Group vehicle trajectory prediction method based on multi-domain view angle

Abstract

The invention discloses a group vehicle trajectory prediction method based on a multi-domain view angle, and the method comprises the steps: enabling a model to convert a historical trajectory into frequency domain representation through the fast Fourier transform in the aspect of a frequency domain, selecting a frequency with the maximum amplitude to capture a potential periodic feature, and carrying out the deep extraction of the feature through the combination of a stacked two-dimensional convolution network and an LSTM (Long Short Term Memory); secondly, in the aspect of a spatial domain, a weighted directed spatial topological graph is constructed, a dynamic spatial relationship is modeled through relative position, speed and acceleration information between vehicles, and spatial features are extracted by using a stacked graph convolutional network with residual connection; then, in the aspect of time domain, constructing a time logic diagram based on the circular finite penetration visibility diagram, capturing time-dependent features by allowing connection across time steps, and extracting time domain features by using residual stacking GCN; finally, the multi-domain features are fused through a coding-decoding structure of the gating circulation unit, trajectory prediction of the group vehicles is achieved, and high-precision prediction of future trajectories of the group vehicles is achieved.

Description

Technical Field

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

[0002] With the rapid development of autonomous driving technology, vehicle safety and operational stability have become core research topics in intelligent transportation systems, and vehicle trajectory prediction is an important part of autonomous driving decision-making and control systems. The main goal of vehicle trajectory prediction is to achieve safe driving of vehicles in complex traffic environments by predicting the motion behavior of surrounding vehicles. However, the current mainstream vehicle trajectory prediction methods still have many deficiencies in feature extraction and prediction accuracy, and further improvements are urgently needed to meet the actual needs of autonomous driving systems.

[0003] Most existing vehicle trajectory prediction methods focus on feature extraction in the time and space domains, and usually ignore 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 periodic and oscillatory features in trajectories. However, current trajectory prediction methods based on deep learning often fail to effectively extract frequency domain features, resulting in an incomplete feature space for vehicle trajectory prediction, which limits the accuracy and robustness of the prediction.

[0004] In addition, traditional trajectory prediction methods mainly predict single vehicles, ignoring the dynamic interaction between vehicles. In actual traffic scenarios, the movement behavior of vehicles is usually significantly affected by neighboring vehicles. It is difficult to achieve high-precision prediction between groups of vehicles by relying solely on single-vehicle prediction, especially in dense traffic scenarios, where the prediction of a single vehicle is often insufficient. Therefore, constructing a dynamic spatial topological map that can reflect the interdependence of group vehicles is of great significance for improving the accuracy of trajectory prediction and its practical application value. In addition to the modeling of spatial relationships, temporal dependency modeling is also a key factor in trajectory prediction. However, current trajectory prediction methods mostly represent temporal dependencies through simple temporal sequences, usually ignoring the modeling of relationships across time steps, resulting in insufficient modeling of long-term dependencies. In addition, with the rapid development of deep learning models, the structure of existing trajectory prediction models tends to be complex. Although complex models improve the prediction effect to a certain extent, they also increase the computational cost, affecting the real-time and generalization capabilities of the models. The complex model structure also makes it difficult to achieve efficient fusion of multi-domain features, affecting the applicability of the model in complex traffic scenarios. Summary of the invention

[0005] In order to overcome the problems of existing vehicle trajectory prediction methods in terms of comprehensiveness of feature extraction, modeling of dynamic interaction relationships of group vehicles, accurate expression of time dependencies, and complexity of model structure, the present invention proposes a group vehicle trajectory prediction method based on a multi-domain perspective, thereby achieving high-precision prediction of the future trajectory of group vehicles.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

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

[0008] Step 1: Extract frequency domain features: Map historical vehicle trajectory data from the time domain to the frequency domain through fast Fourier transform (FFT), calculate the amplitude of the spectrum data, select the key frequency component with the highest amplitude, and reconstruct the trajectory data into a two-dimensional tensor according to the selected frequency. The rows and columns represent the intra-cycle changes and inter-cycle changes respectively, so as to extract the intra-cycle and inter-cycle features; use the 2D convolutional neural network (CNN) combined with the long short-term memory network (LSTM) to further extract the frequency domain features;

[0009] Step 2: Extract spatial domain features: The constructed weighted directed spatial topology graph is used 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 weights between them are calculated. The node features in the GCN aggregation topology graph are used to extract the spatial interaction features of multiple vehicles in the spatial domain to capture more detailed group vehicle motion characteristics.

