A multi-domain perspective-based group vehicle trajectory prediction method

Abstract

A kind of group vehicle trajectory prediction method based on multi-domain perspective, first, in frequency domain, model utilizes fast Fourier transform to convert historical trajectory into frequency domain representation, selects the frequency of maximum amplitude to capture potential periodic characteristics, combines stacked two-dimensional convolution network and LSTM to extract features in depth;Second, in spatial domain, construct weighted directed spatial topology graph, model dynamic spatial relationship by the relative position, speed and acceleration information between vehicles, use stacked graph convolution network with residual connection to extract spatial features;Then, in time domain, based on circular finite penetration visibility graph, construct time logic graph, by allowing cross-time step connection to capture time-dependent features, use residual stacked GCN to extract time domain features;Finally, by the gate recurrent unit encoding-decoding structure fusion multi-domain feature, realize the trajectory prediction of group vehicle, the future trajectory of group vehicle is high-precision predicted by the present application.

Description

TECHNICAL FIELD

[0001] The present application belongs to the cross field of group vehicle trajectory prediction and deep learning, and is a group vehicle trajectory prediction method based on multi-domain feature fusion. BACKGROUND

[0002] With the rapid development of automatic driving technology, the safety and stability of vehicles have become the core research topic in intelligent transportation systems, and vehicle trajectory prediction is an important part of the automatic driving decision and control system. The main goal of vehicle trajectory prediction is to predict the motion behavior of surrounding vehicles to achieve safe driving of vehicles in complex traffic environments. However, the current mainstream vehicle trajectory prediction methods still have many shortcomings in feature extraction and prediction accuracy, and further improvement is needed to meet the actual needs of the automatic driving system.

[0003] Most of the existing vehicle trajectory prediction methods mainly focus on feature extraction in the time and spatial domains, often ignoring the potential value of frequency domain features. The motion behavior of vehicles often has periodic characteristics, and the introduction of frequency domain features can help identify the periodicity and oscillation characteristics in the trajectory. However, current deep learning-based trajectory prediction methods often fail to effectively extract frequency domain features, resulting in an incomplete feature space for vehicle trajectory prediction, limiting the accuracy and robustness of the prediction.

[0004] In addition, traditional trajectory prediction methods mainly focus on predicting a single vehicle, ignoring the dynamic interaction between vehicles. In actual traffic scenarios, the motion behavior of vehicles is often significantly influenced by neighboring vehicles, and relying solely on single vehicle prediction cannot achieve high-precision prediction between group vehicles, especially in dense traffic scenarios, where single vehicle prediction often falls short. Therefore, constructing a dynamic spatial topology graph that reflects the interdependence of group vehicles is of great significance to improve the accuracy of trajectory prediction and the practical application value. In addition to the modeling of spatial relationships, time-dependent modeling is also a key factor in trajectory prediction. However, current trajectory prediction methods often represent time dependence through simple time sequence, usually ignoring the modeling of cross-time step relationships, resulting in insufficient modeling of long-time dependence. In addition, with the rapid development of deep learning models, the structure of existing trajectory prediction models tends to be complex, although complex models have improved prediction results to some extent, but also increased the computational cost, affecting the real-time performance and generalization ability of the model. Complex model structure also makes it difficult to efficiently fuse multi-domain features, affecting the applicability of the model in complex traffic scenarios. SUMMARY

[0005] In order to overcome the problems of the existing vehicle trajectory prediction method in the comprehensiveness of feature extraction, the modeling of dynamic interaction relationship of group vehicles, the accurate expression of time dependence and the complexity of model structure, the present application proposes a group vehicle trajectory prediction method based on multi-domain perspective, so as to realize high-precision prediction of future trajectory of group vehicles.

