An Integrated Prediction-Planning Method for Autonomous Driving Based on Traffic Heterogeneous Graphs

Abstract

This invention relates to an integrated prediction-planning method for autonomous driving based on traffic heterogeneous graphs, belonging to the field of autonomous vehicle technology. The method includes: S1: Modeling the dynamic features of agents and the interaction features between traffic participants, and designing lane graph node feature representations based on graph convolutional networks; S2: Employing an encoder-decoder architecture, capturing the interaction features between lane nodes and agents through a graph attention mechanism, and constructing a multi-agent trajectory prediction model based on traffic heterogeneous graphs after feature fusion; S3: Based on the future position information of surrounding vehicles output by the trajectory prediction model, designing an objective function and various constraints to optimize the vehicle's actions. Compared with traditional planning methods, this invention offers superior prediction performance, safety, and driving efficiency.

Description

Technical Field

[0001] This invention belongs to the field of autonomous vehicle technology and relates to an autonomous driving prediction-planning integration method based on traffic heterogeneous graphs. Background Technology

[0002] In recent years, autonomous driving has become a research hotspot in the fields of artificial intelligence and transportation. Autonomous driving systems aim to enable vehicles to autonomously perceive, make decisions, and control themselves in complex road environments. Trajectory prediction, by analyzing road scenes, traffic rules, and the dynamic behavior of traffic participants, infers the future movement trend of the target, providing reliable prior information for downstream processes. Motion planning, after receiving the predicted information, combines vehicle dynamics constraints, the driving environment, and the driving objective to generate a future trajectory that meets safety, comfort, and efficiency requirements. Currently, mainstream motion planning methods can be divided into rule-based, optimization-based, and learning-based methods. While traditional rule-based methods are logically clear and highly interpretable, they rely on expert experience, require manual task decomposition and definition of triggering conditions, and struggle to handle multi-agent interactions and dynamic environmental changes. Optimization-based methods define the driving task as a mathematical optimization problem, solving it by setting an objective function and constraints. Learning-based methods utilize deep learning and reinforcement learning techniques to implicitly learn target strategies from data, enabling flexible adaptation to complex traffic scenarios.

[0003] Regardless of whether it's a modular or end-to-end autonomous driving framework, the two core modules—trajectory prediction and motion planning—must be tightly coupled in their functional implementation to improve feature transfer efficiency and reduce accumulated errors during information transmission. However, current mainstream motion planning methods do not adequately model traffic scenarios using upstream information and fail to fully utilize the rich information provided by the trajectory prediction module. Therefore, there is an urgent need for an efficient prediction-planning integration method that fuses interactive prediction information to deeply explore the coupling mechanism between prediction information and motion planning, thereby further improving the driving safety and traffic efficiency of autonomous vehicles. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide an integrated prediction-planning method for autonomous driving based on traffic heterogeneous graphs. This method effectively models multimodal interactions in complex interactive scenarios and efficiently integrates surrounding vehicle prediction information. Through optimization, it ensures that the vehicle makes safe and reasonable motion plans. Compared with traditional planning methods, this invention offers superior prediction performance, safety, and driving efficiency.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] An autonomous driving prediction-planning integration method based on traffic heterogeneous graphs specifically includes the following steps:

[0007] S1: Model the dynamic features of intelligent agents and the interaction features between traffic participants, and design a lane graph node feature representation based on graph convolutional networks;

[0008] S2: Adopting an encoder-decoder architecture, it captures the interaction features between lane nodes and agents through a graph attention mechanism, and constructs a multi-agent trajectory prediction model based on traffic heterogeneous graphs after feature fusion.

[0009] S3: Based on the future location information of surrounding vehicles output by the trajectory prediction model, design an objective function and various constraints to optimize the vehicle's actions.

[0010] Furthermore, step S1 specifically includes the following steps:

[0011] S11: Define current traffic information, including lane information. and the historical trajectory information of n agents at time t

[0012]

[0013] Where t0 is the length of the historical trajectory; This represents the motion state information of the i-th agent at time t. These represent the vehicle's current coordinates, velocity components in the x and y directions, and heading angle, respectively.

