Surrounding vehicle trajectory prediction method and system based on cro-intentformer and integrated with vehicle driving intent

Abstract

Disclosed in the present invention are a trajectory prediction method and system based on Cro-IntentFormer and integrated with a vehicle driving intent. The method comprises: first, preprocessing vehicle trajectories collected by a sensor to obtain original data that meets a model input condition; constructing a physical relationship graph using vehicles as nodes and the distance between the vehicles as the basis for whether there is an edge between the two vehicle nodes, and inputting the physical relationship graph and the original data into a spatiotemporal feature extraction module to obtain spatiotemporal features of the trajectories; inputting a spatiotemporal feature matrix into an intent prediction module to obtain predicted intents of the vehicles; re-constructing a semantic relationship graph on the basis of intent information output by the intent prediction module, and inputting the semantic relationship graph and the original data into the spatiotemporal feature extraction module to obtain semantic features of the trajectories; and integrating the spatiotemporal features and the semantic features of the trajectories, and inputting the integrated features into a decoder to obtain predicted trajectories of surrounding vehicles. The present invention takes into consideration the similarity between trajectories of vehicles having similar behavior intents, improving the interpretability and the prediction precision of a model.

Description

Vehicle trajectory prediction method and system based on Cro-IntentFormer integrating vehicle driving intention Technical Field

[0001] The present application relates to the field of intelligent driving technology, and specifically to a vehicle trajectory prediction method and system based on Cro-IntentFormer that integrates vehicle driving intentions. Background Art

[0002] With the growing number of autonomous vehicles in my country, human-machine co-driving has become a trend in future transportation. In complex and ever-changing traffic environments, autonomous vehicles' ability to understand complex driving scenarios and predict the future intentions and trajectories of surrounding traffic participants is crucial to ensuring their safety and serves as the foundation for downstream trajectory planning tasks.

[0003] Current trajectory prediction methods typically only consider the impact of traffic participants within a certain distance of the target vehicle on its future trajectory. Furthermore, when modeling the interaction between the target vehicle and surrounding vehicles, they prioritize the distance between them. However, they ignore potential factors, such as the influence of vehicle driving intention on the strength of inter-vehicle interactions, and fail to fully extract the interaction characteristics between the target vehicle and surrounding vehicles. Furthermore, current methods typically separate the prediction of vehicle driving intention from the prediction of vehicle trajectory, failing to effectively incorporate the prediction of vehicle driving intention into the trajectory prediction task. Furthermore, most existing trajectory prediction methods apply attention mechanisms between time frames, ignoring the connections between adjacent trajectory segments.

[0004] Summary of the Invention

[0005] This paper provides a method and system for predicting vehicle trajectories based on Cro-IntentFormer, integrating vehicle driving intent. This method explicitly considers the impact of vehicle driving intent on future trajectories, incorporating both inter-vehicle distances and similarities in behavioral intent into the interactive modeling of inter-vehicle relationships. This method enables real-time prediction of future trajectories during driving, effectively ensuring safe driving in complex and changing traffic environments and providing a reference for downstream trajectory planning tasks.

[0006] A vehicle trajectory prediction method based on Cro-IntentFormer and integrating vehicle driving intention is characterized by comprising the following steps:

[0007] S1. Preprocessing the raw data of the vehicle trajectory collected by the sensor;

[0008] S2. Construct a physical relationship graph with vehicles as nodes and the distance between vehicles as the basis for whether there is an edge between two vehicle nodes, and input it and the original data into the spatio-temporal feature extraction module to obtain the spatio-temporal features of the trajectory;

[0009] S3. Input the spatio-temporal feature matrix into the intention prediction module to obtain the predicted intention of the vehicle;

[0010] S4. Reconstruct a semantic relationship graph according to the intention information output by the intention prediction module, and input it and the original data into the spatio-temporal feature extraction module to obtain the semantic features of the trajectory;

[0011] S5. Fuse the spatio-temporal features and semantic features of the trajectory, and input them into the decoder to obtain the predicted trajectories of the vehicles around the target vehicle.

