CN119942790B - An Integrated Prediction-Planning Method for Autonomous Driving Based on Traffic Heterogeneous Graphs - Google Patents
An Integrated Prediction-Planning Method for Autonomous Driving Based on Traffic Heterogeneous GraphsInfo
- Publication number
- CN119942790B CN119942790B CN202510100416.9A CN202510100416A CN119942790B CN 119942790 B CN119942790 B CN 119942790B CN 202510100416 A CN202510100416 A CN 202510100416A CN 119942790 B CN119942790 B CN 119942790B
- Authority
- CN
- China
- Prior art keywords
- vehicle
- lane
- node
- information
- feature
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Traffic Control Systems (AREA)
Abstract
本发明涉及一种基于交通异构图的自动驾驶预测‑规划集成方法，属于自动驾驶汽车技术领域。该方法包括：S1：建模智能体动态特征以及交通参与者之间的交互特征，并设计基于图卷积网络的车道图节点特征表示；S2：采用编码‑解码架构，通过图注意力机制捕捉车道节点与智能体之间的交互特征，进行特征融合后构建基于交通异构图的多智能体轨迹预测模型；S3：基于轨迹预测模型输出的周围车辆未来位置信息，设计目标函数及多种约束来进行自车动作最优化求解。与传统规划方法相比，本发明预测性能、安全性以及行驶效率更加优异。
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
The invention belongs to the technical field of automatic driving automobiles, and relates to an automatic driving prediction-planning integration method based on traffic heterograms.
Background
In recent years, automatic driving has become one of research hotspots in the fields of artificial intelligence and transportation. The automatic driving system aims at realizing autonomous perception, decision and control of the vehicle in a complex road environment. The track prediction deduces the future motion trend of the target by analyzing the road scene, the traffic rules and the dynamic behaviors of traffic participants, and provides reliable prior information for downstream. After the motion planning receives the prediction information, the future track meeting safety, comfort and high efficiency is generated by combining the vehicle dynamics constraint, the driving environment and the driving target. Currently mainstream motion planning methods can be classified into rule-based, optimization-based and learning-based methods. Traditional rule-based methods, although clear in logic and strong in interpretation, rely on expert experience, require human task decomposition and defining trigger conditions, and are difficult to handle multi-body interactions and environmental dynamics. The driving task is defined as a mathematical optimization problem based on an optimization method, and the optimization solution is carried out by setting an objective function and constraint conditions. The learning-based method utilizes deep learning, reinforcement learning and other technologies to implicitly learn the target strategy from the data, and can flexibly adapt to complex traffic scenes.
The track prediction and motion planning two core modules are tightly coupled in function realization no matter in a modularized or end-to-end automatic driving frame, so that the feature transmission efficiency is improved and the accumulated error in the information transmission process is reduced. However, modeling of the traffic scene by the upstream information received by the current mainstream motion planning method is not enough, and the rich information provided by the track prediction module is not fully utilized. Therefore, a prediction planning integration method for efficiently fusing interactive prediction information is needed, and a coupling mechanism between the prediction information and the motion planning is deeply excavated, so that the running safety and the passing efficiency of the automatic driving vehicle are further improved.
Disclosure of Invention
In view of the above, the invention aims to provide an automatic driving prediction-planning integrated method based on traffic heterograms, which effectively models multimode interaction in complex interaction scenes, efficiently fuses peripheral vehicle prediction information and ensures safe and reasonable motion planning of a vehicle through optimization solution. Compared with the traditional planning method, the method has more excellent prediction performance, safety and running efficiency.
In order to achieve the above purpose, the present invention provides the following technical solutions:
an automatic driving prediction-planning integration method based on traffic abnormal composition specifically comprises the following steps:
s1, modeling dynamic characteristics of an intelligent agent and interactive characteristics among traffic participants, and designing a lane graph node characteristic representation based on a graph convolution network;
S2, capturing interaction characteristics between the lane nodes and the intelligent agents through a graph attention mechanism by adopting an encoding-decoding architecture, and constructing a multi-intelligent agent track prediction model based on traffic abnormal composition after feature fusion;
And S3, designing an objective function and various constraints based on the future position information of the surrounding vehicle output by the track prediction model to perform optimization solution of the vehicle motion.