[0010] Step 3: Extract features in the time domain: By constructing a time logic graph based on the Circular Limited Penetrable Visibility Graph (CLPVG), first determine the circular system between each pair of time points, and then calculate the number of penetrations on the connection path to determine whether it is visible; if the visibility condition is met, add the corresponding edge to the time logic graph; finally, perform feature aggregation on the time logic graph through the residual connected GCN;

[0011] Step 4, feature fusion and trajectory generation: The features extracted from the frequency domain, space domain and time domain are fused in multiple domains to form a comprehensive feature representation for prediction. First, the frequency domain features, space domain features and time domain features are integrated into a high-dimensional feature vector through feature concatenation operation to achieve mutual complementation and coordinated expression of features in each domain; then, the fused feature vector is input into the GRU generator. The GRU effectively captures the temporal dependency in the fused features through its internal memory and gating mechanism. Through the encoding and decoding process, the GRU generator gradually outputs the trajectory point sequence of each future time step to achieve the prediction of the future trajectory of the group of vehicles.

[0012] The technical concept of the present invention is as follows: First, 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-term dependencies. Subsequently, a deep neural network structure for multi-feature extraction and aggregation is built to effectively obtain multi-domain representation features. Finally, multi-domain features are fused through the gated recurrent unit (GRU) to generate future trajectories, achieving high-precision prediction of group vehicles.

[0013] The beneficial effects of the present invention are as follows: through the multi-domain feature fusion and generation method, high-precision prediction of group vehicle trajectories is achieved, the dynamic interaction relationship between group vehicles can be effectively captured, and the group behavior prediction accuracy, safety and decision-making reliability of the autonomous driving system in complex traffic scenarios are significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a schematic diagram of the overall framework of group vehicle trajectory prediction based on a multi-domain perspective.

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

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

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

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

[0019] The present invention is further described below in conjunction with the accompanying drawings.

[0020] Reference Figure 1 to Figure 5, a group vehicle trajectory prediction method based on a multi-domain perspective, given a group of historical trajectory data V of vehicles, which contains observations within T time steps, the trajectory state of each time step contains location information, speed, 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 representation 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 future T time steps based on the observation data V of the past T time steps p The trajectory within each time step is outputted as 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 the trajectory prediction model.

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

[0027] Step 1: Implement frequency domain feature extraction. The process is as follows:

[0028] Step 1.1: Trajectory frequency domain feature transformation

[0029] The frequency domain analysis is used to mine the periodic features in the vehicle trajectory to enhance the accuracy of trajectory prediction. First, FFT is applied to the historical trajectory data V to transform it from the time domain to the frequency domain to reveal the potential periodic features, as shown in Equation (5). Then, the first k main frequencies {f1,…,f k}, extract the significant periodic features of the trajectory, and select the main frequency as shown in formula (6). Each main frequency f i Corresponding to a cycle In order to further analyze the periodic changes, the trajectory data is reconstructed into a two-dimensional tensor V containing the intra-cycle and inter-cycle changes j , 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) means that the amplitude of each frequency component is obtained by Fourier transforming the data V; Amp means the amplitude of the calculated frequency component; Average means averaging the amplitude to obtain the overall spectrum distribution; Padding (V) fills the trajectory data to adapt to the period length p j and the frequency component f i , Reconstruct tensors according to period and frequency;

[0034] Step 1.2: Frequency domain feature extraction network construction

[0035] Reconstructed tensor A stacked two-dimensional convolutional network (CNN) with residual connections is used for feature extraction to capture the detailed features within and between cycles, 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 represent the vector composed of four weight matrices and the number of layers of the convolutional network, respectively. e represents the e-th layer of the two-dimensional convolutional network (2D CNN), represents the initial value, Indicates the output result;

