input of the missing vehicle.

[0125] Furthermore, the three graphs defined by distance and semantic feature vector similarity include:

[0126] The first graph is a connectivity graph based on spatial proximity. Calculate the reciprocal of the Euclidean distance between vehicles as the initial weights, and normalize the initial weights for each vehicle. If the normalized weights exceed a preset threshold... If , then a connection is established between the two vehicles, indicating a spatial connectivity relationship;

[0127] The second graph is a connection graph based on semantic behavior similarity. Calculate the cosine similarity of semantic feature vectors between vehicles as the initial weights, and then normalize them. If the normalized weights exceed a preset threshold... If , then a connection is established, representing a semantic connectivity relationship;

[0128] The third figure is a comprehensive connectivity graph based on spatial-semantic fusion. The normalized weights corresponding to the first two graphs are weighted and summed according to a preset ratio to obtain the fusion weight, and then the fusion weight is normalized again; if the fusion weight exceeds a preset threshold... If a connection is established, it represents a comprehensive interactive connectivity relationship.

[0129] Furthermore, the multi-relationship graph attention mechanism in the graph interaction fusion module includes:

[0130] For the connection graph based on spatial proximity Connection graph based on semantic behavior similarity and a comprehensive connectivity graph based on spatial-semantic fusion Each of them constructs an independent graph attention layer;

[0131] In each graph attention layer, for any vehicle node in the graph Its semantic feature vector After a learnable linear transformation, it becomes the query vector, and all neighboring nodes connected to it in the relational graph are included. semantic feature vector After undergoing learnable linear transformations, the vectors are used as key and value vectors, respectively. The attention weights are then calculated as follows:

[0132] ,

[0133] in, Indicates the type of relationship diagram. In order to be in Vehicle nodes in the relationship diagram Its neighboring nodes Attention weights Vehicle nodes and neighboring nodes semantic feature vectors, for The learnable projection matrix corresponding to the relationship graph. for vector dimension, Along the neighbor node dimension Normalize, In order to be in Vehicle nodes in the relationship diagram The set of connected neighboring nodes;

[0134] Based on the calculated attention weights, the semantic feature vectors of neighboring nodes are weighted and aggregated to obtain the th... Vehicle nodes under each relationship Contextual representation:

[0135] ,

[0136] in, In order to be in Vehicle nodes in the relationship diagram Contextual representation, In relation Vehicle nodes in the diagram Its neighboring nodes Attention weights In order to be in Vehicle nodes in the relationship diagram The set of connected neighbor nodes, The projection matrix of the value vector. Neighboring nodes semantic feature vectors;

[0137] Context representation obtained from the three relation diagrams The data is stitched together, and vehicle nodes are generated through an output projection layer. Vehicle characterization :

[0138] ,

[0139] in, For vehicle nodes Vehicle characterization, This represents vector concatenation. To output the projection matrix, The context representations are obtained under the three relational graphs.

[0140] Furthermore, the trajectory decoding module includes:

[0141] The trajectory decoding module includes a feedforward neural network composed of multiple fully connected layers. It takes the vehicle representation of each vehicle node output in step S3 as input, processes it layer by layer through linear transformation and the ReLU activation function, and finally outputs a sequence containing... A prediction vector of 1 element;

[0142] In the prediction vector, each pair of consecutive elements represents the future nth... The predicted latitude and longitude coordinates at each time step, arranged in chronological order, constitute the future coordinates of the target vehicle. A trajectory sequence at each time step.

[0143] Furthermore, the combination of the prediction loss function and the contrastive learning loss function includes:

[0144] Define the prediction loss function Used to measure the model's predicted trajectory With the actual future trajectory The difference between them is used as the average of the sum of squared Euclidean distances at each time step as the loss:

[0145] ,

[0146] in, For predicting the loss function, The time step length for predicting the trajectory, For the true trajectory in the 1st The position coordinates of each time step. To predict the trajectory in the 1st The position coordinates of each time step. The square of the Euclidean distance between the predicted point and the true point;

[0147] The prediction loss function and the contrastive learning loss function are weighted and summed to construct a comprehensive loss function:

[0148] ,

[0149] in, These are adjustable hyperparameters used to balance trajectory prediction accuracy and feature representation quality. For the comprehensive loss function, For predicting the loss function, To learn the loss function by comparison;

[0150] Based on the aforementioned comprehensive loss function The gradient is calculated using the backpropagation algorithm, and the Adam optimizer is used to jointly update all trainable parameters of the trajectory encoding module, graph interaction fusion module, and trajectory decoding module.