[0012] Furthermore, the preprocessing of the vehicle trajectory in step S1 includes:

[0013] Standardize the format of the trajectory data collected by on-vehicle sensors, and use timestamps and vehicle IDs to identify the trajectory information;

[0014] Clean the standardized trajectory data, remove outliers and duplicate data, fill in missing values, and perform noise reduction on the cleaned data;

[0015] Label the intention of the trajectory according to the heading angle and longitudinal and lateral speeds of the vehicle at each time step. The behavioral intentions of the vehicle include going straight a1, changing lanes to the left a2, and changing lanes to the right a3;

[0016] Use the time window T to divide the processed trajectory data, and the divided data is used as the original data input to the model.

[0017] Furthermore, the specific steps for constructing the physical relationship graph in step S2 are as follows:

[0018] Select the vehicles observed at time t as the nodes V of the graph

[0021] ,

[0020] , ij ,

[0019] ,<​​​​​​​​​​​​​​​​​​​

[0022] The original data at time t Input into the time information fusion network in sequence to learn the time dependency of the vehicle's own trajectory and output the feature matrix B∈R after feature extraction N×L×dmodel ;

[0023] in, is the historical status information of vehicle i, T h is the preset historical time window, L is the number of time segments after network coding; the state information of vehicle i at time t The position coordinates of the vehicle at time t Intent Information Heading angle and speed Feature composition;

[0024] After obtaining the feature matrix B that integrates the time dependency of the vehicle’s own trajectory through the time information fusion network, the adjacency relationship of the vehicle nodes in the physical relationship graph G1 is re-established in each time segment, and the attention operation is performed on the feature matrix B between vehicles to learn the spatial relationship of the trajectory, and the trajectory spatiotemporal feature matrix Z of all nodes in the physical relationship graph G1 is obtained. 1 ∈R N×L×dmodel .

[0025] Furthermore, the calculation process of the time information fusion network is as follows:

[0026] The data points of each vehicle’s historical state information h in each feature dimension are divided into seg time segments, namely: h i,c ={h t,c |(i-1)×L seg <t≤i×L seg}

[0027] Where C is the number of features of vehicle trajectories in the original data, h i,c The length of feature c is L seg The i-th time segment of ;

[0028] And use a learnable linear matrix and Perform feature embedding and position encoding on the obtained time segment to obtain the encoded feature vector m i,c ,Right now:

[0029] Among them, L is the historical time step L seg The total number of time segments after division;

[0030] The encoded feature vector m is subjected to multi-head attention calculation and residual connection in the time dimension and feature dimension respectively to obtain the feature matrix m that combines the time segment and input feature dimension dim ∈R L×C×dmodel ,Right now:

[0031] In the formula, MSA(Q,K,V) represents the multi-head attention operation, LayerNorm represents the layer normalization operation, MLP represents the multi-layer feedforward neural network, and m i,: ∈R C×dmodel Represents the feature matrix of all feature dimensions of time segment i, m :,c ∈R L×dmodel Represents the feature matrix of all time segments of feature dimension c;

[0032] Finally, the feature matrix m that combines the time segment and input feature dimensions is dim ∈R L×C×dmodel Perform additive aggregation operations on the feature dimension to obtain the feature matrix B that integrates the time dependency of the trajectory of vehicle i i ∈R L×dmodel .

[0033] Furthermore, the specific steps of step S3 are:

[0034] The trajectory spatiotemporal feature matrix Z of all nodes in the physical relationship graph G1 1 ∈R N×L×dmodel Perform additive aggregation operations on the time dimension;

[0035] The trajectory spatiotemporal feature matrix after the addition aggregation operation is passed through a fully connected network and normalized using the Softmax function to obtain the predicted intention vector ω of vehicle i i ={atten1, atten2, atten3}; atten1, atten2, atten3 represent the probabilities of the vehicle going straight, changing lanes to the left, and changing lanes to the right, respectively.