Further, the step S1 specifically includes the following steps:
s11, defining current traffic information including driving lane information Historical track information of n intelligent agents at t moment
Wherein t 0 is the historical track time length; The motion state information of the ith agent at the moment t, Respectively representing coordinate information of the current moment of the vehicle, speed components in the x and y directions and course angle information;
S12, for the node characteristics of the intelligent agent, adopting a method based on a cyclic neural network to dynamically encode three traffic participant types of vehicles, bicycles and pedestrians, wherein for the intelligent agent A i, the encoding dynamic characteristics are as follows:
Wherein f i t represents the dynamics characteristics of the vehicle after the vehicle history track information is encoded, and GRU (graph) represents encoding through a cyclic neural network;
S13, modeling the interaction characteristics of the agents through a graph neural network, and firstly, designing dynamic screening conditions for each agent to select neighbor nodes:
wherein, the For the screened circumference radius, V i t is the current speed of the target node, L vehicle is the length of the vehicle, and lambda R is an empirically selected constant value;
S14, after the graph structure data representation is constructed, a graph attention mechanism is introduced to further strengthen the interaction characteristics among the vehicle nodes, and firstly, for a target node i and a neighbor node j screened by the target node i, the splicing characteristics z i,j of the directed edges of the target node i are expected to be obtained:
zi,j＝ReLU(Wa[hi‖ei,j‖hj])
Where h i,hj is the coding feature of the node, e i,j is the edge embedding feature from node i to node j, W a is the attention line transformation matrix, and II is the concatenation of features;
s15, after feature stitching of each edge of the target node is obtained, the attention score is obtained through normalization of softmax, and the feature of the target node is updated:
Wherein A i represents the updated target node characteristics, And the number of neighbors of the node i is represented, k represents the number of neighbor agent nodes in the screening range of the target node i, alpha i,j represents the importance degree of the node j to the node i, and W b is a full-connection layer.
Further, the step S2 specifically includes the following steps:
s21, based on the driving lane information defined in step S11 Converting the data into list data by a polynomial interpolation method, sampling an interpolation point every 1m as a lane node, and specifically representing as follows:
wherein m represents the number of lanes of the current scene, i is more than or equal to 1 and less than or equal to m, and j represents the total number of nodes of a single lane;
S22, based on the obtained lane node information, for the lane Adding a lane information matrix to expand lane node characteristics:
Wherein, l u,lv is a set of index values of the Lane nodes, lane pre matrix indicates that the Lane node with index l v is a preceding node of the corresponding Lane node with index l u, lane suc represents a subsequent node, lane left,Laneright represents a left adjacent node and a right adjacent node respectively;
s23, for the established position information relation, carrying out characteristic enhancement by utilizing a graph convolution idea, and updating the characteristics of the target node by using the following formula:
wherein, the As a feature of the node of the target lane,In order to perform the linear layer of feature mapping,The specific calculation mode is as follows:
wherein, the Is a full connection layer for performing linear transformation on target features;
S24, in the prediction task, the feature that only adjacent lane nodes are gathered for the target lane node is insufficient in the aspect of lane information characterization, and the relation needs to be further expanded between more distant nodes, so that the pre and the suc type nodes are expanded by k steps, and node feature update is changed into:
The lane node indexed by l v_k is the successor/successor kth node indexed by l u_k, and the k value is usually 6 within the effective range of the lane node sequence; A linear transformation layer corresponding to the expansion node;
s25, introducing an attention mechanism to model vehicle-lane interaction characteristics, and obtaining an agent characteristic A and lane node characteristics Under the condition of (1), traversing the intelligent body nodes and the lane nodes, screening out node pairs with the distance meeting the range requirement to obtain a corresponding index list, c i,cj, recording the distance information of each pair of nodes meeting the condition, carrying out linear transformation on the screened characteristic information to obtain an intelligent body inquiry characteristic query and a distance characteristic dist, further obtaining characteristic cascade connection of the intelligent body nodes and the interactive lane nodes by the distance information based on the information, and carrying out characteristic superposition:
Ctx=Wc(query‖dist‖ctx)
Wherein Ctx is the selected lane node feature, ctx is the fusion vector of the query feature, the distance feature and the lane node aggregation feature, and after being accumulated with the corresponding agent node feature, the vehicle lane interaction feature la, W l、Wc and the regularization layer Norm are obtained through the activation layer Relu All-connected layer for carrying out linear transformation on corresponding characteristics
S26, carrying out dimension processing on the interaction characteristics of the intelligent agent node i and the lane, fusing the interaction characteristics with the previous intelligent agent dynamic characteristics and the interaction characteristics, and decoding a future prediction track under the intelligent agent prediction time domain P:
Enc=fuse(f+A+LA)
And LA i is the vehicle lane interactive characteristic after the average pooling operation, and the comprehensive characteristic code Enc is obtained after the characteristic fusion by combining the obtained intelligent body dynamic characteristic f and the interactive characteristic A. Also obtaining future trajectories of agents through a cyclic neural network based decoder LSTM
Further, the step S3 specifically includes the following steps:
S31, defining a state quantity h and a control quantity u of a system by combining a bicycle model of the vehicle:
Where a in u represents acceleration of the vehicle, δ is steering angle of the front wheel, v_s is speed along the reference path in the arc coordinate system, s_v is relaxation factor, and variables in h are current position coordinates x, y and heading angle of the vehicle, respectively And a vehicle speed v;
s32, comprehensively considering key factors influencing automatic driving, and designing an objective function J (u t,ht,st) based on model predictive control:
J(ut,ht,st)＝Jfollow(ht,st)+Jv(ht)+Ju(ut)+JLF(ht,st)
Wherein s t is the position coordinate of the vehicle t moment in an arc coordinate system, J follow(ht,st) is a path following objective function, J v(ht) is a speed maintaining objective function, J u(ut) is a control action amount objective function, J LF(ht,st) is a road potential field objective function, and the specific calculation formulas of the objective functions are as follows:
wherein, the AndRepresenting error values of the real vehicle position and the approximate vehicle position in the arc length direction and the vehicle side direction in path following respectively, w s and w l are weight matrixes,S t is the reference arc length coordinate at the vehicle position at the current moment, which is the state quantity of the vehicle at the moment t;
wherein, the The square term representing the variable is multiplied by a weight value, J v represents a tracking target speed term, v t is a self-speed at t moment, v ref represents a target speed, and J u is an objective function formed by control variables of a solver and comprises acceleration a t, steering angle delta t, arc speed v_s and relaxation factor s_v;
The method is characterized in that the method is an objective function designed based on road constraint, and is expected to normalize the movement of a vehicle in a lane and drive near the center line of the lane, and meanwhile, a lower cost value is given to an adjacent lane compared with the lane boundary to ensure that the possibility of lane changing of the vehicle is considered when the vehicle is in obstacle avoidance, wherein y L is the distance between the vehicle and the lane boundary, L width is the lane width, and w LF is a weight coefficient matrix of the term;
s33, when the objective function is optimally solved through model predictive control, partial variables in the objective function are required to be constrained, so that the objective function is ensured to be in a reasonable calculation range, and the specific constraints are as follows:
hmin≤ht≤hmax
umin≤ut≤umax
wherein h min、hmax、umin、umax and Upper and lower limits of the state quantity and the control quantity respectively,As the heading angle of the current vehicle,Heading angle for an approximate point of the vehicle on the reference path;
S34, in the planning step, after Zhou Che track prediction information is packaged and processed, the track prediction information is used as a dynamic obstacle and is constrained by a vehicle planning controller formed through a contour error model, and the predicted track of Zhou Che is defined as follows:
wherein, the For the output of the trajectory prediction model,The future motion trail j of surrounding vehicles is the number of neighbor intelligent agents in the screening range;
Next, the shape of the vehicle is enveloped by a circle with three centers on the vehicle center line, denoted as The ellipse is used to represent the surrounding vehicle, and for each dynamic obstacle information, the ellipse is used to represent the surrounding vehicleWhich represents the coordinates of the position thereof,Representing a rotation matrix, a j and b j represent major and minor axes of an envelope ellipse of the vehicle, based on which the obstacle avoidance constraints of the controller can be converted to zero intersection of the envelope circle footprint of the vehicle itself with the surrounding vehicle representing the ellipse footprint:
wherein, the In order to avoid the representation of the obstacle constraint,The difference between the center of the vehicle circle and the center of Zhou Che ellipse in the x, y direction is t epsilon [0, P ], P is the prediction time domain, alpha=a+r i and beta=b+r i are the union of the original surrounding vehicle ellipse and the vehicle circle, so as to approximate the minkowski sum of the surrounding vehicle ellipse, and r i is the radius of the vehicle envelope circle;
S35, based on the objective function and the constraint condition, in a prediction time domain P of a prediction network, the motion planning problem of the vehicle can be converted into a rolling optimization type secondary optimal planning problem, which can be expressed as:
wherein, the And after the optimal control sequence is obtained, finishing iterative updating of the self-vehicle state and the initial value of the solver through the vehicle model, and entering the next round of optimization solving.
The method has the advantages that the method is used for capturing the multi-mode interaction characteristics in the complex traffic task efficiently, simultaneously fusing the model predictive control theory, dynamically updating the system state by introducing the predictive information of the surrounding vehicles, and generating reasonable and safe motion planning for the own vehicle. Compared with the traditional planning method, the method has the advantages that the interactive modeling precision is remarkably improved, and the traffic efficiency and the behavior robustness are further optimized on the premise of ensuring the safety.
(1) The invention encodes the dynamic characteristics of the vehicle history characteristics based on the cyclic neural network and captures the behavior interaction between vehicles through a multi-head attention mechanism as the vehicle nodes. Meanwhile, a dynamic neighbor screening method is designed, and an interaction diagram among traffic participants is more efficiently constructed.
(2) The invention constructs a novel lane node information matrix based on the vehicle driving lane information, and designs a corresponding graph convolution mode to model and expand the driving characteristics of the lane. The attention mechanism is introduced to model interaction characteristics of vehicle nodes and weighted lane nodes, so that the vehicle is helped to better understand the current driving environment and make more accurate track behavior prediction.
(3) The invention introduces the output information of the interactive prediction model into the safety motion planning of the own vehicle through the thought of model prediction control, the own vehicle predicts the future state, supplements the prediction information as system state update, and designs an objective function and constraint conditions to solve the current optimal acceleration and steering angle of the vehicle. The prediction information is efficiently utilized, and meanwhile, safe and reasonable self-vehicle path planning can be made.
Additional advantages, objects, and features of the invention will be set forth in part in the description which follows and in part will become apparent to those having ordinary skill in the art upon examination of the following or may be learned from practice of the invention. The objects and other advantages of the invention may be realized and obtained by means of the instrumentalities and combinations particularly pointed out in the specification.
