A vehicle trajectory prediction method based on a global space-time graph network

Abstract

This invention discloses a vehicle trajectory prediction method based on a global spatiotemporal graph network. The prediction method includes: constructing a road vehicle spatial topology network based on the distance relationships between road vehicle trajectory data and adding a jump coefficient; constructing a temporal logic network for road vehicles themselves using complex network methods; and constructing a global spatiotemporal graph matrix by combining the road vehicle spatial topology network; using the obtained global spatiotemporal graph matrix and historical vehicle trajectory data as input to a graph operation module to generate graph features; and encoding and decoding using a Seq2Seq network to obtain the position information of each future time step, thereby realizing the prediction of road vehicle trajectory data. This method can fully extract the hidden features of the graph features and predict the position information of future steps, effectively improving the accuracy of road vehicle trajectory prediction.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation, and in particular to a vehicle trajectory prediction method based on a global spatiotemporal graph network. Background Technology

[0002] As autonomous driving technology continues to advance, people increasingly expect and rely on progress in this area, as safe and reliable systems can help autonomous vehicles avoid many traffic accidents. However, in reality, interference from neighboring vehicles and complex road conditions make safety a challenging issue. Therefore, for the safety of autonomous vehicles, vehicle trajectory prediction plays a crucial role in ensuring the vehicle's accurate trajectory.

[0003] Current vehicle trajectory prediction methods mainly include: feedback neural network method, long short-term memory network method, Kalman filter algorithm, etc.; however, the above methods have difficulty obtaining information about surrounding neighboring vehicles. Summary of the Invention

[0004] In view of the above problems, the present invention is proposed to provide a vehicle trajectory prediction method based on a global spatiotemporal graph network that overcomes or at least partially solves the above problems.

[0005] According to one aspect of the present invention, a vehicle trajectory prediction method based on a global spatiotemporal graph network is provided, the prediction method comprising: Based on the distance relationships between road vehicle trajectory data and by incorporating a jump coefficient, a road vehicle spatial topology network is constructed. Using complex network methods, a temporal logic network of road vehicles was constructed; and combined with the spatial topology network of road vehicles, a global spatiotemporal graph matrix was constructed. The obtained global spatiotemporal graph matrix and historical vehicle trajectory data are used as inputs to the graph operation module to generate graph features; and the Seq2Seq network is used for encoding and decoding to obtain the location information of each future time step, thereby realizing the prediction of road vehicle trajectory data.

[0006] Optionally, the step of constructing a road vehicle spatial topology network based on the distance relationships between road vehicle trajectory data and by adding a jump coefficient specifically includes: Construct a road vehicle network based on a road vehicle network and a threshold-based road vehicle adjacency matrix; Add a step coefficient, reconstruct the adjacency matrix, and construct the spatial topology matrix.

[0007] Optionally, the construction of the road vehicle network and the threshold-based road vehicle adjacency matrix specifically includes: Record the historical road vehicle network as G= ( Jt , B t ),in J t ={ j t1 , j t2 , …, j tN} is the first t Vehicles gather on the road at any time. N This refers to the number of vehicles on the road. B t It is a set of vehicle connection relationships. t The selected historical time step T The first in t time, t ∈ T ; Built on t Time based on threshold k Road vehicle adjacency matrix ,in k For the selected distance threshold, the vehicle j tn With vehicles j tm distance H The calculation formula is as follows: in j tn =( x tn , y tn ), j tm =( x tm , y tm ), x tn for t time n The horizontal axis position coordinates of car number 1 y tn for t time n The longitudinal axis position coordinates of car number 1 x tm and y tm Similarly; vehicle j n With vehicles j tm The connection relationship is formed as follows: .