[0151] Furthermore, the weighted aggregation based on the contribution of each participating vehicle includes:

[0152] The central server uses the amount of local historical trajectory data used by each participating vehicle in this round of training as a measure of its contribution and calculates the aggregation weight.

[0153] For the vehicle ensemble participating in this round of training ,vehicle Aggregate weights Based on the amount of its local historical trajectory data The ratio to the sum of the total data for all participating vehicles is determined as follows:

[0154] ,

[0155] in, Indicates vehicle The number of trajectory sequences used for updating the global collaborative prediction model in this round of training. Indicates vehicle The number of trajectory sequences used for updating the global collaborative prediction model in this round of training. For vehicles The aggregate weights used for updating the global collaborative prediction model in this round of training. For the assembly of vehicles participating in this round of training;

[0156] The central server is based on the calculated weights For all participating vehicles, upload local collaborative prediction model parameters Perform a weighted average to generate new global collaborative prediction model parameters. :

[0157] ,

[0158] in, To generate new global collaborative prediction model parameters, For vehicles The aggregate weights used for updating the global collaborative prediction model in this round of training. For vehicles Uploaded local collaborative prediction model parameters, The vehicles participating in this round of training are assembled.

[0159] Furthermore, the model deployment phase includes the following steps:

[0160] Each vehicle continuously collects data from its onboard sensors in real time. Motion data at each time step (including position, velocity, acceleration, and heading angle) constitute a short-term historical trajectory sequence, which is then normalized using the same Z-score normalization method as in step S1.

[0161] The normalized short-time historical trajectory sequence is input into the trajectory encoding module of the final collaborative prediction model, which is deployed locally and trained through federated learning, to obtain the semantic feature vector of the vehicle. ;

[0162] This vehicle broadcasts its own semantic feature vector via vehicle-to-vehicle (V2V) communication. And simultaneously receive from the surroundings Semantic feature vectors of neighboring vehicle broadcasts ;

[0163] Based on the semantic feature vectors of this vehicle and neighboring vehicles, and following the method defined in S3, three relationship graphs are constructed based on spatial distance, semantic similarity, and fusion weight, and a multi-relationship graph structure is constructed locally.

[0164] The constructed multi-relationship graph and the semantic feature vectors of all vehicles are input into the graph interaction fusion module of the local model. The multi-relationship graph attention mechanism is used for message passing and feature aggregation to generate vehicle representations containing information about the surrounding traffic environment.

[0165] The vehicle representation is input into the trajectory decoding module of the local model, which then directly outputs the future trajectory of the vehicle and its neighboring vehicles. A sequence of predicted trajectories at each time step.

[0166] Figure 2 This invention provides an architecture for a multi-vehicle cooperative trajectory prediction method based on federated learning, as shown in the figure. The system includes multiple participating vehicle clients and a central server. Each participating vehicle deploys a cooperative prediction model locally, which consists of three core modules:

[0167] The trajectory encoding module is used to encode the vehicle's own historical trajectory sequence into a semantic feature vector with driving intention discrimination.

[0168] Graph interaction fusion module: It acquires the semantic features of surrounding vehicles through vehicle-to-vehicle communication, constructs a multi-relationship graph based on spatial distance and semantic similarity, and uses graph attention mechanism to fuse contextual information;

[0169] Trajectory decoding module: Based on the fused vehicle representation, decode and generate trajectory prediction results for the next T time steps.

[0170] During the federated learning training process, each vehicle completes model training locally and uploads the complete parameters of its local model to the central server. The server performs weighted aggregation on all uploaded model parameters to generate a global model, and distributes the updated global model to each participating vehicle for the next round of training. Through multiple iterations, a converged collaborative prediction model is finally obtained, achieving privacy protection and knowledge sharing among multiple vehicles.

[0171] Figure 3 This is a block diagram of a multi-vehicle cooperative trajectory prediction system based on federated learning provided in an embodiment of the present invention. The multi-vehicle cooperative trajectory prediction system based on federated learning described below and the multi-vehicle cooperative trajectory prediction method based on federated learning described above can be referred to and correspond to each other.

[0172] Data acquisition and preprocessing module: This module collects historical trajectory data of the vehicle itself and normalizes it using the Z-score standardization method, dividing it into fixed time windows. The trajectory sequence;

[0173] Trajectory encoding calculation module: used to perform a contrastive learning task locally, encoding the input trajectory sequence into a semantic feature vector with driving intent discriminative capabilities;

[0174] The graph interaction fusion computing module is used to obtain semantic feature vectors of surrounding vehicles through vehicle-to-vehicle communication, construct a multi-relationship graph based on spatial distance and semantic similarity, and use the multi-relationship graph attention mechanism for message passing and feature aggregation to generate vehicle representations.