[0036] Furthermore, the specific steps of constructing the semantic relationship graph and obtaining the semantic features of the trajectory in step S4 are:

[0037] Filter the vehicles observed at time t as nodes V of the graph i ;

[0038] According to the predicted intention vector ω of vehicle i i , select the behavior with the highest probability as the vehicle’s future intention a;

[0039] Explicitly consider the impact of vehicle driving intention on the vehicle's future trajectory, establish connections between vehicle nodes with the same intention, and obtain the semantic adjacency matrix A2;

[0040] According to the connection relationship between nodes, a semantic relationship graph G2 = {V, R 2};

[0041] The semantic relationship graph G2 and the original data obtained in step S1 are simultaneously input into the spatiotemporal feature extraction module to obtain the trajectory semantic feature matrix Z of all nodes in the semantic relationship graph G2. 2 ∈R N×L×dmodel .

[0042] Furthermore, the specific steps of step S5 are:

[0043] The importance w1, w2 of the spatiotemporal features and semantic features of the trajectory to all vehicle nodes are obtained respectively, namely:

[0044] Where q represents the learnable semantic-level attention vector, and tanh represents the hyperbolic tangent activation function;

[0045] Normalize the importance w1, w2 of the spatiotemporal features and semantic features of the trajectory to all vehicle nodes to obtain the feature weights β of the spatiotemporal features and semantic features of the trajectory i ,Right now:

[0046] The spatiotemporal features and semantic features of the trajectory are weighted and summed to obtain the feature matrix J of the spatiotemporal information and semantic information of the fused trajectory, that is: J = β1Z 1 +β2Z 2

[0047] The feature matrix J is input into the decoder to obtain the predicted trajectory F of the vehicles around the target vehicle = {f1, f2, ..., f n},in is the future trajectory of vehicle i, T f is a preset future time window.

[0048] The prediction system for the vehicle trajectory prediction method based on Cro-IntentFormer and integrating vehicle driving intention is characterized by comprising:

[0049] The information collection and processing device, including on-board sensors, roadside sensors and a data processing module, is used to obtain the position and speed information of the vehicle and surrounding vehicles in real time, and standardize, clean and pre-process the collected data to obtain raw data that meets the input conditions of the surrounding vehicle trajectory prediction model;

[0050] The vehicle trajectory prediction model includes a spatiotemporal feature extraction module, an intention prediction module, a feature fusion module and a decoder. During vehicle driving, based on the raw data obtained by the information acquisition and processing device, it explicitly considers the impact of the vehicle's driving intention on the vehicle's future trajectory and predicts the future trajectory of vehicles around the vehicle.

[0051] Furthermore, the prediction system also includes a danger warning device, which warns vehicles that may have a collision risk with the future driving path of the own vehicle based on the future trajectory of the surrounding vehicle obtained by the surrounding vehicle trajectory prediction model.

[0052] The present invention proposes a method and system for predicting vehicle trajectories based on Cross-IntentFormer, which integrates vehicle driving intentions. The system uses the CrossFormer model to extract the temporal features of vehicle trajectories. It is a neural network model based on the attention mechanism that can effectively capture dependencies across time segments and across input feature dimensions, and can fully learn the information between adjacent trajectory segments. At the same time, the system explicitly considers the impact of vehicle driving intentions on the vehicle's future trajectory, and incorporates the similarity between the distance between vehicles and the behavioral intentions into the interactive modeling of vehicle relationships, thereby improving the model's interpretability and prediction accuracy, and providing a strong guarantee for the safe driving of vehicles in complex and changing traffic environments.

[0053] The beneficial effects of the present invention are:

[0054] (1) A vehicle trajectory prediction method that explicitly considers vehicle driving intentions is proposed. The distance between vehicles and the similarity between behavioral intentions are simultaneously incorporated into the interaction modeling of vehicle relationships. The similarity of trajectories between vehicles with similar intentions is deeply explored, and the prediction of vehicle driving intentions and vehicle trajectory are coupled to improve the prediction accuracy of the model.