Drawings
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in the following preferred detail with reference to the accompanying drawings, in which:
FIG. 1 is an overall flow chart of an automated driving prediction-planning integration method based on traffic heterographs of the present invention;
FIG. 2 is a flow chart for building a multi-agent trajectory prediction model based on a traffic heterogeneous map;
Fig. 3 is a flow chart of an automatic driving planning method integrating prediction information.
Detailed Description
Other advantages and effects of the present invention will become apparent to those skilled in the art from the following disclosure, which describes the embodiments of the present invention with reference to specific examples. The invention may be practiced or carried out in other embodiments that depart from the specific details, and the details of the present description may be modified or varied from the spirit and scope of the present invention. It should be noted that the illustrations provided in the following embodiments merely illustrate the basic idea of the present invention by way of illustration, and the following embodiments and features in the embodiments may be combined with each other without conflict.
Referring to fig. 1-3, the invention provides an automatic driving prediction-planning integration method based on traffic heterograms, which firstly models the dynamic characteristics of an intelligent agent and the interactive characteristics among traffic participants, and designs a lane graph node characteristic representation based on a graph rolling network as shown in fig. 1. And secondly, capturing interaction between the lane nodes and the intelligent agents through an attention mechanism, and decoding and outputting the predicted track of the multi-intelligent agent after feature fusion. Finally, the prediction information is introduced through the system state establishment and real-time updating of the model prediction control, and the objective function and various constraints are designed to perform the optimization solution of the vehicle motion.
As shown in fig. 2, constructing a multi-agent trajectory prediction model based on traffic heterograms, specifically comprising the following steps:
s21, based on the driving lane information defined in step S11 Converting the data into list data by a polynomial interpolation method, sampling an interpolation point every 1m as a lane node, and specifically representing as follows:
wherein m represents the number of lanes of the current scene, i is more than or equal to 1 and less than or equal to m, and j represents the total number of nodes of a single lane;
S22, based on the obtained lane node information, for the lane Adding a lane information matrix to expand lane node characteristics:
Wherein, l u,lv is a set of index values of the Lane nodes, lane pre matrix indicates that the Lane node with index l v is a preceding node of the corresponding Lane node with index l u, lane suc represents a subsequent node, lane left,Laneright represents a left adjacent node and a right adjacent node respectively;
s23, for the established position information relation, carrying out characteristic enhancement by utilizing a graph convolution idea, and updating the characteristics of the target node by using the following formula:
wherein, the As a feature of the node of the target lane,In order to perform the linear layer of feature mapping,The specific calculation mode is as follows:
wherein, the Is a full connection layer for performing linear transformation on target features;
S24, in the prediction task, the feature that only adjacent lane nodes are gathered for the target lane node is insufficient in the aspect of lane information characterization, and the relation needs to be further expanded between more distant nodes, so that the pre and the suc type nodes are expanded by k steps, and node feature update is changed into:
Wherein, the lane node indexed by l v_k is the successor/successor kth node indexed by l u_k, the k value is usually 6 within the effective range of the lane node sequence, A linear transformation layer corresponding to the expansion node;
s25, introducing an attention mechanism to model vehicle-lane interaction characteristics, and obtaining an agent characteristic A and lane node characteristics Under the condition of (1), traversing the intelligent body nodes and the lane nodes, screening out node pairs with the distance meeting the range requirement to obtain a corresponding index list, c i,cj, recording the distance information of each pair of nodes meeting the condition, carrying out linear transformation on the screened characteristic information to obtain an intelligent body inquiry characteristic query and a distance characteristic dist, further obtaining characteristic cascade connection of the intelligent body nodes and the interactive lane nodes by the distance information based on the information, and carrying out characteristic superposition:
Ctx=Wc(query‖dist‖ctx)
Wherein Ctx is the selected lane node feature, ctx is the fusion vector of the query feature, the distance feature and the lane node aggregation feature, and after being accumulated with the corresponding agent node feature, the vehicle lane interaction feature la, W l、Wc and the regularization layer Norm are obtained through the activation layer Relu All are full-connection layers which perform linear transformation on corresponding characteristics;
S26, carrying out dimension processing on the interaction characteristics of the intelligent agent node i and the lane, fusing the interaction characteristics with the previous intelligent agent dynamic characteristics and the interaction characteristics, and decoding a future prediction track under the intelligent agent prediction time domain P:
Enc=fuse(f+A+LA)