[0175] Trajectory decoding calculation module: used to decode the fused vehicle representation into the future values ​​of the target vehicle and its neighboring vehicles. The position sequence of each time step;

[0176] Model aggregation and update module: used to upload the parameters of the locally trained collaborative prediction model to the central server, and receive the global collaborative prediction model parameters issued by the central server to replace the local model for the next round of training;

[0177] The central server is independently configured to receive collaborative prediction model parameters uploaded by each participating vehicle, calculate the aggregation weight based on the local data volume of each vehicle, perform a weighted average of the collaborative prediction model parameters, generate updated global collaborative prediction model parameters, and distribute them to each participating vehicle.

[0178] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0179] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-vehicle cooperative trajectory prediction method based on federated learning, characterized in that, The steps are as follows: S1: Collect and store historical trajectory data of the target vehicle, including position, speed, acceleration, heading angle, and timestamp information. Use the Z-score normalization method to normalize each dimension of the historical trajectory data, and divide the normalized historical trajectory data into fixed time windows. The trajectory sequence is obtained, and each target vehicle maintains a collaborative prediction model locally, which includes a trajectory encoding module, a graph interaction fusion module, and a trajectory decoding module; S2: Design a contrastive learning task, input the trajectory sequence into the trajectory encoding module to obtain the semantic feature vector of the trajectory, and train the trajectory encoding module by minimizing the contrastive learning loss function; S3: The target vehicle communicates with neighboring vehicles to obtain the semantic feature vectors of the neighboring vehicles' trajectories, calculates the distance between each vehicle, defines three graphs based on the distance and semantic feature vector similarity, and combines the multi-relationship graph attention mechanism in the graph interaction fusion module to perform message passing and feature aggregation on the three graphs to generate vehicle representations. S4: Input the vehicle representation into the trajectory decoding module to obtain the future trajectory of the target vehicle and its neighboring vehicles. The trajectory prediction sequence at each time step is combined with the prediction loss function and the contrastive learning loss function, and the parameters of the collaborative prediction model of the target vehicle are jointly updated through the backpropagation algorithm. S5: Participating vehicles upload the locally trained collaborative prediction model parameters to the central server. The central server performs weighted aggregation based on the contribution of each participating vehicle to generate a global collaborative prediction model. The global collaborative prediction model parameters are then distributed to each participating vehicle to replace its local collaborative prediction model parameters. S6: Repeat S1–S5 until the global collaborative prediction model converges to obtain the final collaborative prediction model; S7: During the model deployment phase, the target vehicle obtains its real-time motion state through onboard sensors, forming a trajectory sequence. The trajectory encoding module of the collaborative prediction model is used to extract semantic feature vectors, and then the semantic feature vectors of the trajectories of neighboring vehicles are obtained. The vehicle representation is obtained through the graph interaction fusion module, and finally the trajectory decoding module outputs the trajectory prediction sequence of the target vehicle and neighboring vehicles. The contrastive learning task includes: Using a trajectory sequence of the target vehicle as an anchor sample, positive samples of the anchor sample are generated through time clipping. Multiple trajectory sequences of the target vehicle are combined into a training batch, and the other trajectory sequences in the batch other than the anchor sample are used as negative samples of the anchor sample. By minimizing the contrastive learning loss function, the training trajectory encoding module learns semantic feature vectors that discriminate driving intentions.

2. The multi-vehicle cooperative trajectory prediction method based on federated learning as described in claim 1, characterized in that, The Z-score normalization method is used to normalize the various dimensions of historical trajectory data, including: The location is divided into two dimensions: longitude and latitude. Velocity, acceleration, and heading angle are each represented as one dimension. These five dimensions are then processed and expressed as follows: Its normalization process is as follows: ， in, For historical trajectory data feature dimensions Standardized eigenvalues For historical trajectory data feature dimensions The original eigenvalues, For historical trajectory data feature dimensions The mean, For historical trajectory data feature dimensions standard deviation This represents the feature dimension of historical trajectory data. .

3. The multi-vehicle cooperative trajectory prediction method based on federated learning as described in claim 1, characterized in that, The time clipping includes: From anchor sample trajectory sequence In the process, a start time step is randomly selected. and end time step ,satisfy and Extracting subsequences As a clipped trajectory segment The length of the clipped trajectory segment; For each time step , The length of the clipped trajectory segment is restored to a fixed time window through a filling mechanism. : for Preserve the original data. for For frames missing at the beginning, a forward expansion strategy is used to copy the starting frame. Fill in, for Missing frames at the end, copy the termination frame. Fill in the blanks.