[0008] Optionally, the step step addition, reconstructing the adjacency matrix, and constructing the spatial topology matrix specifically includes: Based on the constructed vehicle adjacency matrix Add a skip coefficient h The reconstructed vehicle adjacency matrix set is obtained , in for of p power ( p ∈0,1,…, h This indicates the indirect distance relationship between vehicles. p Indicates the degree of indirect relationship. The elements in the matrix are represented as And create a mask matrix for all matrices in the matrix set. The elements in the mask matrix are represented as The calculation method is as follows: Ultimately, for Search all mask matrices in the system for vehicles. v tn With vehicles v tm If the connection between the two edges exists once, then The reconstructed adjacency matrix is ​​obtained. and spatial topological matrix set ={ , ,…, }

[0009] Optionally, the construction of a temporal logic network for road vehicles using complex network methods, and the construction of a global spatiotemporal graph matrix by combining the road vehicle spatial topology network, specifically includes: Using the finite traversal visibility method, the correlation between road vehicles at each moment is analyzed, and the temporal logic network of the vehicles is constructed. A global spatiotemporal graph matrix is ​​constructed by combining the spatial topology network of road vehicles.

[0010] Optionally, the method of using finite traversal visibility to analyze the correlation between road vehicles at each moment and construct the vehicle's own temporal logic network specifically includes: To represent the time-dependent nature of a vehicle, the first... n The historical trajectory data of a vehicle is defined as follows: V (n) =[ v 1 (n) , v 2 (n) ,…, v T (n) ], At the sampling point i and j Choose any sampling point between c Perform the calculation: in v i (n) , v j (n) and v c (n) They represent the first n Vehicle at the sampling point i , j and c The trajectory information data. If it satisfies Then it is considered v i (n) and v j (n) The line of sight can penetrate once; right i and j Perform the above calculation on all sampling points between them, and define: in L To accumulate the number of penetrations, The set penetration threshold; When two sampling points are visible, an edge connection is established to obtain the time logic network of a single vehicle. Q (n) =[ V (n) , E (n) ]( n ∈{1,…, N}), N This represents the total number of vehicles. E It refers to the connection relationship between different sampling points; Repeat the above calculations to obtain the time logic network for all vehicles: And obtain the time logical adjacency matrix of each vehicle. For the first n vehicle time point i With time nodes j : .

[0011] Optionally, the construction of the global spatiotemporal graph matrix by combining the road vehicle spatial topology network specifically includes: By combining temporal logic networks and spatial topology networks, a global spatiotemporal graph matrix of the vehicle group is obtained. The first part is: The diagonal of the vehicle t The spatial topological adjacency matrix at time t; the second part is The rest; Based on time logic network From the vehicle's own temporal logical adjacency matrix To obtain the temporal connection relationship of each vehicle, we obtain the temporal logical adjacency matrix { A″ 11 , A″ 12 ,…, A″ 21, A″ 22 ,…, A″ ( T-1}T , A ″ TT}, and splice them together to form The rest; A″ ij It is a diagonal matrix representing the sampling time. i and sampling time j The temporal logical adjacency matrix of the vehicles is represented as follows: .

[0012] Optionally, the step of using the obtained global spatiotemporal graph matrix and historical vehicle trajectory data as input to the graph operation module to generate graph features; and then encoding and decoding them through a Seq2Seq network to obtain the location information for each future time step, thereby realizing the prediction of road vehicle trajectory data, specifically includes: Construct a graph operation module to obtain graph features based on global spatiotemporal data; Construct a Seq2Seq network to encode graph features and generate future trajectory data based on hidden features; Define the loss function for the global spatiotemporal graph network model: Future trajectory data obtained from a global spatiotemporal graph network With real future trajectory data The loss function for the global spatiotemporal graph network model is defined as follows: .