[0038] Then will The time dimension is reconstructed into So that k features can be aggregated into a tensor, as shown in formula (9), the obtained features Further extraction is performed through LSTM to obtain feature information from the frequency domain perspective, as shown in formula (11); the output of LSTM is expressed as V F =[h1,h2,…,h T ];

[0039]

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

[0041] Step 2: Implement spatial domain feature extraction. The process is as follows:

[0042] Step 2.1: Construction of weighted directed spatial topology graph

[0043] The spatial topological relationship of vehicle interactions at any time is mined through spatial domain analysis to enhance the accuracy of trajectory prediction. The graph uses the distance, relative position, speed and acceleration between vehicles to represent the relative state and actual distance of the vehicles and construct the connection relationship between vehicles. The edges of the spatial graph are given 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 range 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. The values ​​1 and 2 represent the front and rear vehicles in the same lane, and 3 and 4 represent the front and rear vehicles in different lanes. Figure 2 ;

[0044]

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

[0046] The relative state of the vehicles is defined by the time required for vehicle i to collide with its adjacent vehicle, as shown in Equation (15). coll and d rela The edge weight between two vehicles is determined as shown in formula (16), so the adjacency matrix containing the connection relationship of all vehicles is obtained, as shown in formula (17). Formula (18) is the normalization processing of the weight to form a normalized adjacency matrix, as shown in formula (19);

[0047]

[0048] Where, t coll is the time required for vehicles i and j to collide in their current state, assuming that their driving directions change to be consistent with the direction of the line connecting the two vehicles, Δa and Δv are the relative acceleration and relative velocity between the vehicles, respectively; represents the edge weight between vehicle i and vehicle j, where I is the unit matrix;

[0049] Step 2.2: Construction of deep network for spatial feature aggregation graph

[0050] The spatial features are extracted by stacking GCN with residual connections. GCN aggregates spatial features through the following formula (20). Graph convolution is an important part of GCN. The features of each node and its neighboring nodes can be aggregated through graph convolution operation. The principle formula is shown in formula (21). Finally, the fully connected layer is used to adjust the feature dimension, and the extracted features are output through the temporal convolution layer to obtain the spatial perspective feature output V. S ;

[0051]

[0052] Where z is the number of GCNs; represents the zth GCN; is the weight matrix of the zth layer; They are the outputs of the first two layers of GCN; the 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 zth layer;

[0053] Step 3: Implement 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 vehicle historical trajectories, the logical relationship between time points is modeled based on the visibility graph of the circular system. First, the circular system of two time points (trajectory points) is introduced, as shown in formula (22);

[0055]

[0056] Where (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 "passed through" on the line from point a to point c; if the line connecting a and c can penetrate point b, the penetration number "NumP" increases by 1. If the final penetration number NumP≤L, points a and c are considered visible and a neighboring edge relationship is established; if the penetration number NumP>L, points a and c are considered invisible and no neighboring edge relationship is established, such as Figure 3 ; Performing this step can construct a time logic graph of the historical trajectory, as shown in formula (23), and the time logic graph containing all vehicles is obtained through iterative calculation, 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] Where E m Represents the connectivity between different trajectory points;

[0060] Step 3.2: Building a deep network for temporal feature aggregation graphs

[0061] The temporal logic graph and the spatial topology graph use the same structure of residual connection stacking GCN. The feature aggregation formula of the temporal perspective is as shown in formula (24), and its graph convolution is defined as formula (25). The dimension of the feature is adjusted through the fully connected layer and the extracted features are output through the temporal convolution layer. Finally, the output V of the temporal perspective is obtained. T ;

[0062]

[0063] Where p is the number of GCNs, is the pth GCN, is the weight matrix of the p-th layer; For network input, is the output of GCN; is the degree matrix of each node, Formula weight matrix;

[0064] Step 4: Implement multi-domain feature fusion and trajectory encoding / decoding. The process is as follows:

[0065] Output the features after multi-domain information extraction / aggregation V F 、V T 、V S According to formula (26), feature fusion is performed to obtain the fused feature matrix H. Subsequently, the GRU is used to encode and decode the features. Each time dimension of the fused features is used to generate corresponding hidden features through GRU, as shown in formula (27). After the encoding is completed, the hidden state H is decoded by the GRU decoder. hid Decode to obtain the expected future trajectory. The decoding process of GRU is shown in equations (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 coordinate of the trajectory point at time t, h t-1 H hid , r t 、h t To reset the gate and update the state of the gate, W r , W z is the weight matrix of the reset gate and update gate, W is The weight matrix in the generation step, ⊙ is element-wise 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, The predicted coordinates of vehicle m at the future time T+t.