[0014] S12: For agent node features, a recurrent neural network-based method is used to dynamically encode the three traffic participant types: vehicles, bicycles, and pedestrians; for agent A... i The dynamic characteristics are encoded as follows:

[0015]

[0016] Among them, f i t This represents the dynamic characteristics after encoding the vehicle's historical trajectory information, where GRU(·) indicates encoding via a recurrent neural network;

[0017] S13: Model the interaction features of agents using graph neural networks. First, for each agent, design dynamic filtering conditions to select neighboring nodes:

[0018]

[0019] in, V is the radius of the circle to be selected. i t L represents the current vehicle speed at the target node. vehicle Let λ be the length of the vehicle.R A constant value chosen based on experience;

[0020] S14: After constructing the graph structure data representation, a graph attention mechanism is introduced to further enhance the interaction features between vehicle nodes; firstly, for the target node i and its selected neighbor nodes j, the desired feature is to obtain the concatenation feature z of its directed edges. i,j :

[0021] z i,j =ReLU(W a [h i ||e i,j ||h j ])

[0022] Among them, h i ,h j For the encoding features of the node, e i,j W represents the edge embedding feature from node i to node j. a Let be the attention linear transformation matrix, and || be the concatenation of features;

[0023] S15: After concatenating the features of each edge of the target node, normalize them using softmax to obtain their attention scores, and then update the features of the target node:

[0024]

[0025] Among them, A i This represents the updated target node characteristics. Let α represent the number of neighbors of node i, k represent the number of neighboring agent nodes within the target node i's selection range, and α represent the number of neighbors of node i. i,j W represents the importance of node j to node i. b It is a fully connected layer.

[0026] Furthermore, step S2 specifically includes the following steps:

[0027] S21: Driving lane information based on step S11 The data is converted into list data using a polynomial interpolation method; an interpolation point is sampled every 1m as a lane node, specifically represented as follows:

[0028]

[0029] Where m represents the number of lanes in the current scene, 1≤i≤m, and j represents the total number of nodes in a single lane;

[0030] S22: Based on the obtained lane node information, for the lane Add lane information matrix to expand lane node features:

[0031]

[0032] Among them, l u ,l v It is a set of lane node index values, Lane pre Matrix representation l v The indexed lane node is l u The index corresponds to the predecessor node of the lane node; similarly, the Lane node... suc Represents subsequent nodes, Lane left Lane right These represent the left neighbor and right neighbor nodes, respectively.

[0033] S23: For the established positional information relationships, feature enhancement is performed using graph convolution, and the target node features are updated using the following formula:

[0034]

[0035] in, For the target lane node features, For linear layers performing feature mapping, The node features after aggregating positional relationships are calculated as follows:

[0036]

[0037] in, It is a fully connected layer that performs a linear transformation on the target features;

[0038] S24: In the prediction task, simply aggregating features of adjacent lane nodes for the target lane node is insufficient for representing lane information. It is necessary to extend this relationship to more distant nodes. Therefore, the `pre` and `suc` type nodes are extended by k steps, and the node feature update becomes:

[0039]

[0040] Among them, l v_k The indexed lane node is l u_k The k-th node preceding / following the index is typically set to 6 within the range that ensures the validity of the lane node sequence. For the linear transformation layer corresponding to the extended node;

[0041] S25: Introduce an attention mechanism to model vehicle-lane interaction features, based on the previously obtained agent features A and lane node features. In this case, first traverse the agent nodes and lane nodes, filter out node pairs whose distance meets the range requirement, and obtain their corresponding index list: c i ,c jFurthermore, the distance information of each pair of nodes that meet the conditions is recorded; subsequently, the selected feature information is linearly transformed to obtain the agent query feature (query) and distance feature (dist). Based on the above information, the features of the interactive lane nodes connected to the agent nodes by distance information are further concatenated and superimposed:

[0042]

[0043] Ctx = W c (query‖dist‖ctx)

[0044]

[0045] Where ctx represents the selected lane node features, and Ctx is a fusion vector of query features, distance features, and lane node aggregation features. After being accumulated with the corresponding agent node features, it passes through the ReLU activation layer and the Norm regularization layer to obtain the vehicle lane interaction features la and W. l W c as well as All are fully connected layers that perform linear transformations on the corresponding features.