[0013] Optionally, the graph construction module, in obtaining graph features based on global spatiotemporal data, specifically includes: The obtained global vehicle spatiotemporal topology matrix Compared with historical vehicle trajectory data X As input to the graph manipulation module; in X ={ X 1, X 2, …, X T},in X t This represents the road vehicle trajectory data at time t. t ∈0, 1, …, T ),and ,in x tn and y tn Indicates the first t time n The location coordinates of vehicle number [number]; Stacked graph operations are used to extract global spatiotemporal features. The graph convolution operation is described as follows: in, yes The degree matrix, H For graph operations input, W This represents the weight matrix for graph convolution operations. l The number of layers for stacking graph operations ultimately yields graph features based on global spatiotemporal data. .

[0014] Optionally, the construction of the Seq2Seq network to encode graph features and generate future trajectory data based on hidden features specifically includes: The Seq2Seq network consists of two GRU networks, acting as an encoder and a decoder, respectively. The encoder encodes the graph features and extracts the hidden features from the GRU network. The calculation method is as follows: GRU stands for Gated Cyclic Unit Network. output This is the output of the last hidden layer. hidden This is the hidden state of the last time step, which is the hidden feature extracted by the encoder from the graph features obtained by the graph operation module. Based on the hidden features extracted by the GRU encoder, the GRU decoder predicts the trajectory data by analyzing the output position data at future time steps. The GRU decoding method is as follows: In the first decoding step, the input b t Position data of the last step in the historical time step X t , h t-1 Graph hidden features extracted by the encoder hidden , z t and r t For output control updates and resetting of the gating state, σ is the sigmoid activation function. W z , W r It is the weight matrix for updating and resetting the gating calculation steps; pass h t-1 and r t Perform matrix multiplication and combine with b t After concatenation, the output is then activated by the tanh function. , Including current input b t information, W s To generate The weight matrix in the steps; Indicates the use of updated gating z t Selectively forget the hidden features of the current input. This indicates that the input includes the current input. b t Information Selective memorization yields hidden states. h t The hidden features are used as input for the next decoding step and output as the vehicle position information at the next time step. By repeatedly decoding, the position information of all future time steps is finally obtained, forming future trajectory data. ,in c The maximum step size for the future time to be predicted.

[0015] This invention provides a vehicle trajectory prediction method based on a global spatiotemporal graph network. The prediction method includes: constructing a road vehicle spatial topology network based on the distance relationships between road vehicle trajectory data and adding a jump coefficient; constructing a temporal logical network for each road vehicle using complex network methods; and constructing a global spatiotemporal graph matrix by combining the road vehicle spatial topology network; using the obtained global spatiotemporal graph matrix and historical vehicle trajectory data as input to a graph operation module to generate graph features; and encoding and decoding these features using a Seq2Seq network to obtain the position information for each future time step, thereby achieving the prediction of road vehicle trajectory data. This method can fully extract the hidden features of the graph features and predict the position information for future steps, effectively improving the accuracy of road vehicle trajectory prediction.

[0016] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a vehicle trajectory prediction method based on a global spatiotemporal graph network, provided as an embodiment of the present invention; Figure 2 This is a structural diagram of the global spatiotemporal graph network model provided in an embodiment of the present invention; Figure 3 The image shows the prediction results of the vehicle trajectory prediction model based on a global spatiotemporal graph network provided in this embodiment of the invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0020] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0022] like Figure 1 As shown, a vehicle trajectory prediction method based on a global spatiotemporal graph network is disclosed. The prediction method includes: Based on the distance relationships between road vehicle trajectory data and by incorporating a jump coefficient, a road vehicle spatial topology network is constructed. Using complex network methods, a temporal logic network of road vehicles was constructed; and combined with the spatial topology network of road vehicles, a global spatiotemporal graph matrix was constructed. The obtained global spatiotemporal graph matrix and historical vehicle trajectory data are used as inputs to the graph operation module to generate graph features; and the Seq2Seq network is used for encoding and decoding to obtain the location information of each future time step, thereby realizing the prediction of road vehicle trajectory data.