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

[0075] (1) Dataset selection

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

[0077] The NGSIM dataset contains some noise because it is generated using video analysis and numerical differentiation methods. To this end, the experiment uses the Savitzky-Golay filter to smooth the vehicle position to improve the accuracy of the trajectory data. The dataset is divided into three 15-minute subsets, and finally divided into training sets, validation sets, and test sets. The original 10 frames of data are resampled into 5 frames, and the 3-second historical trajectory data of each sequence is used to predict the trajectory of the next 5 seconds.

[0078] (2) Performance comparison experiment

[0079] During the experiment, the Adam optimizer was selected for training, the learning rate was set to 0.001, the training batch size was 32, and the total number of training theories was 100. The input and output sizes of LSTM are 4 and 8 respectively, and k in argTopk is 5; the convolution kernel size of all temporal convolutional layers in the graph convolutional network is 1×3, the step size is set to 1, and appropriate padding is performed to maintain the shape of the feature. In order to avoid overfitting, 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, the size of all GRU network layers is 64 dimensions, and the hidden state is activated by a nonlinear rectified linear unit. The total loss function during training is shown in formula (35);

[0080]

[0081] The evaluation indicators of the experimental process are root mean square error (RMSE), average displacement error (ADE) and final displacement error (FDE), and their calculation formulas are shown in Equations (36)-(38);

[0082]

[0083]

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

[0085]

[0086] Continued

[0087]

[0088] Table 1

[0089] The results of comparative experiments show that in terms of single target vehicle trajectory prediction, the model's 1-second short-term prediction accuracy is improved by 70.33%, and the 5-second long-term prediction accuracy is improved by 55.58%; in terms of group vehicle prediction, the model's 1-second short-term prediction accuracy is improved by 48.13%, and the 5-second long-term prediction accuracy is improved by 35.78%. This shows that the proposed method is superior to the baseline method in both robustness and accuracy in single target and group vehicle prediction, and can achieve more accurate group trajectory prediction performance.

[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 space topology map were experimentally analyzed. Distance threshold: In the spatial perspective, the distance threshold d used when constructing the spatial topology map will affect the prediction accuracy of the model. The experiment tested the cases where the d value was 0, 15, 25, 35, and 50 feet. The results showed that the prediction accuracy of the model was the highest when d = 25; Value selection in CLPVG: In the construction of the time graph, the information based on acceleration can better reflect the time dependency of the vehicle trajectory. The experiment shows that the prediction accuracy of the CLPVG based on acceleration is about 3.68% higher than that of the model based on speed; The selection of k value in the frequency perspective: In the frequency perspective, the selection of the first k frequencies with the highest amplitude will affect the prediction effect. The experiment tested different k values ​​and found that the prediction accuracy was best when k = 5. Further increasing k will introduce feature redundancy and affect the model effect; Weighted directed space topology map, replace the weighted directed graph with a normal space topology map, and set all weights to 1. The experimental results show that the accuracy in trajectory prediction is reduced by 5.60%.

[0092] (4) Visual performance analysis

[0093] The visualization experiment shows the model's group vehicle trajectory prediction effect in different scenarios, such as Figure 4 , 5 The results show that the model can accurately predict the future trajectory while maintaining the lane, but the prediction accuracy decreases when the vehicle suddenly changes lanes or slows down. This shows that the model is robust for trajectory prediction in a stable state, but has certain limitations in dealing with sudden behaviors.

[0094] The contents described in the embodiments of this specification are merely enumerations of implementation forms of the inventive concept and are for illustrative purposes only. The protection scope of the present invention should not be considered to be limited to the specific forms described in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be thought of by ordinary technicians in this field based on the inventive concept.