[0006] The technical scheme adopted by the present application to solve its technical problems is:

[0007] A group vehicle trajectory prediction method based on multi-domain perspective, comprising the following steps:

[0008] Step 1, frequency domain feature extraction is realized: the historical vehicle trajectory data is mapped from the time domain to the frequency domain through fast Fourier transform (FFT), the amplitude of the frequency spectrum data is calculated, the key frequency component with the highest amplitude is selected, the trajectory data is reconstructed into a two-dimensional tensor according to the selected frequency, the row and the column represent the change within the period and the change between the periods respectively, so as to extract the features within the period and between the periods; the 2D convolutional neural network (CNN) is combined with the long short-term memory network (LSTM) to further extract the frequency domain features;

[0009] Step 2, spatial domain feature extraction is realized: the weighted directed spatial topology graph is constructed to dynamically represent the spatial relationship between group vehicles, when constructing the topology graph, the vehicles within a set distance around the target vehicle are regarded as its neighbors, and the connection weight between them is calculated; the node features in the topology graph are aggregated by GCN to extract the spatial interaction features between multiple vehicles in the spatial domain, so as to capture more detailed group vehicle motion characteristics;

[0010] Step 3, time domain feature extraction is realized: the time logic graph based on the limited penetration visibility graph (CLPVG) is constructed, first, the circular system between each pair of time points is determined, then the penetration number on the connection path is calculated, and whether it is visible is judged based on this; if the visibility condition is met, the corresponding edge is added in the time logic graph; finally, the feature aggregation is carried out on the time logic graph by the GCN with residual connection;

[0011] Step 4, feature fusion and trajectory generation: the features extracted in the frequency domain, the spatial domain and the time domain are fused in multiple domains to form a comprehensive feature representation for prediction, first, the frequency domain features, the spatial domain features and the time domain features are integrated into a high-dimensional feature vector through a feature splicing operation to realize mutual complementation and collaborative expression of the domain features; then, the fused feature vector is input into a GRU generator, the GRU effectively captures the time sequence dependence in the fused features through its internal memory and gating mechanism, and through the encoding and decoding process, the GRU generator gradually outputs the trajectory point sequence of each future time step to realize prediction of the future trajectory of the group vehicle.

[0012] The technical concept of the application is: first, the Fourier transform is introduced in the frequency domain to capture the periodic characteristics of the trajectory; a weighted directed topological graph is constructed in the spatial domain to dynamically represent the interaction relationship between vehicles; a limited penetration visibility graph is used in the time domain to model long-time dependence. Then, a deep neural network structure is built for multi-feature extraction and aggregation to effectively obtain multi-domain representation features. Finally, the future trajectory is generated by fusing the multi-domain features through the gated recurrent unit (GRU), realizing high-precision prediction of the group vehicle.

[0013] The beneficial effects of the application are: through the multi-domain feature fusion and generation method, high-precision prediction of the group vehicle trajectory is realized, which can effectively capture the dynamic interaction relationship between the group vehicles, significantly improving the prediction accuracy, safety and decision reliability of the automatic driving system in complex traffic scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is a schematic diagram of the overall framework of group vehicle trajectory prediction based on multiple domain perspectives.

[0015] Figure 2 is a schematic diagram of vehicle relative coordinate calculation.

[0016] Figure 3 is a circular system and penetration diagram for modeling a limited penetration visibility time graph.

[0017] Figure 4 is the trajectory prediction visualization result under scenario 1.

[0018] Figure 5 is the trajectory prediction visualization result under scenario 2. DETAILED DESCRIPTION

[0019] The application will be further described below in conjunction with the drawings.

[0020] REFERENCE Figures 1-5A multi-domain perspective-based group vehicle trajectory prediction method, given a set of group vehicle historical trajectory data V, which contains observation values within T time steps, the trajectory state of each time step contains position information, speed and acceleration and other features as shown in formula (1), the trajectory observation sequence of each vehicle m is shown in formula (2), and the state of vehicle m at each time step t is shown in formula (3);

[0021] V = [V 1 ,V 2 ,…,V M ] (1)

[0022]

[0023] The trajectory prediction task is to predict the trajectory within the future T p time steps based on the observation data V of the past T time steps, and output V' as shown in formula (4);

[0024] V′ = [V′ 1 ,V′ 2 ,…,V′ M ] = Γ([V 1 ,V 2 ,…,V M ]) (4)

[0025] In the formula, Γ is a trajectory prediction model.