[0023] A vehicle trajectory prediction method based on a global spatiotemporal graph network includes the following steps: The steps for constructing the road vehicle spatial topology network are based on real-time collected road vehicle location information, calculating the spatial distance between each neighboring vehicle; establishing adjacency relationships between vehicle nodes according to a preset distance threshold, and introducing a jump expansion coefficient to broaden the coverage of adjacency relationships, thus completing the construction of the vehicle spatial topology matrix at the current moment.

[0024] The steps for constructing the global spatiotemporal graph matrix employ the finite traversal visibility algorithm to model the motion dependencies between time-series nodes of a single vehicle, generating the vehicle's own temporal logic network. The aforementioned spatial topology network and temporal logic network are then fused to generate the global spatiotemporal graph matrix, enabling unified extraction and representation of the spatiotemporal features of historical trajectories across all time periods.

[0025] The graph operation module and Seq2Seq network are constructed to complete the trajectory prediction step. The historical trajectory data of the target vehicle and the global spatiotemporal graph matrix are input into the graph operation module, and the global spatiotemporal graph features are extracted through graph convolution operation. A Seq2Seq network based on GRU structure is built, in which the encoder encodes the graph features and outputs the hidden layer features. The encoder hidden features and the vehicle coordinates of the previous time step are input into the decoder GRU to infer the current position time step by time. The future full-time trajectory sequence of the target vehicle is output through cyclic decoding.

[0026] Example 1: A vehicle trajectory prediction method based on a global spatiotemporal graph network includes the following steps: (1) Steps for constructing a road vehicle spatial topology network: Step 1: Construct a road vehicle network and a threshold-based road vehicle adjacency matrix. Record the historical road vehicle network as G= ( J t , B t ),in J t ={ j t1 , j t2 , …, j tN} is the first t Vehicles gather on the road at any time. N This refers to the number of vehicles on the road. B t It is a set of vehicle connection relationships. t The selected historical time step T The first in t time( t ∈ T ); built on t Time based on threshold k Road vehicle adjacency matrix ,in k For the selected distance threshold, the vehicle j tn With vehicles j tm distance H The calculation formula is as follows: in j tn =( x tn , y tn ), j tm =( x tm , y tm ), x tn for t time n The horizontal axis position coordinates of car number 1 y tn for t time n The longitudinal axis position coordinates of car number 1 x tm and y tm Similarly.

[0027] vehicle j n With vehicles j tmThe connection relationship is formed as follows: Step 2: Add skip step coefficients, reconstruct the adjacency matrix, and construct the spatial topology matrix. Based on the vehicle adjacency matrix constructed in step one Add a skip coefficient h The reconstructed vehicle adjacency matrix set is obtained ,in for of p power ( p ∈0,1,…, h This indicates the indirect distance relationship between vehicles. p Indicates the degree of indirect relationship. The elements in the matrix are represented as And create a mask matrix for all matrices in the matrix set. The elements in the mask matrix are represented as The calculation method is as follows: Ultimately, for Search all mask matrices in the matrix, as long as the vehicle v tn With vehicles v tm If the connection between the two edges exists once, then The reconstructed adjacency matrix is ​​obtained. and spatial topological matrix set ={ , , …, }