[0046] S26: For the interaction features between agent node i and the lane, after dimensionality processing, it is fused with the previous agent dynamic features and interaction features to decode the agent's future predicted trajectory in the time domain P:

[0047]

[0048] Enc = fuse(f + A + LA)

[0049]

[0050] Among them, LA i The vehicle lane interaction features after average pooling are combined with the previously obtained agent dynamics features f and interaction features A to obtain a comprehensive feature code Enc after feature fusion. Similarly, the agent's future trajectory is obtained through a recurrent neural network-based LSTM decoder.

[0051] Furthermore, step S3 specifically includes the following steps:

[0052] S31: Based on the single-vehicle model, define the system's state variables h and control variables u:

[0053]

[0054] In u, 'a' represents the vehicle's acceleration, 'δ' is the front wheel steering angle, 'v_s' is the velocity along the reference path in the arc coordinate system, 's_v' is the relaxation factor, and the variables in h are the vehicle's current position coordinates (x, y) and heading angle, respectively. and vehicle speed v;

[0055] S32: Taking into account the key factors affecting autonomous driving, design the objective function J(u) based on model predictive control. t ,h t ,s t ):

[0056] J(u t ,h t ,s t ) = J follow (h t ,s t )+J v (h t )+J u (u t )+J LF (h t ,s t )

[0057] Among them, s t J represents the vehicle's position coordinates in the arc coordinate system at time t; follow (h t ,s t J is the path following objective function. v (h t Let J be the velocity preservation objective function. u (u t J is the objective function for controlling the motion quantity. LF (h t ,s t Let be the objective function of the road potential field; the specific calculation formulas for each objective function are as follows:

[0058]

[0059] in, and These represent the errors between the actual vehicle position and the approximate vehicle position in the arc length direction and the vehicle lateral direction, respectively, during path following; w s and w l This is the weight matrix. Let s be the state quantity of the vehicle at time t. t It is the reference arc length coordinate of the vehicle's current position;

[0060]

[0061] in, J represents the product of the squared term of the variable and the weight value. v This represents the velocity term of the target being tracked, v t Let v be the vehicle's speed at time t. ref J represents the target velocity. u The objective function is composed of the solver's control variables, including the acceleration a. t Steering angle δ t Arc velocity v_s and relaxation factor s_v;

[0062]

[0063] The above equation is the objective function for road constraint-based design. We aim to regulate vehicle motion within the lane and keep it close to the lane centerline, while assigning lower cost values ​​to adjacent lanes compared to lane boundaries to ensure the vehicle considers lane-changing possibilities when avoiding obstacles; where y L L is the distance from the vehicle to the lane boundary line. width w is the lane width LF This is the weight coefficient matrix for this item;

[0064] S33: When optimizing the objective function using model predictive control, it is also necessary to constrain some of the variables to ensure that they are within a reasonable calculation range. The specific constraints are as follows:

[0065] h min ≤h t ≤h max

[0066] u min ≤u t ≤u max

[0067]

[0068] Among them, h min h max u min u max and Define the upper and lower limits for state variables and control variables respectively. This is the current heading angle of the vehicle. This is the heading angle of the vehicle at an approximate point on the reference path;

[0069] S34: In the planning step, the predicted trajectory information of the vehicle around the perimeter is packaged and processed, and then used as a dynamic obstacle. A contour error model is used to form a vehicle planning controller for constraint. First, the predicted trajectory of the vehicle around the perimeter is defined as:

[0070]

[0071] in, The output of the trajectory prediction model, j represents the future movement trajectory of surrounding vehicles, and j represents the number of neighboring intelligent agents within the filtering range.

[0072] Next, we enclose the shape of the vehicle using three circles centered on the vehicle's centerline, denoted as . The orbital radius is represented by an ellipse. For each dynamic obstacle, the information is represented by... Indicate its position coordinates, Let a represent the rotation matrix. j and b j The major and minor axes represent the vehicle's envelope ellipse. Based on this, the controller's obstacle avoidance constraints can be transformed into the intersection of the region occupied by the vehicle's envelope circle and the regions occupied by the ellipses of surrounding vehicles being zero.

