Automatic driving prediction-planning integration method based on traffic heterogeneous graph

Abstract

The invention relates to an automatic driving prediction-planning integration method based on a traffic heterogeneous graph, and belongs to the technical field of automatic driving automobiles. The method comprises the following steps: S1, modeling intelligent agent dynamic features and interaction features between traffic participants, and designing lane map node feature representation based on a graph convolutional network; s2, an encoding-decoding architecture is adopted, interaction features between lane nodes and agents are captured through a graph attention mechanism, and after feature fusion is carried out, a multi-agent trajectory prediction model based on a traffic heterogeneous graph is constructed; and S3, on the basis of future position information of surrounding vehicles output by the trajectory prediction model, designing a target function and various constraints to carry out self-vehicle action optimization solution. Compared with a traditional planning method, the method is more excellent in prediction performance, safety and driving efficiency.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving vehicles and relates to an autonomous driving prediction-planning integrated method based on a traffic heterogeneous graph. Background Art

[0002] In recent years, autonomous driving has become one of the research hotspots in the field of artificial intelligence and transportation. The autonomous driving system aims to achieve autonomous perception, decision-making and control of vehicles in complex road environments. Trajectory prediction infers the future motion trend of the target by analyzing road scenes, traffic rules and the dynamic behavior of traffic participants, providing reliable prior information for downstream. After receiving the prediction information, motion planning combines vehicle dynamics constraints, driving environment and driving goals to generate future trajectories that meet safety, comfort and efficiency. The current mainstream motion planning methods can be divided into rule-based, optimization-based and learning-based methods. Although the traditional rule-based method has clear logic and strong interpretability, it relies on expert experience, requires manual task decomposition and definition of trigger conditions, and is difficult to handle multi-agent interactions and dynamic changes in the environment. The optimization-based method defines the driving task as a mathematical optimization problem, and optimizes and solves it by setting objective functions and constraints. The learning-based method uses technologies such as deep learning and reinforcement learning to implicitly learn target strategies from data, and can flexibly adapt to complex traffic scenarios.

[0003] Whether it is a modular or end-to-end autonomous driving framework, the two core modules of trajectory prediction and motion planning need to be tightly coupled in terms of functional implementation to improve the efficiency of feature transmission and reduce the cumulative error in the information transmission process. However, the upstream information received by the current mainstream motion planning methods is insufficient for modeling traffic scenarios, and fails to fully utilize the rich information provided by the trajectory prediction module. Therefore, there is an urgent need for a prediction planning integration method that efficiently integrates interactive prediction information and deeply explores the coupling mechanism between prediction information and motion planning, so as to further improve the driving safety and traffic efficiency of autonomous driving vehicles. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide an integrated method for prediction-planning of autonomous driving based on a traffic heterogeneous graph, which effectively models multi-modal interactions in complex interactive scenarios, efficiently integrates surrounding vehicle prediction information, and ensures that the vehicle makes safe and reasonable motion planning through optimization. Compared with traditional planning methods, the present invention has better prediction performance, safety and driving efficiency.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] An integrated method for prediction and planning of autonomous driving based on a heterogeneous traffic graph includes the following steps:

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

[0008] S2: Adopting the encoder-decoder architecture, the graph attention mechanism is used to capture the interaction features between lane nodes and agents, and after feature fusion, a multi-agent trajectory prediction model based on the traffic heterogeneous graph is constructed;

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

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

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

[0012]

[0013] Among them, t0 is the length of the historical trajectory; is the motion state information of the ith agent at time t, Respectively represent the vehicle's current coordinate information, velocity components in the x and y directions, and heading angle information;

[0014] S12: For the agent node features, a recurrent neural network-based method is used to dynamically encode the three types of traffic participants: vehicles, bicycles, and pedestrians. i , encoding its dynamic characteristics as:

[0015]

[0016] Among them, f i t represents the dynamic characteristics after encoding the historical trajectory information of the vehicle, and GRU(·) represents encoding through a recurrent neural network;

[0017] S13: Modeling the interaction characteristics of intelligent agents through graph neural networks. First, for each intelligent agent, dynamic screening conditions are designed to select neighbor nodes:

[0018]

[0019] in, is the radius of the screening circle, V i t is the current speed of the target node, L vehicle is the length of the vehicle, λR is a constant value selected empirically;

[0020] S14: After constructing the graph structure data representation, the graph attention mechanism is introduced to further enhance the interaction features between vehicle nodes; first, for the target node i and its selected neighbor node j, it is expected to obtain the splicing 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 is the encoding feature of the node, e i,j represents the edge embedding feature from node i to node j, W a is the attention linear transformation matrix, ‖ is the cascade of features;

[0023] S15: After concatenating the features of each edge of the target node, normalize it through softmax to get its attention score, and update the target node features:

[0024]

[0025] Among them, A i represents the updated target node features, represents the number of neighbors of node i, k represents the number of neighboring agent nodes within the screening range of target node i, α i,j represents the importance of node j to node i, W b is a fully connected layer.

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

[0027] S21: Based on the driving lane information defined in step S11 It is converted into list data through polynomial interpolation method; an interpolation point is sampled every 1m as a lane node, which is specifically expressed as:

[0028]

[0029] Among them, 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, Add lane information matrix to expand lane node features:

[0031]

[0032] Among them, l u ,l v Is a collection of lane node index values, Lane pre Matrix representation l v The lane node indexed is l u The index of the corresponding lane node's predecessor node, similarly, Lane suc Represents the subsequent node, Lane left ,Lane right Represent the left and right neighbor nodes respectively;

[0033] S23: For the established location information relationship, the graph convolution idea is used to enhance the features, and the following formula is used to update the target node features:

[0034]

[0035] in, is the target lane node feature, is the linear layer for feature mapping, is the node feature after the aggregation position relationship, and the specific calculation method is:

[0036]

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

[0038] S24: In the prediction task, for the target lane node, only the features of the adjacent lane nodes are gathered, which is insufficient in representing the lane information. This relationship needs to be further extended to nodes farther away. 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 lane node indexed is l u_k The index of the previous / successor kth node. The k value is usually 6 to ensure the lane node sequence is valid. is the linear transformation layer corresponding to the extended node;

[0041] S25: Introduce the attention mechanism to model the vehicle-lane interaction features. In the above, we get the agent feature A and lane node feature. In the case of , first traverse the agent nodes and lane nodes, filter out the node pairs whose distances meet the range requirements, and obtain their corresponding index lists: c i ,c j, and record the distance information of each pair of nodes that meet the conditions; subsequently, perform linear transformation on the filtered feature information to obtain the agent query feature query and the distance feature dist. Based on the above information, further obtain the feature cascade that connects the agent node with the interactive lane node by distance information and perform feature superposition:

[0042]

[0043] Ctx=W c (query‖dist‖ctx)

[0044]

[0045] Among them, ctx is the filtered lane node feature, Ctx is the fusion vector of query feature, distance feature and lane node aggregation feature, which is accumulated with the corresponding agent node feature, and then passes through the activation layer Relu and the regularization layer Norm to obtain the vehicle lane interaction feature la, W l , W c as well as These are all fully connected layers that perform linear transformations on the corresponding features.

[0046] S26: For the interaction features between the agent node i and the lane, after dimension processing, they are fused with the previous agent dynamic features and interaction features to decode the future prediction trajectory under the agent prediction domain P:

[0047]

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

[0049]

[0050] Among them, LA i The vehicle lane interaction feature after the average pooling operation is combined with the previously obtained intelligent agent dynamic feature f and interaction feature A to obtain the comprehensive feature encoding Enc after feature fusion. The future trajectory of the intelligent agent is also obtained through the decoder LSTM based on the recurrent neural network

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

[0052] S31: Combined with the vehicle's single-vehicle model, define the system's state quantity h and control quantity u:

[0053]

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

[0055] S32: Comprehensively consider the key factors affecting autonomous driving and design the objective function J(u 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 is the position coordinate of the vehicle in the arc coordinate system at time t; follow (h t ,s t ) is the path following objective function, J v (h t ) is the velocity maintenance objective function, J u (u t ) is the objective function of the control action, J LF (h t ,s t ) is the road potential field objective function; the specific calculation formulas of each objective function are as follows:

[0058]

[0059] in, and They represent the error values ​​between the real vehicle position and the approximate vehicle position in the arc length direction and the vehicle side direction during path following; w s and w l is the weight matrix, is the state of the vehicle at time t, s t is the reference arc length coordinate of the vehicle’s current position;

[0060]

[0061] in, Indicates that the square term of the variable is multiplied by the weight value, J v represents the tracking target velocity term, v t is the speed of the vehicle at time t, v ref represents the target speed, J u is the objective function composed of the control variables of the solver, including the acceleration a t , steering angle δ t , arc velocity v_s and relaxation factor s_v;

[0062]

[0063] The above formula is the objective function designed based on road constraints. We hope to regulate the movement of the vehicle in the lane and drive close to the center line of the lane. At the same time, we give the adjacent lane a lower cost value than the lane boundary to ensure that the vehicle considers the possibility of lane change when avoiding obstacles. In the formula, y L is the distance from the vehicle to the lane boundary, L width is the lane width, w LF is the weight coefficient matrix of this item;

[0064] S33: When optimizing and solving the objective function through model predictive control, some variables need to be constrained 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 The upper and lower limits of the state quantity and the control quantity respectively, is the current vehicle heading angle, is the heading angle of the vehicle at the approximate point on the reference path;

[0069] S34: In the planning step, the prediction information of the surrounding vehicle trajectory is packaged and processed, and then used as a dynamic obstacle and constrained by forming an ego vehicle planning controller through a contour error model. First, the prediction trajectory of the surrounding vehicle is defined as:

[0070]

[0071] in, is the output of the trajectory prediction model, is the future motion trajectory of the surrounding vehicles, j is the number of neighboring agents within the screening range;

[0072] Next, use three circles whose centers are on the centerline of the vehicle to envelop the shape of the vehicle, denoted as Use an ellipse to represent the perimeter vehicle. For each dynamic obstacle information, use Represents its position coordinates, represents the rotation matrix, a j and b j Represents the major and minor axes of the vehicle envelope ellipse. Based on this, the obstacle avoidance constraint of the controller can be converted into the intersection of the area occupied by the envelope circle of the vehicle and the area occupied by the ellipse representing the surrounding vehicles is zero:

[0073]

[0074] in, To express the obstacle avoidance constraint, is the difference between the center of the self-circle and the center of the surrounding ellipse in the x and y directions, t∈[0,P], P is the prediction time domain; α=a+r i And β = b + r i is the union of the original surrounding vehicle ellipse and the ego vehicle circle, which is used to approximate the Minkowski sum of the surrounding vehicle ellipse, r i is the radius of the vehicle envelope circle;

[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 converted into a quadratic optimal planning problem of rolling optimization type, which can be specifically expressed as:

[0076]

[0077] in, It 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 solver initial value are iteratively updated through the vehicle model to enter the next round of optimization solution.

[0078] The beneficial effect of the present invention is that the method of the present invention efficiently captures the multimodal interaction characteristics in complex traffic tasks, and integrates model predictive control theory, dynamically updates the system state by introducing the prediction information of surrounding vehicles, and generates a reasonable and safe motion plan for the vehicle. Compared with traditional planning methods, the present invention significantly improves the accuracy of interactive modeling, and further optimizes traffic efficiency and behavioral robustness while ensuring safety.

[0079] (1) Based on the recurrent neural network, the present invention encodes the dynamic characteristics of vehicle history features and uses a multi-head attention mechanism as a vehicle node to capture the behavioral interactions between vehicles. At the same time, a dynamic neighbor screening method is designed to more efficiently construct the interaction graph between traffic participants.

[0080] (2) Based on the vehicle driving lane information, the present 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. The introduction of the attention mechanism to model the interaction characteristics of the vehicle node and the weighted lane node helps the vehicle better understand the current driving environment and make more accurate trajectory behavior predictions.

[0081] (3) Based on the output information of the interactive prediction model, the present invention introduces it into the safe motion planning of the ego vehicle through the idea of ​​model predictive control. The ego vehicle predicts the future state, supplements the prediction information as the system state update, and designs the objective function and constraint conditions to solve the current optimal acceleration and steering angle of the vehicle. While efficiently utilizing the prediction information, safe and reasonable ego vehicle path planning can be made.

[0082] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0084] Figure 1 It is an overall flow chart of the automatic driving prediction-planning integrated method based on the traffic heterogeneous graph of the present invention;

[0085] Figure 2 Flowchart for building a multi-agent trajectory prediction model based on a traffic heterogeneous graph;

[0086] Figure 3 Flowchart of the autonomous driving planning method that integrates prediction information. DETAILED DESCRIPTION

[0087] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways 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 only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0088] See also Figure 1 to Figure 3 , the present invention provides an integrated method for prediction-planning of autonomous driving based on a traffic heterogeneous graph, such as Figure 1 As shown in the figure, this method first models the dynamic characteristics of the intelligent agent and the interaction characteristics between traffic participants, and designs a lane graph node feature representation based on a graph convolutional network. Secondly, the interaction between lane nodes and intelligent agents is captured through the attention mechanism, and the predicted trajectory of multiple intelligent agents is decoded and output after feature fusion. Finally, the prediction information is introduced by establishing and updating the system state of the model predictive control in real time, and the objective function and multiple constraints are designed to optimize the self-vehicle action.

[0089] like Figure 2 As shown in the figure, a multi-agent trajectory prediction model based on traffic heterogeneous graph is constructed, which specifically includes the following steps:

[0090] S21: Based on the driving lane information defined in step S11 It is converted into list data through polynomial interpolation method; an interpolation point is sampled every 1m as a lane node, which is specifically expressed as:

[0091]

[0092] Among them, 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, Add lane information matrix to expand lane node features:

[0094]

[0095] Among them, l u ,l v Is a collection of lane node index values, Lane pre Matrix representation l v The lane node indexed is l u The index of the corresponding lane node's predecessor node, similarly, Lane suc Represents the subsequent node, Laneleft ,Lane right Represent the left and right neighbor nodes respectively;

[0096] S23: For the established location information relationship, the graph convolution idea is used to enhance the features, and the following formula is used to update the target node features:

[0097]

[0098] in, is the target lane node feature, is the linear layer for feature mapping, is the node feature after the aggregation position relationship, and the specific calculation method is:

[0099]

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

[0101] S24: In the prediction task, for the target lane node, only the features of the adjacent lane nodes are gathered, which is insufficient in representing the lane information. This relationship needs to be further extended to nodes farther away. 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 lane node indexed is l u_k The index of the previous / subsequent k-th node. To ensure the lane node sequence is valid, the k value is usually 6. is the linear transformation layer corresponding to the extended node;

[0104] S25: Introduce the attention mechanism to model the vehicle-lane interaction features. In the above, we get the agent feature A and lane node feature. In the case of , first traverse the agent nodes and lane nodes, filter out the node pairs whose distances meet the range requirements, and obtain their corresponding index lists: c i ,c j , and record the distance information of each pair of nodes that meet the conditions; subsequently, perform linear transformation on the filtered feature information to obtain the agent query feature query and the distance feature dist. Based on the above information, further obtain the feature cascade that connects the agent node with the interactive lane node by distance information and perform feature superposition:

[0105]

[0106] Ctx=W c(query‖dist‖ctx)

[0107]

[0108] Among them, ctx is the filtered lane node feature, Ctx is the fusion vector of query feature, distance feature and lane node aggregation feature, which is accumulated with the corresponding agent node feature, and then passes through the activation layer Relu and the regularization layer Norm to obtain the vehicle lane interaction feature la, W l , W c as well as They are all fully connected layers that perform linear transformation on the corresponding features;

[0109] S26: For the interaction features between the agent node i and the lane, after dimension processing, they are fused with the previous agent dynamic features and interaction features to decode the future prediction trajectory under the agent prediction domain P:

[0110]

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

[0112]

[0113] Among them, LA i The vehicle lane interaction feature after the average pooling operation is combined with the previously obtained intelligent agent dynamic feature f and interaction feature A to obtain the comprehensive feature encoding Enc after feature fusion. The future trajectory of the intelligent agent is also obtained through the decoder LSTM based on the recurrent neural network

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

[0115] S31: Combined with the vehicle's single-vehicle model, define the system's state quantity h and control quantity u:

[0116]

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

[0118] S32: Comprehensively consider the key factors affecting autonomous driving and design the objective function J(u 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 is the position coordinate of the vehicle in the arc coordinate system at time t; follow (h t ,s t ) is the path following objective function, J v (h t ) is the velocity maintenance objective function, J u (u t ) is the objective function of the control action, J LF (h t ,s t ) is the road potential field objective function; the specific calculation formulas of each objective function are as follows:

[0121]

[0122] in, and They represent the error values ​​between the real vehicle position and the approximate vehicle position in the arc length direction and the vehicle side direction during path following; w s and w l is the weight matrix, is the state of the vehicle at time t, s t is the reference arc length coordinate of the vehicle’s current position;

[0123]

[0124] in, Indicates that the square term of the variable is multiplied by the weight value, J v represents the tracking target velocity term, v t is the speed of the vehicle at time t, v ref represents the target speed, J u is the objective function composed of the control variables of the solver, including the acceleration a t , steering angle δ t , arc velocity v_s and relaxation factor s_v;

[0125]

[0126] The above formula is the objective function designed based on road constraints. We hope to regulate the movement of the vehicle in the lane and drive close to the center line of the lane. At the same time, we give the adjacent lane a lower cost value than the lane boundary to ensure that the vehicle considers the possibility of lane change when avoiding obstacles. In the formula, y L is the distance from the vehicle to the lane boundary, L width is the lane width, w LF is the weight coefficient matrix of this item;

[0127] S33: When optimizing and solving the objective function through model predictive control, some variables need to be constrained 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 The upper and lower limits of the state quantity and the control quantity respectively, is the current vehicle heading angle, is the heading angle of the vehicle at the approximate point on the reference path;

[0132] S34: In the planning step, the prediction information of the surrounding vehicle trajectory is packaged and processed, and then used as a dynamic obstacle and constrained by forming an ego vehicle planning controller through a contour error model. First, the prediction trajectory of the surrounding vehicle is defined as:

[0133]

[0134] in, is the output of the trajectory prediction model, is the future motion trajectory of the surrounding vehicles, j is the number of neighboring agents within the screening range;

[0135] Next, use three circles whose centers are on the centerline of the vehicle to envelop the shape of the vehicle, denoted as Use an ellipse to represent the perimeter vehicle. For each dynamic obstacle information, use Represents its position coordinates, represents the rotation matrix, a j and bj Represents the major and minor axes of the vehicle envelope ellipse. Based on this, the obstacle avoidance constraint of the controller can be converted into the intersection of the area occupied by the envelope circle of the vehicle and the area occupied by the ellipse representing the surrounding vehicles is zero:

[0136]

[0137] in, To express the obstacle avoidance constraint, is the difference between the center of the self-circle and the center of the surrounding ellipse in the x and y directions, t∈[0,P], P is the prediction time domain; α=a+r i And β = b + r i is the union of the original surrounding vehicle ellipse and the ego vehicle circle, which is used to approximate the Minkowski sum of the surrounding vehicle ellipse, r i is the radius of the vehicle envelope circle;

[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 converted into a quadratic optimal planning problem of rolling optimization type, which can be specifically expressed as:

[0139]

[0140] in, It 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 solver initial value are iteratively updated through the vehicle model to enter the next round of optimization solution.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. An integrated method for prediction and planning of autonomous driving based on a heterogeneous traffic graph, characterized in that: The method specifically comprises the following steps: S1: Model the dynamic characteristics of intelligent agents and the interaction characteristics between traffic participants, and design a lane graph node feature representation based on graph convolutional networks; S2: Adopting the encoder-decoder architecture, the graph attention mechanism is used to capture the interaction features between lane nodes and agents, and after feature fusion, a multi-agent trajectory prediction model based on the traffic heterogeneous graph is constructed; S3: Based on the future position information of surrounding vehicles output by the trajectory prediction model, the objective function and various constraints are designed 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 driving lane information And the historical trajectory information of n agents at time t Among them, t0 is the length of the historical trajectory; is the motion state information of the ith agent at time t, where Respectively represent the vehicle's current coordinate information, velocity components in the x and y directions, and heading angle information; S12: For the agent node features, a recurrent neural network-based method is used to dynamically encode the three types of traffic participants: vehicles, bicycles, and pedestrians. i , encoding its dynamic characteristics as: Among them, f i t represents the dynamic characteristics after encoding the historical trajectory information of the vehicle, and GRU(·) represents encoding through a recurrent neural network; S13: Modeling the interaction characteristics of intelligent agents through graph neural networks. First, for each intelligent agent, dynamic screening conditions are designed to select neighbor nodes: in, is the radius of the screening circle, V i t is the current speed of the target node, L vehicle is the length of the vehicle, λ R is a constant value selected empirically; S14: After constructing the graph structure data representation, the graph attention mechanism is introduced to further enhance the interaction features between vehicle nodes; first, for the target node i and its selected neighbor node j, it is expected to obtain the splicing 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 is the encoding feature of the node, e i,j represents the edge embedding feature from node i to node j, W a is the attention linear transformation matrix, ‖ is the cascade of features; S15: After concatenating the features of each edge of the target node, normalize it through softmax to get its attention score, and update the target node features: Among them, A i represents the updated target node features, represents the number of neighbors of node i, k represents the number of neighboring agent nodes within the screening range of target node i, α i,j represents the importance of node j to node i, W b is a fully connected layer.