[0055] (2) The CrossFormer network is used as a temporal information fusion network to extract the temporal dependencies of vehicle trajectories. The network's time slice partitioning mechanism is used to fully learn the information between adjacent trajectory segments. An attention operation is then performed on the trajectories between vehicles within each time segment to learn the spatial relationships between the trajectories. This application extracts the spatiotemporal features of vehicle trajectories in layers, making better use of the data and improving the interpretability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] FIG1 is a block diagram of the vehicle trajectory prediction system based on Cro-IntentFormer and integrating vehicle driving intention according to the present invention.

[0057] FIG2 is a flow chart of the vehicle trajectory prediction method based on Cro-IntentFormer and integrating vehicle driving intention according to the present invention.

[0058] Figure 3 is the architecture diagram of the vehicle trajectory prediction model.

[0059] FIG4 is a diagram showing the “circumferential vehicle trajectory” interface of the circumferential vehicle trajectory prediction system of the invention.

[0060] FIG5 is a diagram showing the “danger warning” interface of the vehicle trajectory prediction system according to the present invention. DETAILED DESCRIPTION

[0061] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited thereto.

[0062] As shown in FIG1 , the vehicle trajectory prediction system based on Cro-IntentFormer and integrating vehicle driving intentions of the present invention includes an information collection and processing device, a vehicle trajectory prediction device, and a danger warning device.

[0063] The information collection and processing device includes onboard sensors, roadside sensors, and a data processing module. It is used to acquire the real-time location and speed information of the vehicle and surrounding vehicles. Trajectory information is identified using timestamps and vehicle IDs. The standardized trajectory data is cleaned to remove outliers and duplicates, fill in missing values, and perform noise reduction. The trajectory is then annotated with intent based on the vehicle's heading angle and lateral and longitudinal speed at each time step. Vehicle behavioral intentions include going straight (a1), changing lanes to the left (a2), and changing lanes to the right (a3). Finally, the processed trajectory data is partitioned using a time window T to obtain raw data that meets the input requirements of the surrounding vehicle trajectory prediction model.

[0064] The surrounding vehicle trajectory prediction device includes a spatiotemporal feature extraction module, an intention prediction module, a feature fusion module and a decoder. During vehicle driving, based on the raw data obtained by the information acquisition and processing device, it explicitly considers the impact of the vehicle's driving intention on the vehicle's future trajectory and predicts the future trajectory of vehicles around the vehicle; and displays the prediction results on the vehicle's central control screen.

[0065] The hazard warning device, based on the future trajectory of the surrounding vehicle obtained by the surrounding vehicle trajectory prediction model, warns vehicles that may have a risk of collision with the future driving path of the own vehicle, provides a reference for human drivers, and serves as a basis for trajectory planning tasks downstream of the vehicle.

[0066] The vehicle trajectory prediction method based on Cro-IntentFormer and integrated with vehicle driving intention described in the present invention is shown in FIG2 . The processing flow is as follows:

[0067] S1. Preprocess the vehicle trajectories collected by the sensor to obtain raw data that meets the model input conditions;

[0068] S2. Take vehicles as nodes and the distances between vehicles as the basis for whether there is an edge between two vehicle nodes to construct a physical relationship graph, and input it and the original data into the spatio-temporal feature extraction module to obtain the spatio-temporal features of the trajectory;

[0069] S3. Input the spatio-temporal feature matrix into the intention prediction module to obtain the predicted intention of the vehicle;

[0070] S4. Reconstruct the semantic relationship graph according to the intention information output by the intention prediction module, and input it and the original data into the spatio-temporal feature extraction module to obtain the semantic features of the trajectory;

[0071] S5. Fuse the spatio-temporal features and semantic features of the trajectory, and input them into the decoder to obtain the predicted trajectories of the vehicles around the target vehicle.

[0072] The CrossFormer network is a neural network model based on the attention mechanism that can effectively capture the dependencies across time segments and across input feature dimensions, and can fully learn the dependencies between adjacent trajectory segments. In this invention, it is used as the trajectory time information fusion network. After using this network to learn the time dependencies of the vehicle trajectories themselves, an attention operation is performed on the trajectories between vehicles within each time segment to learn the spatial dependencies of the trajectories. In this invention, the spatio-temporal features of the vehicle trajectories are extracted layer by layer, maximizing the interpretability of the model.