[0026] The multi-domain perspective-based group vehicle trajectory prediction method comprises the following steps:

[0027] Step 1, realize frequency domain feature extraction, the process is as follows:

[0028] Step 1.1, trajectory frequency domain feature transformation

[0029] Through frequency domain analysis to mine the periodic characteristics in the vehicle trajectory, to enhance the accuracy of trajectory prediction. First, apply FFT to the historical trajectory data V, convert it from time domain to frequency domain to reveal the potential periodic characteristics, as shown in formula (5); then, by selecting the first k main frequencies {f1,…,f k} with the largest amplitude value, the significant periodic characteristics of the trajectory are extracted, the selection of main frequencies is shown in formula (6), and each main frequency f i corresponds to a period In order to further analyze the periodic changes, the trajectory data is reconstructed into a two-dimensional tensor V j containing intra-period and inter-period changes, as shown in formula (7);

[0030] A = Average[Amp(FFT(V))] (5)

[0031] {f1…, f k}=argTopk(A) (6)

[0032]

[0033] In the formula, FFT(V) represents the Fourier transform of data V to obtain the amplitude of each frequency component; Amp represents the calculation of the amplitude of the frequency component; Average represents the average of the amplitude to obtain the overall frequency spectrum distribution; Padding(V) fills the trajectory data to adapt to the period length p j and frequency component f i , According to the period and frequency, the tensor is reconstructed;

[0034] Step 1.2, frequency domain feature extraction network construction

[0035] The reconstructed tensor The residual connection stacked two-dimensional convolutional network CNN is used for feature extraction to capture the detailed features within and between periods, as shown in formula (8); the convolution operation captures the subtle changes of the trajectory in the periodic structure through the local perception mechanism, and outputs the feature tensor after convolution processing;

[0036]

[0037] In the formula, and e respectively represent the vector composed of four weight matrices and the number of layers of the convolutional network, Conv e represents the e-th layer two-dimensional convolutional network (2D CNN), represents the initial value, represents the output result;

[0038] Then the time dimension of is reconstructed into So that k features can be aggregated into a tensor, as shown in formula (9), and the obtained feature is further extracted by LSTM to obtain the feature information in the frequency domain, as shown in formula (11); the output of LSTM is represented as V F =[h1,h2,…,h T ];

[0039]

[0040] In the formula, Trunc is a truncation operation, which is used to make shape consistent with the vehicle trajectory; W L is the LSTM weight matrix, and h t is the hidden state vector of the time step state t;

[0041] Step 2, spatial domain feature extraction is implemented, the process is as follows:

[0042] Step 2.1, weighted directed spatial topology graph construction

[0043] The spatial topology relationship of vehicles interacting at any time is mined through spatial domain analysis to enhance the accuracy of trajectory prediction. The graph calculates the relative state and actual distance of vehicles by distance, relative position, speed and acceleration between vehicles, and constructs the connection relationship between vehicles. The edges of the spatial graph are assigned corresponding weights to represent the direction and closeness of the connection between vehicles. First, the neighbor vehicles around the target vehicle are defined, and the vehicles within the specified distance d are set as neighbors. The actual relative distance between vehicles is defined as shown in equations (12)-(14). P marks the relative position, and values 1 and 2 represent front and rear vehicles in the same lane, and 3 and 4 represent front and rear vehicles in different lanes, as shown in Figure 2 ;

[0044]

[0045] In the formula, Δx ij and Δy ij represent the lateral and longitudinal relative distances of vehicles i and j, respectively; l i and l j are the lengths of the two vehicles; and Δw ij is the difference in vehicle width.