[0073]

[0074] in, To represent obstacle avoidance constraints, Let α be the difference between the center of the self-circle and the center of the ellipse in the x and y directions, t∈[0,P], where P is the prediction time domain; α=a+r i and β=b+r i Let r be the union of the original surrounding vehicle ellipse and the vehicle's circle, which is used to approximate the Minkowski sum of the surrounding vehicle ellipse. i Let be the radius of the envelope circle of the vehicle.

[0075] S35: Based on the above objective function and constraints, within the prediction time domain P of the prediction network, the vehicle motion planning problem can be transformed into a rolling optimization quadratic optimal planning problem, which can be specifically expressed as:

[0076]

[0077] in, This represents the optimal control quantity solved by the controller in the prediction time domain. After obtaining the optimal control sequence, the vehicle state and the initial value of the solver are iteratively updated through the vehicle model, and then the next round of optimization is entered.

[0078] The beneficial effects of this invention are as follows: The method efficiently captures the multimodal interaction characteristics in complex traffic tasks, while integrating model predictive control theory. By incorporating predictive information from surrounding vehicles, it dynamically updates the system state, generating a reasonable and safe motion plan for the vehicle. Compared with traditional planning methods, this invention significantly improves the accuracy of interaction modeling and further optimizes traffic efficiency and behavioral robustness while ensuring safety.

[0079] (1) This invention encodes the historical features of vehicles using a recurrent neural network and captures the behavioral interactions between vehicles as vehicle nodes through a multi-head attention mechanism. A dynamic neighbor selection method is also designed to more efficiently construct the interaction graph between traffic participants.

[0080] (2) Based on vehicle driving lane information, this invention constructs a novel lane node information matrix and designs a corresponding graph convolution method to model and expand the driving characteristics of the lane. An attention mechanism is introduced to model the interaction features between vehicle nodes and weighted lane nodes, helping vehicles to better understand the current driving environment and make more accurate trajectory behavior predictions.

[0081] (3) This invention is based on the output information of the interactive prediction model and introduces it into the safe motion planning of the autonomous vehicle through the idea of ​​model predictive control. The autonomous vehicle supplements the prediction information as a system state update by predicting the future state, and designs the objective function and constraints to solve for the current optimal acceleration and steering angle of the vehicle. It can make safe and reasonable autonomous vehicle path planning while making efficient use of prediction information.

[0082] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0083] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0084] Figure 1 This is an overall flowchart of the autonomous driving prediction-planning integration method based on traffic heterogeneity graphs of the present invention.

[0085] Figure 2 A flowchart for building a multi-agent trajectory prediction model based on traffic heterogeneous graphs;

[0086] Figure 3 A flowchart of an autonomous driving planning method that integrates predictive information. Detailed Implementation

[0087] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0088] Please see Figures 1-3 This invention provides an integrated prediction-planning method for autonomous driving based on traffic heterogeneous graphs, such as... Figure 1 As shown, this method first models the dynamic features of the agents and the interaction features among traffic participants, and designs a lane graph node feature representation based on a graph convolutional network. Secondly, it captures the interaction between lane nodes and agents through an attention mechanism, performs feature fusion, and decodes to output the predicted trajectory of the multi-agent system. Finally, it introduces prediction information by establishing and updating the system state of model predictive control in real time, and designs an objective function and various constraints to optimize the vehicle's actions.

[0089] like Figure 2 As shown, the construction of a multi-agent trajectory prediction model based on traffic heterogeneous graphs includes the following steps:

[0090] S21: Driving lane information based on step S11 The data is converted into list data using a polynomial interpolation method; an interpolation point is sampled every 1m as a lane node, specifically represented as follows:

[0091]

[0092] Where m represents the number of lanes in the current scene, 1≤i≤m, and j represents the total number of nodes in a single lane;

[0093] S22: Based on the obtained lane node information, for the lane Add lane information matrix to expand lane node features:

[0094]

[0095] Among them, l u ,l v It is a set of lane node index values, Lane pre Matrix representation l v The indexed lane node is l u The index corresponds to the predecessor node of the lane node; similarly, the Lane node... suc Represents subsequent nodes, Laneleft Lane right These represent the left neighbor and right neighbor nodes, respectively.