And LA i is the vehicle lane interactive characteristic after the average pooling operation, and the comprehensive characteristic code Enc is obtained after the characteristic fusion by combining the obtained intelligent body dynamic characteristic f and the interactive characteristic A. Also obtaining future trajectories of agents through a cyclic neural network based decoder LSTM
As shown in fig. 3, the automatic driving planning method for fusing the prediction information specifically includes the following steps:
S31, defining a state quantity h and a control quantity u of a system by combining a bicycle model of the vehicle:
Where a in u represents acceleration of the vehicle, δ is steering angle of the front wheel, v_s is speed along the reference path in the arc coordinate system, s_v is relaxation factor, and variables in h are current position coordinates x, y and heading angle of the vehicle, respectively And a vehicle speed v;
s32, comprehensively considering key factors influencing automatic driving, and designing an objective function J (u t,ht,st) based on model predictive control:
J(ut,ht,st)＝Jfollow(ht,st)+Jv(ht)+Ju(ut)+JLF(ht,st)
Wherein s t is the position coordinate of the vehicle t moment in an arc coordinate system, J follow(ht,st) is a path following objective function, J v(ht) is a speed maintaining objective function, J u(ut) is a control action amount objective function, J LF(ht,st) is a road potential field objective function, and the specific calculation formulas of the objective functions are as follows:
wherein, the AndRepresenting error values of the real vehicle position and the approximate vehicle position in the arc length direction and the vehicle side direction in path following respectively, w s and w l are weight matrixes,S t is the reference arc length coordinate at the vehicle position at the current moment, which is the state quantity of the vehicle at the moment t;
wherein, the The square term representing the variable is multiplied by a weight value, J v represents a tracking target speed term, v t is a self-speed at t moment, v ref represents a target speed, and J u is an objective function formed by control variables of a solver and comprises acceleration a t, steering angle delta t, arc speed v_s and relaxation factor s_v;
The method is characterized in that the method is an objective function designed based on road constraint, and is expected to normalize the movement of a vehicle in a lane and drive near the center line of the lane, and meanwhile, a lower cost value is given to an adjacent lane compared with the lane boundary to ensure that the possibility of lane changing of the vehicle is considered when the vehicle is in obstacle avoidance, wherein y L is the distance between the vehicle and the lane boundary, L width is the lane width, and w LF is a weight coefficient matrix of the term;
s33, when the objective function is optimally solved through model predictive control, partial variables in the objective function are required to be constrained, so that the objective function is ensured to be in a reasonable calculation range, and the specific constraints are as follows:
hmin≤ht≤hmax
umin≤ut≤umax
wherein h min、hmax、umin、umax and Upper and lower limits of the state quantity and the control quantity respectively,As the heading angle of the current vehicle,Heading angle for an approximate point of the vehicle on the reference path;
S34, in the planning step, after Zhou Che track prediction information is packaged and processed, the track prediction information is used as a dynamic obstacle and is constrained by a vehicle planning controller formed through a contour error model, and the predicted track of Zhou Che is defined as follows:
wherein, the For the output of the trajectory prediction model,J is the number of neighbor intelligent agents in the screening range for the future motion trail of surrounding vehicles;
Next, the shape of the vehicle is enveloped by a circle with three centers on the vehicle center line, denoted as The ellipse is used to represent the surrounding vehicle, and for each dynamic obstacle information, the ellipse is used to represent the surrounding vehicleWhich represents the coordinates of the position thereof,Representing a rotation matrix, a j and b j represent major and minor axes of an envelope ellipse of the vehicle, based on which the obstacle avoidance constraints of the controller can be converted to zero intersection of the envelope circle footprint of the vehicle itself with the surrounding vehicle representing the ellipse footprint:
wherein, the In order to avoid the representation of the obstacle constraint,The difference between the center of the vehicle circle and the center of Zhou Che ellipse in the x, y direction is t epsilon [0, P ], P is the prediction time domain, alpha=a+r i and beta=b+r i are the union of the original surrounding vehicle ellipse and the vehicle circle, so as to approximate the minkowski sum of the surrounding vehicle ellipse, and r i is the radius of the vehicle envelope circle;
S35, based on the objective function and the constraint condition, in a prediction time domain P of a prediction network, the motion planning problem of the vehicle can be converted into a rolling optimization type secondary optimal planning problem, which can be expressed as:
wherein, the And after the optimal control sequence is obtained, finishing iterative updating of the self-vehicle state and the initial value of the solver through the vehicle model, and entering the next round of optimization solving.
Finally, it is noted that the above embodiments are only for illustrating the technical solution of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that modifications and equivalents may be made thereto without departing from the spirit and scope of the present invention, which is intended to be covered by the claims of the present invention.