[0046] The relative state of the vehicle is defined by the time required for vehicle i to collide with its adjacent vehicle, as shown in equation (15). The edge weight between the two vehicles is determined based on the collision time t coll and d rela , as shown in equation (16). Thus, the adjacency matrix containing all vehicle connection relationships is obtained, as shown in equation (17). Equation (18) is the weight normalization process, which forms the normalized adjacency matrix, as shown in equation (19).

[0047]

[0048] In the formula, t coll is the time required for vehicles i and j to collide in the current state, assuming that their driving directions are changed to be consistent with the direction of the connection between the two vehicles, and Δa and Δv are the relative acceleration and relative speed between the vehicles, respectively. represents the edge weight between vehicles i and j, and I is the unit matrix.

[0049] Step 2.2, spatial feature aggregation graph deep network construction

[0050] The spatial features are extracted by using a stack of GCNs with residual connections. The GCN aggregates the spatial features by the following formula (20). The graph convolution is an important component of the GCN. The features of each node and its adjacent nodes can be aggregated by the graph convolution operation. The principle formula is shown in formula (21). Finally, the feature dimension is adjusted by using a fully connected layer, and the extracted features are output by a time convolution layer, and then the spatial perspective feature output V is obtained S ;

[0051]

[0052] In the formula, z is the number of GCNs; represents the zth GCN; is the weight matrix of the zth layer; are the outputs of the first two layers of GCN respectively; the initial input is the initial state information of the vehicle; is the output of the GCN residual connection; D is a degree matrix describing each node; is the weight matrix of the zth layer;

[0053] Step 3, realize the time domain feature extraction, the process is as follows:

[0054] Step 3.1, time logic graph construction. In order to effectively establish the time interaction logic between the historical trajectories of vehicles, the logical relationship between time points is modeled based on the visibility graph of the circular system. First, introduce the circular system of two time points (trajectory points), as shown in formula (22);

[0055]

[0056] In the formula, (t, v) is the time and speed corresponding to any trajectory point b between trajectory points a and c; α is an adjustment parameter used to control the flexibility of the circular system. The penetration number is defined as the number of trajectory points that can be "penetrated" on the connecting line from point a to point c; if the connecting line connecting a and c can penetrate point b, the penetration number "NumP" increases by 1, and if the final penetration number NumP≤L, a point and c point are considered to be visible, and the adjacent edge relationship is established; if the penetration number NumP>L, a point and c point are considered to be invisible, and the adjacent edge relationship is not established, as shown in Figure 3 ; This step can construct a time logic graph of a historical trajectory, as shown in formula (23). Through iterative calculation, the time logic graph containing all vehicles is obtained, as shown in formula (24);

[0057] L m =[V m ,E m +1] m∈{1,…,M} (23)

[0058] L=[L 1 ,L2 ,…,L M ]

[0059] E m represents the connectivity between different trajectory points;

[0060] Step 3.2, time feature aggregation graph depth network building

[0061] The time logic graph and the space topology graph adopt the same structure of residual connection stacked GCN, the feature aggregation formula of the time perspective is as formula (24), and the definition of graph convolution is as formula (25); the dimension of the feature is adjusted through the full connection layer and the extracted feature is output through the time convolution layer. Finally, the output V T of the time perspective is obtained.