[0028] (2) Steps for constructing a vehicle-based global spatiotemporal graph matrix: Step 1: Using the finite cross-travel visibility method, analyze the correlation between road vehicles at each moment to construct the vehicle's own temporal logic network. To represent the time-dependent nature of a vehicle, the first... n The historical trajectory data of a vehicle is defined as follows: V (n) =[ v 1 (n) , v 2 (n) ,…, v T (n) At the sampling point i and j Choose any sampling point between c Perform the following calculations: in v i (n) , v j (n) and v c (n) They represent the first n Vehicle at the sampling point i , j and c The trajectory information data. If it satisfies Then it is considered v i (n) and v j (n) It can be seen through once. To i and j Perform the above calculation on all sampling points between them, and define: in L To accumulate the number of penetrations, The penetration threshold is set. When two sampling points are visible, an edge connection is established between them, thus obtaining the temporal logic network of a single vehicle. Q (n) =[ V (n) , E (n) ]( n ∈{1,…, N}), N This represents the total number of vehicles. E This represents the connection relationships between different sampling points. Finally, by repeating the above calculations, the temporal logic network of all vehicles can be obtained: And obtain the time logical adjacency matrix of each vehicle. For the first n vehicle time point i With time nodes j : Step 2: Construct a global spatiotemporal graph matrix by combining the road vehicle spatial topology network. By combining temporal logic networks and spatial topology networks, a global spatiotemporal graph matrix of the vehicle group is obtained. It consists of two parts: the first part is The diagonal of the vehicle tThe spatial topological adjacency matrix at time t; the second part is The rest. According to the time logic network. From the vehicle's own temporal logical adjacency matrix To obtain the temporal connection relationship of each vehicle, we obtain the temporal logical adjacency matrix { A″ 11 , A″ 12 ,…, A″ 21, A″ 22 ,…, A″ ( T-1}T , A″ TT}, and is composed of its splicing. The rest of the text. A″ ij It is a diagonal matrix representing the sampling time. i and sampling time j The temporal logical adjacency matrix of the vehicles is represented as follows: .

[0029] The structure diagram of the global spatiotemporal graph network model is as follows: Figure 2 As shown.

[0030] (3) Construct a graph operation module and a Seq2Seq network to implement the steps of encoding graph features and generating future trajectory data from the encoded hidden features: Step 1: Construct a graph operation module to obtain graph features based on global spatiotemporal context. The obtained global vehicle spatiotemporal topology matrix Compared with historical vehicle trajectory data X As input to the graph manipulation module, where X ={ X 1, X 2, …, X T},in X t This represents the road vehicle trajectory data at time t. t ∈0, 1, …, T ),and ,in x tn and y tn Indicates the first t time n The location coordinates of vehicle number 1; and stacked graph operations are used to extract global spatiotemporal features. The graph convolution operation can be described as follows: in yes The degree matrix, H For graph operations input, W This represents the weight matrix for graph convolution operations. l The number of layers for stacking graph operations ultimately yields graph features based on global spatiotemporal data. .

[0031] Step 2: Construct a Seq2Seq network to encode graph features and generate future trajectory data based on hidden features. The Seq2Seq network consists of two GRU networks, acting as an encoder and a decoder, respectively. The encoder encodes the graph features and extracts the hidden features from the GRU network. The calculation method is as follows: GRU stands for Gated Cyclic Unit Network. output This is the output of the last hidden layer. hidden This represents the hidden state at the last time step, which is the hidden feature extracted by the encoder from the graph features obtained by the graph operation module.

[0032] Based on the hidden features extracted by the GRU encoder, the GRU decoder predicts the trajectory data by analyzing the output position data at future time steps. The GRU decoding method is as follows: In the first decoding step, the input b t Position data of the last step in the historical time step X t , h t-1 Graph hidden features extracted by the encoder hidden , z t and r t For output control updates and resetting of the gating state, σ is the sigmoid activation function. W z , W r It updates and resets the weight matrix of the gating calculation steps; through h t-1 and r t Perform matrix multiplication and combine with bt After concatenation, the output is then activated by the tanh function. , Including current input b t information, W s To generate The weight matrix in the steps; Indicates the use of updated gating z t Selectively forget the hidden features of the current input. This indicates that the input includes the current input. b t Information Selective memorization yields hidden states. h t The hidden features are used as input for the next decoding step and output as the vehicle position information at the next time step. By repeatedly decoding, the position information of all future time steps is finally obtained, forming future trajectory data. ,in c The maximum step size for the future time to be predicted.

[0033] Step 3: Define the loss function for the global spatiotemporal graph network model Future trajectory data obtained from a global spatiotemporal graph network With real future trajectory data The loss function for the global spatiotemporal graph network model is defined as follows: .