[0073] Specifically, referring to the architecture diagram of the surrounding vehicle trajectory prediction model shown in Figure 3, input the original data preprocessed by the information acquisition and processing device into the surrounding vehicle trajectory prediction model for the following prediction processing:

[0074] S2.1 Construct the physical relationship graph G1.

[0075] Select the vehicles observed at time t as the nodes V of the graph i , calculate the physical distance d between vehicles at time t according to the preset physical distance threshold D ij , if d ij < D, it is considered that there is an edge e between node i and node j ij , and establish a physical adjacency matrix A1 according to the physical distances between vehicles; establish a physical relationship graph G1 = {V, E 1} according to the connection relationship between nodes and nodes.

[0076] S2.2 Input the physical relationship graph G1 and the original data into the spatio-temporal feature extraction module to obtain the spatio-temporal features of the trajectory.

[0077] The original data of the vehicles at time t in the node set of the physical relationship graph G1 Input into the time information fusion network in sequence to learn the time dependency of the vehicle's own trajectory and output the feature matrix B∈R after feature extraction N×L×dmodel ,in is the historical status information of vehicle i, T h is the preset historical time window. The status information of vehicle i at time t The position coordinates of the vehicle at time t Intent Information Heading angle and speed composition;

[0078] The calculation method for learning the time dependency of the vehicle's own trajectory using the time information fusion network is as follows:

[0079] The data points of each vehicle’s historical state information h in each feature dimension are divided into seg time segments, namely: h i，c ={h t，c |(i-1)×L seg <t≤i×L seg}

[0080] Where C is the number of features of vehicle trajectories in the original data, h i,c The length of feature c is L seg ith time segment of ; and using a learnable linear matrix and Perform feature embedding and position encoding on the obtained time segment to obtain the encoded feature vector m i,c ,Right now:

[0081] Where L is the historical time step L seg The total number of time segments after division;

[0082] The encoded feature vector m is subjected to multi-head attention calculation and residual connection in the time dimension and feature dimension respectively to obtain the feature matrix m that combines the time segment and input feature dimension dim ∈R L×C×dmodel ,Right now:

[0083] In the formula, MSA(Q,K,V) represents the multi-head attention operation, LayerNorm represents the layer normalization operation, MLP represents the multi-layer feedforward neural network, and m i,: ∈R C×dmodelRepresents the feature matrix of all feature dimensions of time segment i, m :,c ∈R L×dmodel Represents the feature matrix of all time segments of feature dimension c;

[0084] Finally, the feature matrix m that combines the time segment and input feature dimensions is dim ∈R L×C×dmodel Perform additive aggregation operations on the feature dimension to obtain the feature matrix B that integrates the time dependency of the trajectory of vehicle i i ∈R L×dmodel .

[0085] The feature matrix B∈R that integrates the time dependency of the vehicle's own trajectory through the time information fusion network is obtained N×L×dmodel Finally, the adjacency relationship is re-established based on the physical distance between vehicles in the last time step of each time segment, and the physical relationship diagram between vehicles in each time segment is obtained. in is the physical relationship diagram between vehicles in the i-th time segment. The adjacency relationship and feature matrix B of the vehicle are used to perform attention operations between vehicles in each time segment to learn the spatial dependency of the trajectory. Finally, the result tensors calculated in each time segment are stacked in the time dimension to obtain the trajectory spatiotemporal feature matrix Z of all nodes in the physical relationship graph G1 1 ∈R N×L×dmodel .

[0086] S3. Input the spatiotemporal feature matrix into the intention prediction module to obtain the predicted intention of the vehicle.