4. The multi-vehicle cooperative trajectory prediction method based on federated learning as described in claim 1, characterized in that, The contrastive learning loss function includes: The cosine similarity between the anchor sample and its positive samples is taken as the positive term, and the cosine similarity between the anchor sample and multiple negative samples is taken as the negative term. The positive and negative terms were respectively subjected to a test with the natural constant as the base and temperature parameter. Weighted by the exponent of the denominator; By normalizing all weighting terms using a normalized exponential function, the relative probabilities of anchor samples and positive samples are obtained. The trajectory encoding module is trained using the negative logarithm that minimizes the relative probability as the contrastive learning loss function; Wherein, the temperature parameter satisfy .

5. The multi-vehicle cooperative trajectory prediction method based on federated learning as described in claim 1, characterized in that, The trajectory encoding module includes: The trajectory encoding module includes an encoder consisting of two stacked long short-term memory networks, which processes anchor sample trajectory sequences. With positive sample trajectory sequence , For a fixed time window, The number of feature dimensions of the trajectory sequence. ; For each input sequence, obtain the hidden state of the Long Short-Term Memory network at each time step, forming a time window of fixed length. The sequence of hidden vectors; The sequence of hidden vectors is flattened into a one-dimensional vector along the time dimension, and then mapped to a fixed dimension through a fully connected layer. The semantic feature vector, where, The preset semantic feature vector length.

6. The multi-vehicle cooperative trajectory prediction method based on federated learning as described in claim 1, characterized in that, The target vehicle communicates with neighboring vehicles, including: The target vehicle uses its position at the end of its anchor sample trajectory sequence as a reference point to calculate the Euclidean distance between it and surrounding vehicles at the same time. Based on this Euclidean distance, within its set maximum sensing and communication range... Internal filtering selects those with a distance less than The candidate neighbor vehicles are selected, and the K nearest vehicles are selected as neighbor vehicles. If the number of candidate neighbor vehicles is less than K, the zero vector with the same dimension as the semantic feature vector is used to replace the feature input of the missing vehicle.

7. The multi-vehicle cooperative trajectory prediction method based on federated learning as described in claim 1, characterized in that, The three graphs defined by distance and semantic feature vector similarity include: The first graph is a connectivity graph based on spatial proximity. Calculate the reciprocal of the Euclidean distance between vehicles as the initial weights, and normalize the initial weights for each vehicle. If the normalized weights exceed a preset threshold... If , then a connection is established between the two vehicles, indicating a spatial connectivity relationship; The second graph is a connection graph based on semantic behavior similarity. Calculate the cosine similarity of semantic feature vectors between vehicles as the initial weights, and then normalize them. If the normalized weights exceed a preset threshold... If , then a connection is established, representing a semantic connectivity relationship; The third figure is a comprehensive connectivity graph based on spatial-semantic fusion. The normalized weights corresponding to the first two graphs are weighted and summed according to a preset ratio to obtain the fusion weight, and then the fusion weight is normalized again; if the fusion weight exceeds a preset threshold... If a connection is established, it represents a comprehensive interactive connectivity relationship.

8. The multi-vehicle cooperative trajectory prediction method based on federated learning as described in claim 7, characterized in that, The multi-relationship graph attention mechanism in the graph interaction fusion module includes: For the connection graph based on spatial proximity Connection graph based on semantic behavior similarity and a comprehensive connectivity graph based on spatial-semantic fusion Each of them constructs an independent graph attention layer; In each graph attention layer, for any vehicle node in the graph Its semantic feature vector After a learnable linear transformation, it becomes the query vector, and all neighboring nodes connected to it in the relational graph are included. semantic feature vector After undergoing learnable linear transformations, the vectors are used as key and value vectors, respectively. The attention weights are then calculated as follows: ， in, Indicates the type of relationship diagram. In order to be in Vehicle nodes in the relationship diagram Its neighboring nodes Attention weights Vehicle nodes and neighboring nodes semantic feature vectors, for The learnable projection matrix corresponding to the relationship graph. for vector dimension, Along the neighbor node dimension Normalize, In order to be in Vehicle nodes in the relationship diagram The set of connected neighboring nodes; Based on the calculated attention weights, the semantic feature vectors of neighboring nodes are weighted and aggregated to obtain the th... Vehicle nodes under each relationship Contextual representation: ， in, In order to be in Vehicle nodes in the relationship diagram Contextual representation, In relation Vehicle nodes in the diagram Its neighboring nodes Attention weights In order to be in Vehicle nodes in the relationship diagram The set of connected neighbor nodes, The projection matrix of the value vector. Neighboring nodes semantic feature vectors; Context representation obtained from the three relation diagrams The data is stitched together, and vehicle nodes are generated through an output projection layer. Vehicle characterization : ， in, For vehicle nodes Vehicle characterization, This represents vector concatenation. To output the projection matrix, The context representations are obtained under the three relational graphs.