[0062]

[0063] In the formula, p is the number of GCN, is the pth GCN, is the weight matrix of the pth layer; is the network input, is the GCN output; is the degree matrix of each node, is the weight matrix;

[0064] Step 4, realizing multi-domain feature fusion and trajectory encoding / decoding, the process is as follows:

[0065] The feature output V F ,V T ,V S is fused according to formula (26) to obtain the fusion feature matrix H; then, the features are encoded and decoded by GRU, and each time dimension of the fusion feature generates the corresponding hidden feature through GRU, as shown in formula (27); after the encoding is completed, the H hid is decoded by the GRU decoder to obtain the expected future trajectory, and the decoding process of the GRU is shown as formula (28)-(31);

[0066] H=Concatg(V F ,V T ,V S ) (26)

[0067] H enc ,H hid =GRU E (H) (27)

[0068]

[0069] In the formula, GRUE is the GRU encoding network; H enc is the output of the last hidden layer; H hid is the hidden state of the last time step; x t is the trajectory point coordinate at time t, h t-1 is H hid , r t , h t is the state of the reset gate and update gate, W r , W z is the weight matrix of the reset gate and update gate, W is is the weight matrix in the generation step, and is element multiplication.

[0070] After decoding, the predicted future trajectory is obtained, as shown in equations (32)-(34).

[0071] V' = [V' 1 ,V' 2 ,…,V' M ] (32)

[0072]

[0073] In the formula, is the predicted coordinate of vehicle m at future T+t.

[0074] The simulation experiment of the embodiment is as follows:

[0075] (1) Selection of data set

[0076] The experiment is based on the NGSIM data set, which contains detailed vehicle trajectory information of two traffic scenes: I-80 highway, collected from the eastbound I-80 section of the San Francisco Bay Area in the United States, containing vehicle driving data for 500 meters; US-101 highway, collected from the southbound US-101 section of the Los Angeles area, containing 640 meters of driving data.

[0077] The NGSIM data set contains some noise due to the use of video analysis and numerical differentiation method. Therefore, the Savitzky-Golay filter is used to smooth the vehicle position to improve the accuracy of the trajectory data. The data set is divided into three 15-minute subsets, and finally divided into training set, validation set and test set. The original 10 frames of data are resampled into 5 frames, and 3 seconds of historical trajectory data of each sequence is used to predict the future 5 seconds of trajectory.

[0078] (2) Performance comparison experiment

[0079] During the experiment, the trained selection Adam optimizer is selected, the learning rate is set to 0.001, the training batch size is 32, and the total training number is 100. The input and output sizes of the LSTM are 4 and 8 respectively, and k in argTopk is 5; the convolution kernel size of all time convolution layers in the graph convolution network is 1x3, the step size is set to 1, and appropriate padding is performed to maintain the shape of the features. In order to avoid overfitting, the Dropout is set to 0.5; the input of the encoder GRU is a 72-dimensional state corresponding to the dimension of the input feature, and the size of all GRU network layers is 64-dimensional, and the hidden state is activated by a nonlinear rectified linear unit. The total loss function during training is shown in equation (35);

[0080]

[0081] The evaluation indicators during the experiment are root mean square error (RMSE), average displacement error (ADE) and final displacement error (FDE), and the calculation formulas are shown in equations (36)-(38);

[0082]

[0083]

[0084] The performance comparison experiment is compared with five baseline trajectory prediction models of Constant Velocity (CV), CS-LSTM, MHA-LSTM, AI-TP and GRIP, and the performance comparison experiment results are shown in Table 1, which is the prediction performance comparison of each model on the NGSIM I-80 and US-101 data set trajectories;

[0085]

[0086] Continued table

[0087]

[0088] Table 1

[0089] The comparison experiment results show that in the single target vehicle trajectory prediction, the 1-second short-term prediction accuracy of the model is improved by 70.33%, and the 5-second long-term prediction accuracy is improved by 55.58%; and in the group vehicle prediction, the accuracy of the model in the 1-second short-term prediction is improved by 48.13%, and the 5-second long-term prediction accuracy is improved by 35.78%. It is shown that the robustness and accuracy of the method in the single target and group vehicle prediction are better than those of the baseline method, and the group trajectory prediction performance can be more accurate.