[0034] The prediction results of the vehicle trajectory prediction model based on the global spatiotemporal graph network are shown in the figure below. Figure 3 As shown.

[0035] Example 2: Data from actual experiments (1) Selecting experimental data The experiment was evaluated using the publicly available large-scale highway NGSIM dataset. The NGSIM dataset is divided into three 15-minute time-segment datasets; we selected one-quarter of the data from the three subsets as the test set, and divided the remaining data into training and validation sets; all data were resampled into 5 frames, and the vehicle trajectory data was divided into 8-second time segments, with the first 3 seconds used as historical trajectory data, and the trajectory data for the next 5 seconds predicted.

[0036] (2) Parameter determination Distance threshold kWe selected a 25-foot encoder, with the GRU receiving 64-dimensional states as input, corresponding to the dimensions of the input map features. All GRU network layers were 64-dimensional, and the hidden states were activated by non-linear ReLU. We optimized the model using Adam with a learning rate of 0.001. The model was trained using PyTorch, with a batch size of 32 and 100 training epochs for both training and validation sets.

[0037] (3) Experimental results The experimental results are evaluated using the root mean square error (RMSE) as an indicator. The formula for calculating RMSE is as follows: in C The number of samples in the test set. and These are the actual future trajectory data and the predicted future trajectory data for the test set, respectively.

[0038] In the results analysis, the errors of all vehicles on the road were evaluated, and the experimental results are as follows: Table 1: Experimental Results of Vehicle Trajectory Prediction on the NGSIM Dataset Beneficial effects: (1) Based on the road vehicle location data at each moment, the constructed spatial topology matrix can mine the spatial positional relationship between vehicles at the current moment. By adding the step coefficient, the search range of the connection is expanded. Experiments show that the spatial correlation between vehicles can be fully mined. (2) Using a complex network of finite traversal view, a temporal logic network is constructed for vehicle trajectories to capture the temporal relationship between vehicles at each moment, fully reflecting the local spatiotemporal influence of the target vehicle within a certain time period. The constructed global spatiotemporal graph matrix forms the global spatiotemporal extraction of historical vehicle trajectory data information. (3) Using the graph spatiotemporal features extracted by the graph operation module, the encoding and decoding through the Seq2Seq network can fully extract the hidden features of the graph features and predict the position information of future steps. Experiments show that the accuracy of road vehicle trajectory prediction can be improved more effectively.

[0039] This invention constructs a spatial topology matrix by using the distance relationship and jump coefficient between road vehicle trajectory data, and combines a temporal logic network constructed from a complex network of finite traversal views to build a global spatiotemporal graph matrix. The graph operation module is used to extract graph features and a Seq2Seq network is used for encoding and decoding to obtain the position information of each future time step, thereby more effectively improving the accuracy of road vehicle trajectory prediction.

[0040] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A vehicle trajectory prediction method based on a global spatiotemporal graph network, characterized in that, The prediction method includes: Based on the distance relationships between road vehicle trajectory data and by incorporating a jump coefficient, a road vehicle spatial topology network is constructed. Using complex network methods, a temporal logic network of road vehicles was constructed; and combined with the spatial topology network of road vehicles, a global spatiotemporal graph matrix was constructed. The obtained global spatiotemporal graph matrix and historical vehicle trajectory data are used as inputs to the graph operation module to generate graph features; and the Seq2Seq network is used for encoding and decoding to obtain the location information of each future time step, thereby realizing the prediction of road vehicle trajectory data.

2. The vehicle trajectory prediction method based on a global spatiotemporal graph network according to claim 1, characterized in that, The construction of the road vehicle spatial topology network based on the distance relationships between road vehicle trajectory data and by adding a jump coefficient specifically includes: Construct a road vehicle network based on a road vehicle network and a threshold-based road vehicle adjacency matrix; Add a step coefficient, reconstruct the adjacency matrix, and construct the spatial topology matrix.