[0087] The trajectory spatiotemporal feature matrix Z of all nodes in the physical relationship graph G1 1 ∈R N×L×dmodel Perform an additive aggregation operation on the time dimension, pass the trajectory spatiotemporal feature matrix after the additive aggregation operation through a fully connected network, and use the Softmax function to normalize it to obtain the predicted intention vector ω of vehicle i. i ={atten1, atten2, atten3}. atten1, atten2, atten3 represent the probability of the vehicle going straight, changing lanes to the left, and changing lanes to the right, respectively.

[0088] Step S4.1: Constructing a semantic relationship graph: The present invention explicitly considers the impact of vehicle driving intentions on future vehicle trajectories, incorporates the distance between vehicles and the similarity between behavioral intentions into the interaction modeling of vehicle relationships, and couples the prediction of vehicle driving intentions with the prediction of vehicle trajectories. The vehicle observed at time t is selected as the node V of the graph. i , according to the predicted intention vector ω of vehicle ii , select the behavior with the highest probability as the future intention a of the vehicle, establish a connection relationship between vehicle nodes with the same intention, and establish a semantic relationship graph G2 = {V,E 2}.

[0089] Step S4.2: Obtain the semantic features of the trajectory: Input the semantic relationship graph G2 and the original data obtained in step S1 into the spatiotemporal feature extraction module at the same time to obtain the trajectory semantic feature matrix Z of all nodes in the semantic relationship graph G2. 2 ∈R N×L×dmodel .

[0090] S5.1 Fusion of spatiotemporal and semantic features of trajectories: Obtain the importance w1, w2 of spatiotemporal and semantic features of trajectories to all vehicle nodes, namely:

[0091] Where q represents the learnable semantic-level attention vector, and tanh represents the hyperbolic tangent activation function;

[0092] Normalize the importance w1, w2 of the spatiotemporal features and semantic features of the trajectory to all vehicle nodes to obtain the feature weights β of the spatiotemporal features and semantic features of the trajectory i ,Right now:

[0093] The spatiotemporal features and semantic features of the trajectory are weighted and summed to obtain the feature matrix J of the spatiotemporal information and semantic information of the fused trajectory, that is: J = β1Z 1 +β2Z 2 .

[0094] S5.2 Decoding to obtain the predicted trajectory of vehicles around the target vehicle: Input the feature matrix J into the decoder to obtain the predicted trajectory F of vehicles around the target vehicle = {f1, f2, ..., f n},in is the future trajectory of vehicle i, T f is a preset future time window.

[0095] Figure 4 shows a specific example of the "Circumvehicle Trajectory" interface displayed by the vehicle trajectory prediction system of the present invention. The interface consists of three parts: Frame 1 displays the current time, weather conditions and temperature information for the city where the vehicle is located, and the remaining battery percentage; Frame 2 displays the current real-time information of the vehicle and surrounding vehicles, as well as the trajectory of surrounding vehicles predicted by the vehicle trajectory prediction system for a period of time in the future. To distinguish between the vehicle and surrounding vehicles, the vehicles in the scene are marked in orange and gray, respectively. The future trajectory of surrounding vehicles, if safe, is indicated by a green dashed line; Frame 3 displays the vehicle's current real-time navigation.

[0096] Figure 5 shows an example of the "Danger Warning" interface displayed by the surrounding vehicle trajectory prediction system of the present invention. When the system indicates that a vehicle's future trajectory, as determined by the system, may pose a collision risk with the vehicle's path, the future trajectory of the corresponding risky vehicle in the interface changes from green to red, and a red light and a warning box appear next to the risky vehicle.

[0097] The above series of descriptions are only specific descriptions of feasible implementation methods of the present application and do not limit the scope of protection of the present application. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention are within the scope of protection of the present invention.

Claims

1. A vehicle trajectory prediction method based on Cro-IntentFormer and integrating vehicle driving intention, characterized by: The following steps are involved: S1. Raw data obtained by preprocessing the vehicle trajectory collected by the sensor; S2. Taking vehicles as nodes and the distance between vehicles as the basis for whether there is an edge between two vehicle nodes, a physical relationship graph is constructed, and the graph and the original data are input into the spatiotemporal feature extraction module to obtain the spatiotemporal features of the trajectory; S3. Input the spatiotemporal feature matrix into the intention prediction module to obtain the predicted intention of the vehicle; S4. Reconstruct the semantic relationship graph according to the intention information output by the intention prediction module, and input it and the original data into the spatiotemporal feature extraction module to obtain the semantic features of the trajectory; S5. The spatiotemporal features and semantic features of the trajectory are integrated and input into the decoder to obtain the predicted trajectory of the vehicles around the target vehicle.