[0090] (3) Ablation verification experiment

[0091] In the ablation experiment, the distance threshold, the value selection in CLPVG, the selection of k value in the frequency perspective, and the weighted directed spatial topology graph are analyzed. Distance threshold: In the spatial perspective, the distance threshold d used when constructing the spatial topology graph will affect the prediction accuracy of the model. The experiment tests the case when d is 0, 15, 25, 35, and 50 feet. The results show that when d = 25, the prediction accuracy of the model is the highest. Value selection in CLPVG: In the construction of the time graph, the information based on acceleration can better reflect the time dependence of the vehicle trajectory. The experiment shows that the prediction accuracy of the model based on acceleration to construct CLPVG is about 3.68% higher than that of the model based on speed. Selection of k value in the frequency perspective: In the frequency perspective, selecting the top k frequencies with the highest amplitude will affect the prediction effect. The experiment tests different k values and finds that when k = 5, the prediction accuracy is the best. Further increasing k will introduce feature redundancy and affect the model effect. Weighted directed spatial topology graph: Replace the weighted directed graph with the ordinary spatial topology graph and set all weights to 1. The experimental results show that the accuracy in trajectory prediction decreases by 5.60%.

[0092] (4) Visual performance analysis

[0093] The visualization experiment shows the group vehicle trajectory prediction effect of the model in different scenarios, such as Figure 4 , 5 The results show that the model can accurately predict the future trajectory while keeping the lane, but the prediction accuracy decreases when the vehicle suddenly changes lanes or decelerates. This indicates that the model has strong robustness in predicting trajectories in stable states, but has certain limitations in dealing with sudden behaviors.

[0094] The content described in the embodiments of the present specification is only a list of implementation forms of the inventive concept, and is only for illustrative purposes. The protection scope of the present application should not be regarded as being limited to the specific forms described in the present embodiments, and the protection scope of the present application also extends to equivalent technical means that can be thought of by those skilled in the art according to the inventive concept.