Claims

1. A group vehicle trajectory prediction method based on a multi-domain perspective, characterized in that: The method comprises the following steps: Step 1: Extract frequency domain features: Map historical vehicle trajectory data from the time domain to the frequency domain through fast Fourier transform (FFT), calculate the amplitude of the spectrum data, select the key frequency component with the highest amplitude, and reconstruct the trajectory data into a two-dimensional tensor according to the selected frequency. The rows and columns represent the intra-cycle changes and inter-cycle changes respectively, so as to extract the intra-cycle and inter-cycle features; use the 2D convolutional neural network (CNN) combined with the long short-term memory network (LSTM) to further extract the frequency domain features; Step 2: Extract spatial domain features: The constructed weighted directed spatial topology graph is used 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 weights between them are calculated. The node features in the GCN aggregation topology graph are used to extract the spatial interaction features of multiple vehicles in the spatial domain to capture more detailed group vehicle motion characteristics. Step 3: Extract features in the temporal domain: By constructing a temporal logic graph based on the Circular Limited Penetrable Visibility Graph (CLPVG), we first determine the circular system between each pair of time points, and then calculate the number of penetrations on the connection path to determine whether it is visible; if the visibility condition is met, add the corresponding edge to the temporal logic graph; finally, perform feature aggregation on the temporal logic graph through the residual connected GCN; Step 4, feature fusion and trajectory generation: The features extracted from the frequency domain, space domain and time domain are fused in multiple domains to form a comprehensive feature representation for prediction. First, the frequency domain features, space domain features and time domain features are integrated into a high-dimensional feature vector through feature concatenation operation to achieve mutual complementation and coordinated expression of features in each domain; then, the fused feature vector is input into the GRU generator. The GRU effectively captures the temporal dependency in the fused features through its internal memory and gating mechanism. Through the encoding and decoding process, the GRU generator gradually outputs the trajectory point sequence of each future time step to achieve the prediction of the future trajectory of the group of vehicles.

2. The method for predicting group vehicle trajectories based on a multi-domain perspective as claimed in claim 1, characterized in that: The process of step 1 is as follows: Step 1.1: Trajectory frequency domain feature transformation The periodic features in vehicle trajectories are mined through frequency domain analysis to enhance the accuracy of trajectory prediction. First, FFT is applied to the historical trajectory data V to transform it from the time domain to the frequency domain to reveal the potential periodic features, as shown in Equation (5). Then, the first k main frequencies {f1,…,f k }, extract the significant periodic features of the trajectory, and select the main frequency as shown in formula (6). Each main frequency f i Corresponding to a cycle In order to further analyze the periodic changes, the trajectory data is reconstructed into a two-dimensional tensor V containing the intra-cycle and inter-cycle changes j , as shown in formula (7); A=Average[Amp(FFT(V))] (5) {f1,…,f k }=argTopk(A) (6) In the formula, FFT(V) means that the amplitude of each frequency component is obtained by Fourier transforming the data V; Amp means the amplitude of the calculated frequency component; Average means averaging the amplitude to obtain the overall spectrum distribution; Padding (V) fills the trajectory data to adapt to the period length p j and the frequency component f i , Reconstruct tensors according to period and frequency; Step 1.2: Frequency domain feature extraction network construction Reconstructed tensor A stacked 2D convolutional network (CNN) with residual connections is used for feature extraction to capture the detailed features within and between cycles, as shown in Equation (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. In the formula, and e represent the vector composed of four weight matrices and the number of layers of the convolutional network, respectively. e represents the e-th layer of the two-dimensional convolutional network, represents the initial value, Indicates the output result; Then will The time dimension is reconstructed into So that k features can be aggregated into a tensor, as shown in formula (9); the obtained features Further extraction is performed through LSTM to obtain feature information from the frequency domain perspective, as shown in formula (11); the output of LSTM is expressed as V F =[h1,h2,…,h T ]; In the formula, Trunc is the truncation operation, which is used to make The shape is consistent with the vehicle trajectory; W L is the LSTM weight matrix, h t is the hidden state vector at time step state t.