[0096] S23: For the established positional information relationships, feature enhancement is performed using graph convolution, and the target node features are updated using the following formula:

[0097]

[0098] in, For the target lane node features, For linear layers performing feature mapping, The node features after aggregating positional relationships are calculated as follows:

[0099]

[0100] in, It is a fully connected layer that performs a linear transformation on the target features;

[0101] S24: In the prediction task, simply aggregating features of adjacent lane nodes for the target lane node is insufficient for representing lane information. It is necessary to extend this relationship to more distant nodes. Therefore, the `pre` and `suc` type nodes are extended by k steps, and the node feature update becomes:

[0102]

[0103] Among them, l v_k The indexed lane node is l u_k The index is the k-th node preceding / following it. To ensure the validity of the lane node sequence, k is typically set to 6. For the linear transformation layer corresponding to the extended node;

[0104] S25: Introduce an attention mechanism to model vehicle-lane interaction features, based on the previously obtained agent features A and lane node features. In this case, first traverse the agent nodes and lane nodes, filter out node pairs whose distance meets the range requirement, and obtain their corresponding index list: c i ,c j Furthermore, the distance information of each pair of nodes that meet the conditions is recorded; subsequently, the selected feature information is linearly transformed to obtain the agent query feature (query) and distance feature (dist). Based on the above information, the features of the interactive lane nodes connected to the agent nodes by distance information are further concatenated and superimposed:

[0105]

[0106] Ctx = W c(query‖dist‖ctx)

[0107]

[0108] Where ctx represents the selected lane node features, and Ctx is a fusion vector of query features, distance features, and lane node aggregation features. After being accumulated with the corresponding agent node features, it passes through the ReLU activation layer and the Norm regularization layer to obtain the vehicle lane interaction features la and W. l W c as well as All are fully connected layers that perform linear transformations on the corresponding features;

[0109] S26: For the interaction features between agent node i and the lane, after dimensionality processing, it is fused with the previous agent dynamic features and interaction features to decode the agent's future predicted trajectory in the time domain P:

[0110]

[0111] Enc = fuse(f + A + LA)

[0112]

[0113] Among them, LA i The vehicle lane interaction features after average pooling are combined with the previously obtained agent dynamics features f and interaction features A to obtain a comprehensive feature code Enc after feature fusion. Similarly, the agent's future trajectory is obtained through a recurrent neural network-based LSTM decoder.

[0114] like Figure 3 As shown, the autonomous driving planning method that integrates predictive information specifically includes the following steps:

[0115] S31: Based on the single-vehicle model, define the system's state variables h and control variables u:

[0116]

[0117] In u, 'a' represents the vehicle's acceleration, 'δ' is the front wheel steering angle, 'v_s' is the velocity along the reference path in the arc coordinate system, 's_v' is the relaxation factor, and the variables in h are the vehicle's current position coordinates (x, y) and heading angle, respectively. and vehicle speed v;

[0118] S32: Taking into account the key factors affecting autonomous driving, design the objective function J(u) based on model predictive control. t ,h t ,s t ):

[0119] J(u t ,h t ,s t ) = J follow (h t ,s t )+J v (h t )+J u (u t )+J LF (h t ,s t )

[0120] Among them, s t J represents the vehicle's position coordinates in the arc coordinate system at time t; follow (h t ,s t J is the path following objective function. v (h t Let J be the velocity preservation objective function. u (u t J is the objective function for controlling the motion quantity. LF (h t ,s t Let be the objective function of the road potential field; the specific calculation formulas for each objective function are as follows:

[0121]

[0122] in, and These represent the errors between the actual vehicle position and the approximate vehicle position in the arc length direction and the vehicle lateral direction, respectively, during path following; w s and w l This is the weight matrix. Let s be the state quantity of the vehicle at time t. t It is the reference arc length coordinate of the vehicle's current position;