Claims

1.A method for predicting trajectories of a group of vehicles based on multi-domain perspectives, the method comprising: The method comprises the following steps: Step 1, realizing frequency domain feature extraction: mapping historical vehicle trajectory data from time domain to frequency domain through fast Fourier transform (FFT), calculating the amplitude of the frequency spectrum data, selecting the key frequency component with the highest amplitude, and reconstructing the trajectory data into a two-dimensional tensor according to the selected frequency, with the rows and columns representing the intra-period variation and inter-period variation, respectively, to extract the intra-period and inter-period features; using a 2D convolutional neural network (CNN) combined with a long short-term memory (LSTM) network, further extracting the frequency domain features; Step 2, realizing spatial domain feature extraction: constructing a weighted directed spatial topology graph to dynamically represent the spatial relationship between group vehicles, considering the vehicles within a certain distance around the target vehicle as its neighbors when constructing the topology graph, and calculating the connection weights between them; using GCN to aggregate the node features in the topology graph to extract the spatial interaction features between multiple vehicles in the spatial domain, to capture more detailed group vehicle motion characteristics; Step 3, realizing time domain feature extraction: constructing a time logic graph based on a circular limited penetrable visibility graph (CLPVG), first determining the circular system between each pair of time points, then calculating the penetration number on the connection path, and judging whether it is visible based on this; if the visibility condition is met, add the corresponding edge in the time logic graph; finally, performing feature aggregation on the time logic graph through a GCN with residual connection; Step 4, feature fusion and trajectory generation: multi-domain fusion of the features extracted in the frequency domain, spatial domain and time domain to form a comprehensive feature representation for prediction, first integrating the frequency domain features, spatial domain features and time domain features into a high-dimensional feature vector through feature concatenation operation, to realize mutual complement and collaborative expression of the domain features; then, inputting the fused feature vector into a GRU generator, which effectively captures the temporal dependency in the fused features through its internal memory and gating mechanism, and gradually outputs the trajectory point sequence at each future time step through the encoding and decoding process, realizing prediction of the future trajectory of the group vehicle. 2.The multi-domain perspective based group vehicle trajectory prediction method of claim 1, wherein, The process of step 1 is as follows: Step 1.1, trajectory frequency domain feature transformation Periodic features in vehicle trajectories are mined through frequency domain analysis to enhance the accuracy of trajectory prediction; first, the historical trajectory data V is applied with FFT to convert it from time domain to frequency domain to reveal potential periodic features, as in equation (5); subsequently, the significant periodic features of the trajectory are extracted by selecting the top k dominant frequencies {f1,…,f k} with the largest amplitude values, the selection of dominant frequencies as in equation (6), each dominant frequency f i corresponds to a period To further analyze the periodic changes, the trajectory data is reconstructed into a two-dimensional tensor V j , as in equation (7); A = Average[Amp(FFT(V))] (5) {f1,…,f k} = argTopk(A) (6) where FFT(V) represents the Fourier transform of data V to obtain the amplitude of each frequency component; Amp represents the calculation of the amplitude of the frequency component; Average represents the averaging of the amplitude to obtain the overall frequency spectrum distribution; Padding(V) pads the track data to fit the period length p j and frequency component f i , Tensor reconstruction according to period and frequency; Step 1.2, frequency domain feature extraction network construction Reconstructed tensor Feature extraction is performed using a stack of 2D convolutional networks CNN with residual connections to capture intra- and inter-periodic details, as in equation (8); the convolutional operation captures subtle changes in the trajectory within the periodic structure through a local perception mechanism, and outputs a feature tensor after convolution wherein, and e respectively represent a vector composed of four weight matrices and the number of layers of the convolutional network, Conv e represents the e-th layer two-dimensional convolutional network, represents an initial value, represents an output result; Then The time dimension is reconstructed as This allows k features to be aggregated into a tensor, as shown in equation (9); the resulting features Further extraction using LSTM is used to obtain feature information from the frequency domain perspective, as shown in Equation (11); the output of LSTM is represented as V F =[h1,h2,…,h T ]; where Trunc is a truncation operation to make shape consistent with the vehicle trajectory; W L is an LSTM weight matrix, h t is a hidden state vector at time step t. 3.The multi-domain perspective based group vehicle trajectory prediction method of claim 1 or 2, wherein, The process of step 2 is as follows: Step 2.1, construction of weighted directed spatial topology graph The spatial topology relationship of vehicles at any time is mined by spatial domain analysis, the relative state and actual distance of vehicles are represented by calculating the distance, relative position, speed and acceleration between vehicles, and the connection relationship between vehicles is constructed; the edges of the spatial graph are assigned with corresponding weights, representing the direction and closeness of the connection between vehicles; first, the neighbor vehicles around the target vehicle are defined, and the vehicles within the specified distance d are set as neighbors; the actual relative distance between vehicles is defined as shown in equations (12)-(14), P is marked as the relative position, and values 1 and 2 represent front and rear vehicles in the same lane, and values 3 and 4 represent front and rear vehicles in different lanes; where Δx ij and Δy ij represent the lateral and longitudinal relative distances of vehicles i and j, respectively; l i and l j are the lengths of the two vehicles; and Δw ij is the difference in the width of the vehicles. The relative state of the vehicle is defined by the time required for vehicle i to collide with its neighboring vehicles, as shown in equation (15), based on the collision time t coll and d rela The edge weight between two vehicles is determined as shown in equation (16); thus obtaining the adjacency matrix containing all the connection relationships of the vehicles, as shown in equation (17); equation (18) is the normalization processing of the weight, forming the normalized adjacency matrix, as shown in equation (19); where t coll is the time required for vehicles i and j to collide at the current state, assuming their driving directions are changed to be consistent with the direction of the line connecting the two vehicles, and Δa and Δv are the relative acceleration and relative speed between the vehicles, respectively; denotes the edge weight between vehicles i and j, and I is the identity matrix. Step 2.2, building of spatial feature aggregation graph deep network The spatial features are extracted by using a stacked GCN with a residual connection. The GCN aggregates the spatial features by using the following formula (20). The graph convolution is an important component of the GCN. The features of each node and its adjacent nodes can be aggregated by the graph convolution operation. The principle formula is as shown in formula (21). Finally, the feature dimension is adjusted by using a fully connected layer, and the extracted features are output by using a time convolution layer, so as to obtain the spatial view feature output V s ; where z is the number of GCNs; represents the z-th GCN; is the weight matrix of the z-th layer; are the outputs of the first two layers of GCN respectively; initial input is the initial state information of the vehicle; is the output of the GCN residual connection; D is the degree matrix describing each node; is the weight matrix of the z-th layer. 4.The multi-domain perspective based group vehicle trajectory prediction method of claim 3, wherein, The process of step 3 is as follows: Step 3.1, construction of time logic graph, to effectively establish the time interaction logic between vehicle historical trajectories, the logical relationship between time points is modeled based on the visibility graph of the circular system, first introduce two time points, namely the circular system of the trajectory point, as shown in equation (22); In the formula, (t, v) is the time and speed corresponding to any trajectory point b between trajectory points a and c; α is an adjustment parameter used to control the flexibility of the circular system; the penetration number is defined as the number of trajectory points that can be "penetrated" on the connecting line from point a to point c; if the connecting line connecting a and c can penetrate point b, the penetration number "NumP" increases by 1, if the final penetration number NumP≤L, a and c are considered to be visible, and the adjacent edge relationship is established; if the penetration number NumP>L, a and c are considered to be invisible, and no adjacent edge relationship is established; this step can construct a time logic graph of a historical trajectory, as shown in equation (23), and the time logic graph containing all vehicles is obtained by iterative calculation, as shown in equation (24); L m = [V m ,E m +1] m∈{1,…,M} (23)L= [L 1 ,L 2 ,…,L M ] wherein E m represents the connectivity between different trajectory points; Step 3.2, building of time feature aggregation graph deep network The time logic diagram and the space topology diagram adopt the same structure of residual connection stacked GCN, the feature aggregation formula of the time perspective is as formula (24), and the graph convolution definition is as formula (25); the dimension of the feature is adjusted through a fully connected layer, and the extracted feature is output through a time convolution layer, and finally the output V of the time perspective is obtained T ; where p is the number of GCNs, is the p-th GCN, is the weight matrix of the p-th layer; is the network input, is the GCN output; is the degree matrix of each node, is the weight matrix. 5.The multi-domain perspective based group vehicle trajectory prediction method of claim 4, wherein, The process of step 4 is as follows: The feature output V after multi-domain information extraction / aggregation F 、V T 、V S The feature fusion is performed according to the form of formula (26) to obtain a fusion feature matrix H. Subsequently, the features are encoded and decoded by the GRU, and each time dimension of the fused features is generated by the GRU to generate the corresponding hidden features, as shown in equation (27); after the encoding is completed, the hidden state H hid is decoded by the GRU decoder to obtain the expected future trajectory, and the decoding process of the GRU is shown in equations (28)-(31); H = Concat(V F ,V T ,V S ) (26) H enc ,H hid = GRU E (H) (27) In the formula, GRU E is a GRU encoding network; H enc is the output of the last hidden layer; H hid is the hidden state of the last time step; x t is the trajectory point coordinate at time t, h t-1 is H hid , r t , h t is the state of the reset gate and the update gate, W r , W z is the weight matrix of the reset gate and the update gate, W is the weight matrix in the generation step, and is element multiplication. After decoding, the predicted future trajectory is obtained, as shown in equations (32)-(34); V′=[V′ 1 ,V′ 2 ,…,V′ M ] (32) In the formula, Predicted coordinates of the formula vehicle m at future T+t.

Images