3. The method for predicting group vehicle trajectories based on a multi-domain perspective according to claim 1 or 2, characterized in that: The process of step 2 is as follows: Step 2.1: Construction of weighted directed spatial topology graph Through spatial domain analysis, the spatial topological relationship of vehicle interaction at any time is mined. The distance, relative position, speed and acceleration between vehicles are used to represent the relative state and actual distance of the vehicles, and the connection relationship between vehicles is constructed. The edges of the spatial graph are given 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 range d are set as neighbors. The actual relative distance between vehicles is defined as shown in equations (12)-(14), where P is marked as the relative position, and values ​​1 and 2 represent the front and rear vehicles in the same lane, and 3 and 4 represent the front and rear vehicles in different lanes. In the formula, Δx ij and Δy ij Respectively represent the lateral and longitudinal relative distances of vehicles i and j; l i and l j is the length of the two vehicles; Δw ij is the difference in vehicle width; The relative state of the vehicles is defined by the time required for vehicle i to collide with its adjacent vehicle, 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 formula (16); thus, the adjacency matrix containing the connection relationship of all vehicles is obtained, as shown in formula (17); formula (18) is the normalization processing of the weight, forming a normalized adjacency matrix, as shown in formula (19); Where, t coll is the time required for vehicles i and j to collide in their current state, assuming that their driving directions change to be consistent with the direction of the line connecting the two vehicles, Δa and Δv are the relative acceleration and relative velocity between the vehicles, respectively; represents the edge weight between vehicle i and vehicle j, where I is the unit matrix; Step 2.2: Construction of deep network for spatial feature aggregation graph The spatial features are extracted by stacking GCN with residual connections. GCN aggregates spatial features through the following formula 20. Graph convolution is an important part of GCN. The features of each node and its neighboring nodes can be aggregated through graph convolution operation. The principle formula is shown in formula (21). Finally, the fully connected layer is used to adjust the feature dimension, and the extracted features are output through the temporal convolution layer to obtain the spatial perspective feature output V S ; Where z is the number of GCNs; represents the zth GCN; is the weight matrix of the zth layer; They are the outputs of the first two layers of GCN; the 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 method for predicting group vehicle trajectories based on a multi-domain perspective as claimed in claim 3, characterized in that: The process of step 3 is as follows: 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, two time points, i.e., the circular system of trajectory points, are introduced, as shown in Equation (22); Where (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 "passed through" on the line from point a to point c; if the line connecting a and c can penetrate point b, the penetration number "NumP" increases by 1. If the final penetration number NumP≤L, points a and c are considered visible and a neighboring edge relationship is established; if the penetration number NumP>L, points a and c are considered invisible and no neighboring edge relationship is established; executing this step can construct a time logic graph of a historical trajectory, as shown in formula (23), and through iterative calculation, a time logic graph containing all vehicles is obtained, as shown in formula (24); L m =[V m ,E m +1]m∈{1,…,M} (23) L=[L 1 ,L 2 ,…,L M ] Where E m Indicates the connectivity between different trajectory points; Step 3.2 / Building a deep network for temporal feature aggregation graph The temporal logic graph and the spatial topology graph use the same structure of residual connection stacking GCN. The feature aggregation formula of the temporal perspective is as shown in formula (24), and its graph convolution is defined as formula (25). The dimension of the feature is adjusted through the fully connected layer and the extracted features are output through the temporal convolution layer, and finally the output V of the temporal perspective is obtained. T ; Where p is the number of GCNs, is the pth GCN, is the weight matrix of the p-th layer; For network input, is the output of GCN; is the degree matrix of each node, The weight matrix.

5. The method for predicting group vehicle trajectories based on a multi-domain perspective as claimed in claim 4, characterized in that: The process of step 4 is as follows: Output the features after multi-domain information extraction / aggregation V F 、V T 、V S According to Formula 26, feature fusion is performed to obtain the fusion feature matrix H; Subsequently, GRU is used to encode and decode the features. Each time dimension of the fused features is used to generate corresponding hidden features through GRU, as shown in formula (27). After encoding, the hidden state H is decoded by GRU decoder. hid Decode to obtain the expected future trajectory. The decoding process of GRU is shown in (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 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 coordinate of the trajectory point at time t, h t-1 H hid , r t 、h t To reset the gate and update the state of the gate, W r , W z is the weight matrix of the reset gate and update gate, W is The weight matrix in the generation step, ⊙ is element-wise 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, The predicted coordinates of vehicle m at the future time T+t.

Images