Claims (3)
1. An automatic driving prediction-planning integration method based on traffic abnormal composition is characterized by comprising the following steps:
S1, modeling dynamic characteristics of agents and interaction characteristics among traffic participants, and designing dynamic screening conditions for each agent to select neighbor nodes:
wherein, the For the screened circumference radius, V i t is the current speed of the target node, L vehicle is the length of the vehicle, and lambda R is an empirically selected constant value;
S2, designing a lane graph node characteristic representation based on a graph rolling network, capturing interaction characteristics between lane nodes and intelligent agents through a graph attention mechanism by adopting an encoding-decoding architecture, and constructing a multi-intelligent agent track prediction model based on traffic abnormal composition after characteristic fusion, wherein the step S2 specifically comprises the following steps:
S21 based on the driving lane information Converting the data into list data by a polynomial interpolation method, sampling an interpolation point every 1m as a lane node, and specifically representing as follows:
wherein m represents the number of lanes 1≤i≤m of the current scene, and j represents the total number of nodes of a single lane;
S22, based on the obtained lane node information, for the lane Adding a lane information matrix to expand lane node characteristics:
Wherein, l u,lv is a set of index values of the Lane nodes, lane pre matrix indicates that the Lane node with index l v is a preceding node of the corresponding Lane node with index l u, lane suc represents a subsequent node, lane left,Laneright represents a left adjacent node and a right adjacent node respectively;
s23, for the established position information relation, carrying out characteristic enhancement by utilizing a graph convolution idea, and updating the characteristics of the target node by using the following formula:
wherein, the As a feature of the node of the target lane,In order to perform the linear layer of feature mapping,The specific calculation mode is as follows:
wherein, the Is a full connection layer for performing linear transformation on target features;
s24, expanding pre and sub type nodes by k steps, and updating node characteristics into:
wherein, the lane node indexed by l v_k is the successor/successor kth node indexed by l u_k; a linear transformation layer corresponding to the expansion node characteristics;
s25, introducing an attention mechanism to model vehicle-lane interaction characteristics, and obtaining an agent characteristic A and lane node characteristics Under the condition of (1), traversing the intelligent body nodes and the lane nodes, screening out node pairs with the distance meeting the range requirement to obtain a corresponding index list, c i,cj, recording the distance information of each pair of nodes meeting the condition, carrying out linear transformation on the screened characteristic information to obtain an intelligent body inquiry characteristic query and a distance characteristic dist, further obtaining characteristic cascade connection of the intelligent body nodes and the interactive lane nodes by the distance information based on the information, and carrying out characteristic superposition:
Ctx=Wc(query‖dist‖ctx)
Wherein Ctx is the selected lane node feature, ctx is the fusion vector of the query feature, the distance feature and the lane node aggregation feature, and after being accumulated with the corresponding agent node feature, the vehicle lane interaction feature la, W l、Wc and the regularization layer Norm are obtained through the activation layer Relu All are full-connection layers which perform linear transformation on corresponding characteristics;
S26, carrying out dimension processing on the interaction characteristics of the intelligent agent node i and the lane, fusing the interaction characteristics with the previous intelligent agent dynamic characteristics and the interaction characteristics, and decoding a future prediction track under the intelligent agent prediction time domain P:
Enc=fuse(f+A+LA)
wherein LA i is vehicle lane interactive feature after average pooling operation, and the obtained intelligent agent dynamic feature f and interactive feature A are combined to obtain comprehensive feature code Enc after feature fusion, and the future track of intelligent agent is obtained by decoder LSTM based on cyclic neural network
And S3, designing an objective function and various constraints based on the future position information of the surrounding vehicle output by the track prediction model to perform optimization solution of the vehicle motion.
2. The automated driving prediction-planning integration method according to claim 1, wherein step S1 specifically comprises the steps of:
s11, defining current traffic information including driving lane information Historical track information of n intelligent agents at t moment
Wherein t 0 is the historical track time length; Motion state information of the ith agent at time t, wherein Respectively representing coordinate information of the ith intelligent agent at the current moment, speed components in x and y directions and course angle information;
S12, for the node characteristics of the intelligent agent, adopting a method based on a cyclic neural network to dynamically encode three traffic participant types of vehicles, bicycles and pedestrians, wherein for the intelligent agent A i, the encoding dynamic characteristics are as follows:
Wherein f i t represents the dynamics characteristics of the vehicle after the vehicle history track information is encoded, and GRU (graph) represents encoding through a cyclic neural network;
S13, modeling the interaction characteristics of the agents through a graph neural network, and firstly, designing dynamic screening conditions for each agent to select neighbor nodes;
S14, after the graph structure data representation is constructed, a graph attention mechanism is introduced to further strengthen the interaction characteristics among the vehicle nodes, and firstly, for a target node i and a neighbor node j screened by the target node i, the splicing characteristics z i,j of the directed edges of the target node i are expected to be obtained:
zi,j＝ReLU(Wa[hi‖ei,j‖hj])
Where h i,hj is the coding feature of the node, e i,j is the edge embedding feature from node i to node j, W a is the attention line transformation matrix, and II is the concatenation of features;
s15, after feature stitching of each edge of the target node is obtained, the attention score is obtained through normalization of softmax, and the feature of the target node is updated:
Wherein A i represents the updated target node characteristics, And the number of neighbors of the node i is represented, k represents the number of neighbor agent nodes in the screening range of the target node i, alpha i,j represents the importance degree of the node j to the node i, and W b is a full-connection layer.