3. The vehicle trajectory prediction method based on a global spatiotemporal graph network according to claim 2, characterized in that, The construction of the road vehicle network based and the threshold-based road vehicle adjacency matrix specifically includes: Record the historical road vehicle network as G= ( J t , B t ),in J t ={ j t1 , j t2 , …, j tN } is the first t Vehicles gather on the road at any time. N This refers to the number of vehicles on the road. B t It is a set of vehicle connection relationships. t The selected historical time step T The first in t time, t ∈ T ; Built on t Time based on threshold k Road vehicle adjacency matrix ,in k For the selected distance threshold, the vehicle j tn With vehicles j tm distance H The calculation formula is as follows: in j tn =( x tn , y tn ), j tm =( x tm , y tm ), x tn for t time n The horizontal axis position coordinates of car number 1 y tn for t time n The longitudinal axis position coordinates of car number 1 x tm and y tm Similarly; vehicle j n With vehicles j tm The connection relationship is formed as follows: 。 4. The vehicle trajectory prediction method based on a global spatiotemporal graph network according to claim 2, characterized in that, The steps of adding a step coefficient, reconstructing the adjacency matrix, and constructing the spatial topology matrix specifically include: Based on the constructed vehicle adjacency matrix Add a skip coefficient h The reconstructed vehicle adjacency matrix set is obtained , in for of p power ( p ∈0,1,…, h This indicates the indirect distance relationship between vehicles. p Indicates the degree of indirect relationship. The elements in the matrix are represented as And create a mask matrix for all matrices in the matrix set. The elements in the mask matrix are represented as The calculation method is as follows: Ultimately, for Search all mask matrices in the system for vehicles. v tn With vehicles v tm If the connection between the two edges exists once, then The reconstructed adjacency matrix is ​​obtained. and spatial topological matrix set ={ , ,…, } 5. The vehicle trajectory prediction method based on a global spatiotemporal graph network according to claim 1, characterized in that, The method utilizes complex network techniques to construct the temporal logic network of road vehicles themselves; and combines this with the spatial topology network of road vehicles to construct a global spatiotemporal graph matrix, specifically including: Using the finite traversal visibility method, the correlation between road vehicles at each moment is analyzed, and the temporal logic network of the vehicles is constructed. A global spatiotemporal graph matrix is ​​constructed by combining the spatial topology network of road vehicles.