2. The method for predicting vehicle trajectory based on Cro-IntentFormer and integrating vehicle driving intention according to claim 1 is characterized in that: The preprocessing of the vehicle trajectory in step S1 includes: Standardize the format of trajectory data collected by vehicle sensors and use timestamps and vehicle IDs to identify trajectory information; Perform data cleaning on the standardized trajectory data, remove outliers and duplicate data, fill in missing values, and perform noise reduction on the cleaned data; The trajectory is labeled with intentions based on the vehicle's heading angle and lateral and longitudinal speed at each time step. The vehicle's behavior intentions include going straight a1, changing lanes to the left a2, and changing lanes to the right a3; The processed trajectory data is divided into two parts using the time window T, and the divided data is used as the original data of the input model.

3. The method for predicting vehicle trajectory based on Cro-IntentFormer and integrating vehicle driving intention according to claim 1, characterized in that: The specific steps of constructing the physical relationship diagram in step S2 are: Filter the vehicles observed at time t as nodes V of the graph i ; Set a physical distance threshold D and calculate the physical distance d between vehicles at time t ij If d ij < D, it is considered that There is an edge e between nodes i and j ij , and establish a physical adjacency matrix A1 based on the physical distance between vehicles; establish a physical relationship graph G1 = {V, E 1 }.

4. The method for predicting vehicle trajectory based on Cro-IntentFormer and integrating vehicle driving intention according to claim 3 is characterized in that: In step S2, the specific steps of inputting the physical relationship graph G1 and the original data into the spatiotemporal feature extraction module to obtain the spatiotemporal features of the trajectory are: The original data at time t Input into the time information fusion network in sequence to learn the time dependency of the vehicle's own trajectory and output the feature matrix B∈R after feature extraction N×L×dmodel ; in, is the historical status information of vehicle i, T h is the preset historical time window, L is the number of time slices after network encoding; the state information of vehicle i at time t The position coordinates of the vehicle at time t Intent Information Heading Angle and speed Characteristic composition; After obtaining the feature matrix B that integrates the time dependency of the vehicle's own trajectory through the time information fusion network, the adjacency relationship of the vehicle nodes in the physical relationship graph G1 is reestablished in each time segment, and the feature matrix B is subjected to attention operation between vehicles to learn the spatial relationship of the trajectory, and the trajectory spatiotemporal feature matrix Z of all nodes in the physical relationship graph G1 is obtained. 1 ∈R N×L×dmodel .

5. The method for predicting vehicle trajectory based on Cro-IntentFormer and integrating vehicle driving intention according to claim 4 is characterized in that: The calculation process of the time information fusion network is as follows: The data points of each vehicle's historical state information h in each feature dimension are divided into time segments of length Lseg, that is: h i,c ={h t,c |(i-1)×L seg <t≤i×L seg } Where C is the number of features of vehicle trajectories in the original data, h i,c The length of feature c is L seg The i-th time segment of ; And use a learnable linear matrix and Perform feature embedding and position encoding on the obtained time segment to obtain the encoded feature vector m i,c ,Right now: Among them, L is the historical time step L seg The total number of time segments after division; The encoded feature vector m is subjected to multi-head attention calculation and residual connection in the time dimension and feature dimension respectively, and a feature matrix m is obtained which combines the time segment and the input feature dimension. dim ∈R L×C×dmodel ,Right now: In the formula, MSA(Q, K, V) represents multi-head attention operation, LayerNorm represents layer normalization operation, and MLP represents multi-head attention operation. Layer feedforward neural network, m i,: ∈R C×dmodel Represents the feature matrix of all feature dimensions of time segment i, m :,c ∈R L×dmodel Represents the feature matrix of all time segments of feature dimension c; Finally, the feature matrix m that combines the time segment and input feature dimensions is dim ∈R L×C×dmodel Performing additive aggregation operations on the feature dimension obtains the feature matrix B that integrates the trajectory time dependency of vehicle i i ∈R L×dmodel .