[0123]

[0124] in, J represents the product of the squared term of the variable and the weight value. v This represents the velocity term of the target being tracked, v t Let v be the vehicle's speed at time t. ref J represents the target velocity. u The objective function is composed of the solver's control variables, including the acceleration a. t Steering angle δ t Arc velocity v_s and relaxation factor s_v;

[0125]

[0126] The above equation is the objective function for road constraint-based design. We aim to regulate vehicle motion within the lane and keep it close to the lane centerline, while assigning lower cost values ​​to adjacent lanes compared to lane boundaries to ensure the vehicle considers lane-changing possibilities when avoiding obstacles; where y L L is the distance from the vehicle to the lane boundary line. width w is the lane width LF This is the weight coefficient matrix for this item;

[0127] S33: When optimizing the objective function using model predictive control, it is also necessary to constrain some of the variables to ensure that they are within a reasonable calculation range. The specific constraints are as follows:

[0128] h min ≤h t ≤h max

[0129] u min ≤u t ≤u max

[0130]

[0131] Among them, h min h max u min u max and Define the upper and lower limits for state variables and control variables respectively. This is the current heading angle of the vehicle. This is the heading angle of the vehicle at an approximate point on the reference path;

[0132] S34: In the planning step, the predicted trajectory information of the vehicle around the perimeter is packaged and processed, and then used as a dynamic obstacle. A contour error model is used to form a vehicle planning controller for constraint. First, the predicted trajectory of the vehicle around the perimeter is defined as:

[0133]

[0134] in, The output of the trajectory prediction model, The future movement trajectory of surrounding vehicles is represented by j, which is the number of neighboring intelligent agents within the filtering range.

[0135] Next, we enclose the shape of the vehicle using three circles centered on the vehicle's centerline, denoted as . The orbital radius is represented by an ellipse. For each dynamic obstacle, the information is represented by... Indicate its position coordinates, Let a represent the rotation matrix. j and bj The major and minor axes represent the vehicle's envelope ellipse. Based on this, the controller's obstacle avoidance constraints can be transformed into the intersection of the region occupied by the vehicle's envelope circle and the regions occupied by the ellipses of surrounding vehicles being zero.

[0136]

[0137] in, To represent obstacle avoidance constraints, Let α be the difference between the center of the self-circle and the center of the ellipse in the x and y directions, t∈[0,P], where P is the prediction time domain; α=a+r i and β=b+r i Let r be the union of the original surrounding vehicle ellipse and the vehicle's circle, which is used to approximate the Minkowski sum of the surrounding vehicle ellipse. i Let be the radius of the envelope circle of the vehicle.

[0138] S35: Based on the above objective function and constraints, within the prediction time domain P of the prediction network, the vehicle motion planning problem can be transformed into a rolling optimization quadratic optimal planning problem, which can be specifically expressed as:

[0139]

[0140] in, This represents the optimal control quantity solved by the controller in the prediction time domain. After obtaining the optimal control sequence, the vehicle state and the initial value of the solver are iteratively updated through the vehicle model, and then the next round of optimization is entered.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An integrated prediction-planning method for autonomous driving based on traffic heterogeneous graphs, characterized in that, The method specifically includes the following steps: S1: Model the dynamic characteristics of the agents and the interaction characteristics among traffic participants; for each agent, design dynamic filtering conditions to select neighboring nodes: in, V is the radius of the circle to be selected. i t L represents the current vehicle speed at the target node. vehicle Let λ be the length of the vehicle. R A constant value chosen based on experience; S2: Design a lane graph node feature representation based on a graph convolutional network; adopt an encoder-decoder architecture, capture the interaction features between lane nodes and agents through a graph attention mechanism, and construct a multi-agent trajectory prediction model based on traffic heterogeneous graphs after feature fusion; Step S2 specifically includes the following steps: S21: Based on driving lane information The data is converted into list data using a polynomial interpolation method; an interpolation point is sampled every 1m as a lane node, specifically represented as follows: Where m represents the number of lanes in the current scenario (1≤i≤m), and j represents the total number of nodes in a single lane. S22: Based on the obtained lane node information, for the lane Add lane information matrix to expand lane node features: Among them, l u ,l v It is a set of lane node index values, Lane pre Matrix representation l v The indexed lane node is l u The index corresponds to the predecessor node of the lane node; similarly, the Lane node... suc Represents subsequent nodes, Lane left Lane right These represent the left neighbor and right neighbor nodes, respectively. S23: For the established positional information relationships, feature enhancement is performed using graph convolution, and the target node features are updated using the following formula: in, For the target lane node features, For linear layers performing feature mapping, The node features after aggregating positional relationships are calculated as follows: in, It is a fully connected layer that performs a linear transformation on the target features; S24: Extend the pre and suc type nodes by k steps, and the node feature update becomes: Among them, l v_k The indexed lane node is l u_k The k-th node that is the predecessor / follower of the index; For the linear transformation layer corresponding to the extended node features; S25: Introduce an attention mechanism to model vehicle-lane interaction features, obtaining agent features A and lane node features. In this case, first traverse the agent nodes and lane nodes, filter out node pairs whose distance meets the range requirement, and obtain their corresponding index list: c i ,c j Furthermore, the distance information of each pair of nodes that meet the conditions is recorded; subsequently, the selected feature information is linearly transformed to obtain the agent query feature (query) and distance feature (dist). Based on the above information, the features of the interactive lane nodes connected to the agent nodes by distance information are further concatenated and superimposed: Ctx=W c (query‖dist‖ctx) Where ctx represents the selected lane node features, and Ctx is a fusion vector of query features, distance features, and lane node aggregation features. After being accumulated with the corresponding agent node features, it passes through the ReLU activation layer and the Norm regularization layer to obtain the vehicle lane interaction features la and W. l W c as well as All are fully connected layers that perform linear transformations on the corresponding features; S26: For the interaction features between agent node i and the lane, after dimensionality processing, it is fused with the previous agent dynamic features and interaction features to decode the agent's future predicted trajectory in the time domain P: Enc = fuse(f + A + LA) Among them, LA i The vehicle lane interaction features obtained after average pooling are combined with the obtained agent dynamics features f and interaction features A, and then fused to obtain the comprehensive feature code Enc. Similarly, the future trajectory of the agent is obtained through a decoder LSTM based on a recurrent neural network. S3: Based on the future location information of surrounding vehicles output by the trajectory prediction model, design an objective function and various constraints to optimize the vehicle's actions.

2. The autonomous driving prediction-planning integrated method according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11: Define current traffic information, including lane information. And the historical trajectory information of n agents at time t Where t0 is the length of the historical trajectory; Let be the motion state information of the i-th agent at time t, where These represent the coordinates, velocity components in the x and y directions, and heading angle of the i-th agent at the current moment, respectively. S12: For agent node features, a recurrent neural network-based method is used to dynamically encode the three traffic participant types: vehicles, bicycles, and pedestrians; for agent A... i The dynamic characteristics are encoded as follows: Among them, f i t This represents the dynamic characteristics after encoding the vehicle's historical trajectory information, where GRU(·) indicates encoding via a recurrent neural network; S13: Model the interaction features of intelligent agents using graph neural networks. First, for each intelligent agent, design dynamic filtering conditions to select neighboring nodes. S14: After constructing the graph structure data representation, a graph attention mechanism is introduced to further enhance the interaction features between vehicle nodes; firstly, for the target node i and its selected neighbor nodes j, the desired feature is to obtain the concatenation feature z of its directed edges. i,j : z i,j =ReLU(W a [h i ‖have been i,j ‖h j ]) Among them, h i ,h j For the encoding features of the node, e i,j W represents the edge embedding feature from node i to node j. a Let be the attention linear transformation matrix, and || be the concatenation of features; S15: After concatenating the features of each edge of the target node, normalize them using softmax to obtain their attention scores, and then update the features of the target node: Among them, A i This represents the updated target node characteristics. Let α represent the number of neighbors of node i, k represent the number of neighboring agent nodes within the target node i's selection range, and α represent the number of neighbors of node i. i,j W represents the importance of node j to node i. b It is a fully connected layer.

3. The autonomous driving prediction-planning integrated method according to claim 2, characterized in that, Step S3 specifically includes the following steps: S31: Based on the single-vehicle model, define the system's state variables h and control variables u: In u, 'a' represents the vehicle's acceleration, 'δ' is the front wheel steering angle, 'v_s' is the velocity along the reference path in the arc coordinate system, 's_v' is the relaxation factor, and the variables in h are the vehicle's current position coordinates (x, y) and heading angle, respectively. and vehicle speed v; S32: Taking into account the key factors affecting autonomous driving, design the objective function J(u) based on model predictive control. t ,h t ,s t ): J(u t ,h t ,s t )=J follow (h t ,s t )+J v (h t )+J u (u t )+J LF (h t ,s t ) Among them, s t J represents the vehicle's position coordinates in the arc coordinate system at time t; follow (h t ,s t J is the path following objective function. v (h t Let J be the velocity preservation objective function. u (u t J is the objective function for controlling the motion quantity. LF (h t ,s t Let be the objective function of the road potential field; the specific calculation formulas for each objective function are as follows: in, and These represent the errors between the actual vehicle position and the approximate vehicle position in the arc length direction and the vehicle lateral direction, respectively, during path following; w s and w l This is the weight matrix. Let s be the state quantity of the vehicle at time t. t It is the reference arc length coordinate of the vehicle's current position; in, J represents the product of the squared term of the variable and the weight value. v This represents the velocity term of the target being tracked, v t Let v be the vehicle's speed at time t. ref J represents the target velocity. u The objective function is composed of the solver's control variables, including the acceleration a. t Steering angle δ t Arc velocity v_s and relaxation factor s_v; The above equation is the objective function for road constraint-based design, aiming to regulate vehicle motion within the lane and close to the lane centerline, while assigning lower cost values ​​to adjacent lanes compared to lane boundaries to ensure the vehicle considers lane-changing possibilities when avoiding obstacles; where y L L is the distance from the vehicle to the lane boundary line. width w is the lane width LF This is the weight coefficient matrix for this item; S33: When optimizing the objective function using model predictive control, it is also necessary to constrain some of the variables to ensure that they are within a reasonable calculation range. The specific constraints are as follows: Among them, h min h max u min u max and Define the upper and lower limits for state variables and control variables respectively. This is the current heading angle of the vehicle. This is the heading angle of the vehicle at an approximate point on the reference path; S34: In the planning step, the predicted trajectory information of the vehicle around the perimeter is packaged and processed, and then used as a dynamic obstacle. A contour error model is used to form a vehicle planning controller for constraint. First, the predicted trajectory of the vehicle around the perimeter is defined as: in, The output of the trajectory prediction model, The future movement trajectory of surrounding vehicles is represented by j, which is the number of neighboring intelligent agents within the filtering range. Next, we enclose the shape of the vehicle using three circles centered on the vehicle's centerline, denoted as . The orbital radius is represented by an ellipse. For each dynamic obstacle, the information is represented by... Indicate its position coordinates, Let a represent the rotation matrix. j and b j The major and minor axes represent the vehicle's envelope ellipse. Based on this, the controller's obstacle avoidance constraint is transformed into the intersection of the region occupied by the vehicle's envelope circle and the regions occupied by the ellipses of surrounding vehicles being zero. in, Representation of obstacle avoidance constraints; Let α be the difference between the center of the self-circle and the center of the ellipse in the x and y directions, t∈[0,P], where P is the prediction time domain; α=a+r i and β=b+r i Let r be the union of the original surrounding vehicle ellipse and the vehicle's circle, which is used to approximate the Minkowski sum of the surrounding vehicle ellipse. i Let be the radius of the envelope circle of the vehicle. S35: Based on the above objective function and constraints, within the prediction time domain P of the prediction network, the vehicle motion planning problem is transformed into a rolling optimization quadratic optimal planning problem, specifically expressed as: in, This represents the optimal control quantity solved by the controller in the prediction time domain. After obtaining the optimal control sequence, the vehicle state and the initial value of the solver are iteratively updated through the vehicle model, and then the next round of optimization is entered.

Images