9. The multi-vehicle cooperative trajectory prediction method based on federated learning as described in claim 1, characterized in that, The trajectory decoding module includes: The trajectory decoding module includes a feedforward neural network composed of multiple fully connected layers. It takes the vehicle representation of each vehicle node output in step S3 as input, processes it layer by layer through linear transformation and the ReLU activation function, and finally outputs a sequence containing... A prediction vector of 1 element; In the prediction vector, each pair of consecutive elements represents the future nth... The predicted latitude and longitude coordinates at each time step, arranged in chronological order, constitute the future coordinates of the target vehicle. A trajectory sequence at each time step.

10. The multi-vehicle cooperative trajectory prediction method based on federated learning as described in claim 1, characterized in that, The combination of the prediction loss function and the contrastive learning loss function includes: Define the prediction loss function Used to measure the model's predicted trajectory With the actual future trajectory The difference between them is used as the average of the sum of squared Euclidean distances at each time step as the loss: ， in, For predicting the loss function, The time step length for predicting the trajectory, For the true trajectory in the 1st The position coordinates of each time step. To predict the trajectory in the 1st The position coordinates of each time step. The square of the Euclidean distance between the predicted point and the true point; The prediction loss function and the contrastive learning loss function are weighted and summed to construct a comprehensive loss function: ， in, These are adjustable hyperparameters used to balance trajectory prediction accuracy and feature representation quality. For the comprehensive loss function, For predicting the loss function, To learn the loss function by comparison; Based on the aforementioned comprehensive loss function The gradient is calculated using the backpropagation algorithm, and the Adam optimizer is used to jointly update all trainable parameters of the trajectory encoding module, graph interaction fusion module, and trajectory decoding module.

11. The multi-vehicle cooperative trajectory prediction method based on federated learning as described in claim 1, characterized in that, The weighted aggregation based on the contribution of each participating vehicle includes: The central server uses the amount of local historical trajectory data used by each participating vehicle in this round of training as a measure of its contribution and calculates the aggregation weight. For the vehicle ensemble participating in this round of training ,vehicle Aggregate weights Based on the amount of its local historical trajectory data The ratio to the sum of the total data for all participating vehicles is determined as follows: ， in, Indicates vehicle The number of trajectory sequences used for updating the global collaborative prediction model in this round of training. Indicates vehicle The number of trajectory sequences used for updating the global collaborative prediction model in this round of training. For vehicles The aggregate weights used for updating the global collaborative prediction model in this round of training. For the assembly of vehicles participating in this round of training; The central server is based on the calculated aggregate weight For all participating vehicles, upload local collaborative prediction model parameters Perform a weighted average to generate new global collaborative prediction model parameters. : ， in, To generate new global collaborative prediction model parameters, For vehicles The aggregate weights used for updating the global collaborative prediction model in this round of training. For vehicles Uploaded local collaborative prediction model parameters, The vehicles participating in this round of training are assembled.

12. A multi-vehicle cooperative trajectory prediction system based on federated learning, characterized in that, Implementing a multi-vehicle cooperative trajectory prediction method based on federated learning as described in any one of claims 1-11, comprising: Data acquisition and preprocessing module: This module collects historical trajectory data of the vehicle itself and normalizes it using the Z-score standardization method, dividing it into fixed time windows. The trajectory sequence; Trajectory encoding calculation module: used to perform a contrastive learning task locally, encoding the input trajectory sequence into a semantic feature vector with driving intent discriminative capabilities; The graph interaction fusion computing module is used to obtain semantic feature vectors of surrounding vehicles through vehicle-to-vehicle communication, construct a multi-relationship graph based on spatial distance and semantic similarity, and use the multi-relationship graph attention mechanism for message passing and feature aggregation to generate vehicle representations. Trajectory decoding calculation module: used to decode the fused vehicle representation into the future values ​​of the target vehicle and its neighboring vehicles. The position sequence of each time step; Model aggregation and update module: used to upload the parameters of the locally trained collaborative prediction model to the central server, and receive the global collaborative prediction model parameters issued by the central server to replace the local model for the next round of training; The central server is independently configured to receive collaborative prediction model parameters uploaded by each participating vehicle, calculate the aggregation weight based on the local data volume of each vehicle, perform a weighted average of the collaborative prediction model parameters, generate updated global collaborative prediction model parameters, and distribute them to each participating vehicle.

Images