6. The method for predicting vehicle trajectory based on Cro-IntentFormer and integrating vehicle driving intention according to claim 1, characterized in that: The specific steps of step S3 are: The trajectory spatiotemporal feature matrix Z of all nodes in the physical relationship graph G1 1 ∈R N×L×dmodel Perform additive aggregation operations on the time dimension; The trajectory spatiotemporal feature matrix after the addition aggregation operation is normalized using the Softmax function after passing through a layer of fully connected network to obtain the predicted intention vector ω of vehicle i i ={atten1, atten2, atten3}; atten1, atten2, atten3 represent the probabilities of the vehicle going straight, changing lanes to the left, and changing lanes to the right, respectively.

7. The method for predicting vehicle trajectory based on Cro-IntentFormer and integrating vehicle driving intention according to claim 1, characterized in that: The specific steps of constructing the semantic relationship graph and obtaining the semantic features of the trajectory in step S4 are: Filter the vehicles observed at time t as nodes V of the graph i ; According to the predicted intention vector ω of vehicle i i , select the behavior with the highest probability as the vehicle’s future intention a; The influence of vehicle driving intention on the future trajectory of the vehicle is explicitly considered, and a connection relationship is established between vehicle nodes with the same intention to obtain the semantic adjacency matrix A2; According to the connection relationship between nodes, a semantic relationship graph G2 = {V, E 2 }; The semantic relationship graph G2 and the original data obtained in step S1 are simultaneously input into the spatiotemporal feature extraction module to obtain the trajectory semantic feature matrix Z of all nodes in the semantic relationship graph G2 2 ∈R N×L×dmodel .

8. The method for predicting vehicle trajectory based on Cro-IntentFormer and integrating vehicle driving intention according to claim 1, characterized in that: The specific steps of step S5 are: The importance w1, w2 of the spatiotemporal features and semantic features of the trajectory to all vehicle nodes are obtained respectively, namely: Where q represents the learnable semantic-level attention vector, and tanh represents the hyperbolic tangent activation function; Normalize the importance w1, w2 of the spatiotemporal features and semantic features of the trajectory to all vehicle nodes to obtain the feature weights β of the spatiotemporal features and semantic features of the trajectory i ,Right now: The spatiotemporal features and semantic features of the trajectory are weighted and summed to obtain the feature matrix of the spatiotemporal information and semantic information of the fused trajectory. Array J, that is: J=β1Z 1 +β2Z 2 The feature matrix J is input into the decoder to obtain the predicted trajectory F of the vehicles around the target vehicle = {f1, f2, ..., f n },in is the future trajectory of vehicle i, T f is a preset future time window.

9. A prediction system for the vehicle trajectory prediction method based on Cro-IntentFormer and integrating vehicle driving intention as described in any one of claims 1 to 8, characterized in that: include: The information collection and processing device includes on-board sensors, road-side sensors and data processing modules, which are used to obtain the position and speed information of the vehicle and surrounding vehicles in real time, and to standardize, clean and pre-process the collected data to obtain the original data that meets the input conditions of the surrounding vehicle trajectory prediction model; The vehicle trajectory prediction model includes a spatiotemporal feature extraction module, an intention prediction module, a feature fusion module and a decoder. During the vehicle driving process, based on the original data obtained by the information acquisition and processing device, the influence of the vehicle driving intention on the future trajectory of the vehicle is explicitly considered, and the future trajectory of the vehicles around the vehicle is predicted.

10. The prediction system according to claim 9, characterized in that It also includes a danger warning device, which warns vehicles that may have a risk of collision with the future driving path of the vehicle based on the future trajectory of the surrounding vehicle obtained by the surrounding vehicle trajectory prediction model.

Images