6. The vehicle trajectory prediction method based on a global spatiotemporal graph network according to claim 5, characterized in that, The method of using finite cross-travel visibility to analyze the correlation between road vehicles at each moment and construct the vehicle's own temporal logic network specifically includes: To represent the time-dependent nature of a vehicle, the first... n The historical trajectory data of a vehicle is defined as follows: V (n) =[ v 1 (n) , v 2 (n) ,…, v T (n) ], At the sampling point i and j Choose any sampling point between c Perform the calculation: in v i (n) , v j (n) and v c (n) They represent the first n Vehicle at the sampling point i , j and c The trajectory information data. If it satisfies Then it is considered v i (n) and v j (n) The line of sight can penetrate once; right i and j Perform the above calculation on all sampling points between them, and define: in L To accumulate the number of penetrations, The set penetration threshold; When two sampling points are visible, an edge connection is established to obtain the time logic network of a single vehicle. Q (n) =[ V (n) , E (n) ]( n ∈{1,…, N }), N This represents the total number of vehicles. E It refers to the connection relationship between different sampling points; Repeat the above calculations to obtain the time logic network for all vehicles: And obtain the time logical adjacency matrix of each vehicle. For the first n vehicle time point i With time nodes j : 。 7. The vehicle trajectory prediction method based on a global spatiotemporal graph network according to claim 5, characterized in that, The construction of the global spatiotemporal graph matrix by combining the road vehicle spatial topology network specifically includes: By combining temporal logic networks and spatial topology networks, a global spatiotemporal graph matrix of the vehicle group is obtained. The first part is: The diagonal of the vehicle t The spatial topological adjacency matrix at time t; the second part is The rest; Based on time logic network From the vehicle's own temporal logical adjacency matrix To obtain the temporal connection relationship of each vehicle, we obtain the temporal logical adjacency matrix { A″ 11 , A″ 12 ,…, A″ 21, A″ 22 ,…, A″ ( T-1}T , A″ TT }, and splice them together to form The rest; A″ ij It is a diagonal matrix representing the sampling time. i and sampling time j The temporal logical adjacency matrix of the vehicles is represented as follows: 。 8. The vehicle trajectory prediction method based on a global spatiotemporal graph network according to claim 1, characterized in that, The obtained global spatiotemporal graph matrix and historical vehicle trajectory data are used as inputs to the graph operation module to generate graph features; The location information for each future time step is obtained through encoding and decoding using a Seq2Seq network, enabling the prediction of road vehicle trajectory data. Specifically, this includes: Construct a graph operation module to obtain graph features based on global spatiotemporal data; Construct a Seq2Seq network to encode graph features and generate future trajectory data based on hidden features; Define the loss function for the global spatiotemporal graph network model: Future trajectory data obtained from a global spatiotemporal graph network With real future trajectory data The loss function for the global spatiotemporal graph network model is defined as follows: 。 9. The vehicle trajectory prediction method based on a global spatiotemporal graph network according to claim 8, characterized in that, The graph construction module obtains graph features based on global spatiotemporal conditions, specifically including: The obtained global vehicle spatiotemporal topology matrix Compared with historical vehicle trajectory data X As input to the graph manipulation module; in X ={ X 1, X 2, …, X T },in X t This represents the road vehicle trajectory data at time t. t ∈0, 1, …, T ),and ,in x tn and y tn Indicates the first t time n The location coordinates of vehicle number [number]; Stacked graph operations are used to extract global spatiotemporal features. The graph convolution operation is described as follows: in, yes The degree matrix, H For graph operations input, W This represents the weight matrix for graph convolution operations. l The number of layers for stacking graph operations ultimately yields graph features based on global spatiotemporal data. .

10. A vehicle trajectory prediction method based on a global spatiotemporal graph network according to claim 8, characterized in that, The construction of the Seq2Seq network to encode graph features and generate future trajectory data based on hidden features specifically includes: The Seq2Seq network consists of two GRU networks, acting as an encoder and a decoder, respectively. The encoder encodes the graph features and extracts the hidden features from the GRU network. The calculation method is as follows: GRU stands for Gated Cyclic Unit Network. output This is the output of the last hidden layer. hidden This is the hidden state of the last time step, which is the hidden feature extracted by the encoder from the graph features obtained by the graph operation module. Based on the hidden features extracted by the GRU encoder, the GRU decoder predicts the trajectory data by analyzing the output position data at future time steps. The GRU decoding method is as follows: In the first decoding step, the input b t Position data of the last step in the historical time step X t , h t-1 Graph hidden features extracted by the encoder hidden , z t and r t For output control updates and resetting of the gating state, σ is the sigmoid activation function. W z , W r It is the weight matrix for updating and resetting the gating calculation steps; pass h t-1 and r t Perform matrix multiplication and combine with b t After concatenation, the output is then activated by the tanh function. , Including current input b t information, W s To generate The weight matrix in the steps; Indicates the use of updated gating z t Selectively forget the hidden features of the current input. This indicates that the input includes the current input. b t Information Selective memorization yields hidden states. h t The hidden features are used as input for the next decoding step and output as the vehicle position information at the next time step. By repeatedly decoding, the position information of all future time steps is finally obtained, forming future trajectory data. ,in c The maximum step size for the future time to be predicted.

Images