3. The automated driving prediction-planning integration method according to claim 2, characterized in that step S3 comprises in particular the steps of:
S31, defining a state quantity h and a control quantity u of a system by combining a bicycle model of the vehicle:
Where a in u represents acceleration of the vehicle, δ is steering angle of the front wheel, v_s is speed along the reference path in the arc coordinate system, s_v is relaxation factor, and variables in h are current position coordinates x, y and heading angle of the vehicle, respectively And a vehicle speed v;
s32, comprehensively considering key factors influencing automatic driving, and designing an objective function J (u t,ht,st) based on model predictive control:
J(ut,ht,st)＝Jfollow(ht,st)+Jv(ht)+Ju(ut)+JLF(ht,st)
Wherein s t is the position coordinate of the vehicle t moment in an arc coordinate system, J follow(ht,st) is a path following objective function, J v(ht) is a speed maintaining objective function, J u(ut) is a control action amount objective function, J LF(ht,st) is a road potential field objective function, and the specific calculation formulas of the objective functions are as follows:
wherein, the AndRepresenting error values of the real vehicle position and the approximate vehicle position in the arc length direction and the vehicle side direction in path following respectively, w s and w l are weight matrixes,S t is the reference arc length coordinate at the vehicle position at the current moment, which is the state quantity of the vehicle at the moment t;
wherein, the The square term representing the variable is multiplied by a weight value, J v represents a tracking target speed term, v t is a self-speed at t moment, v ref represents a target speed, and J u is an objective function formed by control variables of a solver and comprises acceleration a t, steering angle delta t, arc speed v_s and relaxation factor s_v;
The method is characterized in that the method is an objective function designed based on road constraint, and is expected to normalize the movement of a vehicle in a lane and drive near the center line of the lane, and meanwhile, a low cost value is given to an adjacent lane compared with the lane boundary to ensure that the possibility of lane change of the vehicle is considered when the vehicle is in obstacle avoidance, wherein y L is the distance between the vehicle and the lane boundary, L width is the lane width, and w LF is a weight coefficient matrix of the term;
s33, when the objective function is optimally solved through model predictive control, partial variables in the objective function are required to be constrained, so that the objective function is ensured to be in a reasonable calculation range, and the specific constraints are as follows:
wherein h min、hmax、umin、umax and Upper and lower limits of the state quantity and the control quantity respectively,As the heading angle of the current vehicle,Heading angle for an approximate point of the vehicle on the reference path;
S34, in the planning step, after Zhou Che track prediction information is packaged and processed, the track prediction information is used as a dynamic obstacle and is constrained by a vehicle planning controller formed through a contour error model, and the predicted track of Zhou Che is defined as follows:
wherein, the For the output of the trajectory prediction model,J is the number of neighbor intelligent agents in the screening range for the future motion trail of surrounding vehicles;
Next, the shape of the vehicle is enveloped by a circle with three centers on the vehicle center line, denoted as The ellipse is used to represent the surrounding vehicle, and for each dynamic obstacle information, the ellipse is used to represent the surrounding vehicleWhich represents the coordinates of the position thereof,Representing a rotation matrix, a j and b j represent major and minor axes of an envelope ellipse of the vehicle, based on which the obstacle avoidance constraint of the controller is converted into zero intersection of the envelope circle footprint of the vehicle itself and the surrounding vehicle representing the ellipse footprint:
wherein, the Representing for obstacle avoidance constraint; The difference between the center of the vehicle circle and the center of Zhou Che ellipse in the x, y direction is t epsilon [0, P ], P is the prediction time domain, alpha=a+r i and beta=b+r i are the union of the original surrounding vehicle ellipse and the vehicle circle, so as to approximate the minkowski sum of the surrounding vehicle ellipse, and r i is the radius of the vehicle envelope circle;
S35, converting the motion planning problem of the vehicle into a rolling optimization type secondary optimal planning problem in a prediction time domain P of a prediction network based on the objective function and the constraint condition, wherein the rolling optimization type secondary optimal planning problem is specifically expressed as follows:
wherein, the And after the optimal control sequence is obtained, finishing iterative updating of the self-vehicle state and the initial value of the solver through the vehicle model, and entering the next round of optimization solving.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202510100416.9A CN119942790B (en) | 2025-01-22 | 2025-01-22 | An Integrated Prediction-Planning Method for Autonomous Driving Based on Traffic Heterogeneous Graphs |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202510100416.9A CN119942790B (en) | 2025-01-22 | 2025-01-22 | An Integrated Prediction-Planning Method for Autonomous Driving Based on Traffic Heterogeneous Graphs |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN119942790A CN119942790A (en) | 2025-05-06 |
| CN119942790B true CN119942790B (en) | 2025-10-31 |
Family
ID=95548207