3. The autonomous driving prediction-planning integrated method according to claim 2, characterized in that: Step S2 specifically includes the following steps: S21: Based on the driving lane information defined in step S11 It is converted into list data through polynomial interpolation method; an interpolation point is sampled every 1m as a lane node, which is specifically expressed as: Among them, 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; S22: Based on the obtained lane node information, Add lane information matrix to expand lane node features: Among them, l u ,l v Is a collection of lane node index values, Lane pre Matrix representation l v The lane node indexed is l u The index of the corresponding lane node's predecessor node, similarly, Lane suc Represents the subsequent node, Lane left ,Lane right Represent the left and right neighbor nodes respectively; S23: For the established location information relationship, the graph convolution idea is used to enhance the features, and the following formula is used to update the target node features: in, is the target lane node feature, is the linear layer for feature mapping, is the node feature after the aggregation position relationship, and the specific calculation method is: in, It is a fully connected layer that performs linear transformation on the target features; S24: Expand the pre and suc type nodes by k steps, and the node feature update becomes: Among them, l v_k The lane node indexed is l u_k The predecessor / successor kth node of the index; It is the linear transformation layer corresponding to the extended node features; S25: Introduce the attention mechanism to model the vehicle-lane interaction characteristics, and obtain the agent feature A and lane node features. In the case of , first traverse the agent nodes and lane nodes, filter out the node pairs whose distances meet the range requirements, and obtain their corresponding index lists: c i ,c j , and record the distance information of each pair of nodes that meet the conditions; subsequently, perform linear transformation on the filtered feature information to obtain the agent query feature query and the distance feature dist. Based on the above information, further obtain the feature cascade that connects the agent node with the interactive lane node by distance information and perform feature superposition: Ctx=W c (query‖dist‖ctx) Among them, ctx is the filtered lane node feature, Ctx is the fusion vector of query feature, distance feature and lane node aggregation feature, which is accumulated with the corresponding agent node feature, and then passes through the activation layer Relu and the regularization layer Norm to obtain the vehicle lane interaction feature la, W l , W c as well as They are all fully connected layers that perform linear transformation on the corresponding features; S26: For the interaction features between the agent node i and the lane, after dimension processing, they are fused with the previous agent dynamic features and interaction features to decode the future prediction trajectory under the agent prediction domain P: Enc=fuse(f+A+LA) Among them, LA i The vehicle lane interaction features after the average pooling operation are combined with the obtained agent dynamic features f and interaction features A to obtain the comprehensive feature encoding Enc after feature fusion; the future trajectory of the agent is also obtained through the decoder LSTM based on the recurrent neural network 4. The autonomous driving prediction-planning integrated method according to claim 3, characterized in that: Step S3 specifically includes the following steps: S31: Combined with the vehicle's single-vehicle model, define the system's state quantity h and control quantity u: Among them, a in u represents the acceleration of the vehicle, δ is the front wheel steering angle, v_s is the speed along the reference path in the arc coordinate system, s_v is the relaxation factor, and the variables in h are the current position coordinates x, y, and heading angle of the vehicle. and vehicle speed v; S32: Comprehensively consider the key factors affecting autonomous driving and design the objective function J(u 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 is the position coordinate of the vehicle in the arc coordinate system at time t; follow (h t ,s t ) is the path following objective function, J v (h t ) is the velocity maintenance objective function, J u (u t ) is the objective function of the control action, J LF (h t ,s t ) is the road potential field objective function; the specific calculation formulas of each objective function are as follows: in, and They represent the error values ​​between the real vehicle position and the approximate vehicle position in the arc length direction and the vehicle side direction during path following; w s and w l is the weight matrix, is the state of the vehicle at time t, s t is the reference arc length coordinate of the vehicle’s current position; in, Indicates that the square term of the variable is multiplied by the weight value, J v represents the tracking target velocity term, v t is the speed of the vehicle at time t, v ref represents the target speed, J u is the objective function composed of the control variables of the solver, including the acceleration a t , steering angle δ t , arc velocity v_s and relaxation factor s_v; The above formula is the objective function designed based on road constraints. It is expected to regulate the movement of the vehicle in the lane and drive close to the center line of the lane. At the same time, the adjacent lane is given a lower cost value than the lane boundary to ensure that the vehicle considers the possibility of lane change when avoiding obstacles. In the formula, y L is the distance from the vehicle to the lane boundary, L width is the lane width, w LF is the weight coefficient matrix of this item; S33: When optimizing and solving the objective function through model predictive control, some variables need to be constrained to ensure that they are within a reasonable calculation range. The specific constraints are as follows: h min ≤h t ≤h max in min in t in max Among them, h min 、h max 、u min 、u max and The upper and lower limits of the state quantity and the control quantity respectively, is the current vehicle heading angle, is the heading angle of the vehicle at the approximate point on the reference path; S34: In the planning step, the prediction information of the surrounding vehicle trajectory is packaged and processed, and then used as a dynamic obstacle and constrained by forming an ego vehicle planning controller through a contour error model. First, the prediction trajectory of the surrounding vehicle is defined as: in, is the output of the trajectory prediction model, is the future motion trajectory of the surrounding vehicles, j is the number of neighboring agents within the screening range; Next, use three circles whose centers are on the centerline of the vehicle to envelop the shape of the vehicle, denoted as Use an ellipse to represent the perimeter vehicle. For each dynamic obstacle information, use Represents its position coordinates, represents the rotation matrix, a j and b j Represents the major and minor axes of the vehicle envelope ellipse. Based on this, the controller's obstacle avoidance constraint is converted to the intersection of the area occupied by the vehicle's envelope circle and the area occupied by the surrounding vehicle representation ellipse to be zero: in, It is the obstacle avoidance constraint representation; is the difference between the center of the self-circle and the center of the surrounding ellipse in the x and y directions, t∈[0,P], P is the prediction time domain; α=a+r i And β = b + r i is the union of the original surrounding vehicle ellipse and the ego vehicle circle, which is used to approximate the Minkowski sum of the surrounding vehicle ellipse, r i is the radius of the vehicle envelope circle; S35: Based on the above objective function and constraints, the vehicle motion planning problem is converted into a quadratic optimal planning problem of rolling optimization type within the prediction time domain P of the prediction network, which is specifically expressed as: in, It 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 solver initial value are iteratively updated through the vehicle model to enter the next round of optimization solution.

Images