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202510100416.9A Active CN119942790B (en) | 2025-01-22 | 2025-01-22 | An Integrated Prediction-Planning Method for Autonomous Driving Based on Traffic Heterogeneous Graphs |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN119942790B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN120636171B (en) * | 2025-08-13 | 2025-10-10 | 湖南工商大学 | Traffic violation detection method, device, equipment and medium based on UV-KGNN network |
| CN121291269B (en) * | 2025-12-11 | 2026-03-17 | 常州星宇车灯股份有限公司 | Intelligent headlamp control method and control system based on driving intention prediction |
| CN121386791B (en) * | 2025-12-24 | 2026-02-24 | 长春工业大学 | Deep learning track prediction method introducing vehicle kinematic constraint |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111559388A (en) * | 2020-03-26 | 2020-08-21 | 吉利汽车研究院（宁波）有限公司 | A target vehicle screening method, device, equipment and storage medium |
| CN113954864A (en) * | 2021-09-22 | 2022-01-21 | 江苏大学 | Intelligent automobile track prediction system and method fusing peripheral vehicle interaction information |
Family Cites Families (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113989330B (en) * | 2021-11-03 | 2025-02-25 | 中国电信股份有限公司 | Vehicle trajectory prediction method, device, electronic device and readable storage medium |
| CN115147790B (en) * | 2022-06-28 | 2024-06-04 | 重庆长安汽车股份有限公司 | Future track prediction method of vehicle based on graph neural network |
| CN115179959B (en) * | 2022-07-18 | 2024-08-02 | 福州大学 | Intelligent driving vehicle behavior prediction method based on adaptive update threshold of driving road |
| US12187324B2 (en) * | 2022-08-31 | 2025-01-07 | Zoox, Inc. | Trajectory prediction based on a decision tree |
| CN119007482A (en) * | 2023-05-17 | 2024-11-22 | 鸿海精密工业股份有限公司 | Method and computing system for predicting vehicle motion |
| CN117031961A (en) * | 2023-09-06 | 2023-11-10 | 重庆大学 | Interactive decision-making planning method for autonomous vehicles based on model predictive control |
| CN117315603A (en) * | 2023-09-11 | 2023-12-29 | 福建师范大学 | Multimodal spatiotemporal model for accurate motion prediction based on visual fusion |
| CN118062052A (en) * | 2024-01-09 | 2024-05-24 | 重庆大学 | A safe motion planning method for autonomous vehicles considering driving behavior |
-
2025
- 2025-01-22 CN CN202510100416.9A patent/CN119942790B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111559388A (en) * | 2020-03-26 | 2020-08-21 | 吉利汽车研究院（宁波）有限公司 | A target vehicle screening method, device, equipment and storage medium |
| CN113954864A (en) * | 2021-09-22 | 2022-01-21 | 江苏大学 | Intelligent automobile track prediction system and method fusing peripheral vehicle interaction information |
Also Published As
| Publication number | Publication date |
|---|---|
| CN119942790A (en) | 2025-05-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN119942790B (en) | An Integrated Prediction-Planning Method for Autonomous Driving Based on Traffic Heterogeneous Graphs | |
| Ma et al. | Reinforcement learning for autonomous driving with latent state inference and spatial-temporal relationships | |
| CN114013443B (en) | Automatic driving vehicle lane change decision control method based on hierarchical reinforcement learning | |
| CN114038212B (en) | Signal lamp control method based on two-stage attention mechanism and deep reinforcement learning | |
| CN114169241B (en) | End-to-end multi-target identification, tracking and prediction method | |
| CN112907967A (en) | Intelligent vehicle lane change decision-making method based on incomplete information game | |
| CN114153213A (en) | Deep reinforcement learning intelligent vehicle behavior decision method based on path planning | |
| CN114372570A (en) | A Multimodal Vehicle Trajectory Prediction Method | |
| CN113657433A (en) | A Multimodal Prediction Method of Vehicle Trajectory | |
| CN117315603A (en) | Multimodal spatiotemporal model for accurate motion prediction based on visual fusion | |
| CN119058729B (en) | A trailer real-time traffic information collection and processing system based on artificial intelligence | |
| US12116016B1 (en) | Decision-making and planning integrated method for nonconservative intelligent vehicle | |
| CN116476825B (en) | Automatic driving lane keeping control method based on safe and reliable reinforcement learning | |
| CN119357627B (en) | Multimodal feature extraction network-enhanced negotiation decision-making method based on spatiotemporal directed graph | |
| CN117657214A (en) | An autonomous driving decision planning method coupled with deep models and rules | |
| CN118917179A (en) | Multi-mode reinforcement learning vehicle decision-making planning method with compensation feedback | |
| CN112550314A (en) | Embedded optimization type control method suitable for unmanned driving, driving control module and automatic driving control system thereof | |
| CN117007066A (en) | Unmanned driving trajectory planning method and related devices integrating multiple planning algorithms | |
| CN119682776B (en) | A method and system for prompting optimal lane-changing timing for vehicles based on pre-trained LLMs trajectory prediction | |
| CN118770285B (en) | A hierarchical decision-making control method for autonomous driving in ramp merging scenarios | |
| Xiaoqiang et al. | Graph convolution reinforcement learning for decision-making in highway overtaking scenario | |
| Chen et al. | A novel generalized meta hierarchical reinforcement learning method for autonomous vehicles | |
| CN114912693B (en) | A motion planning method for autonomous vehicles based on multimodal prediction | |
| CN111443701A (en) | Unmanned vehicle/robot behavior planning method based on heterogeneous deep learning | |
| CN121291466A (en) | A unified framework for solving the consistency between trajectory prediction and planning in autonomous driving based on a world model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |