An unmanned vehicle behavior decision-making method for a human-vehicle mixed road environment

Abstract

The application discloses a kind of unmanned vehicle behavior decision-making methods for human-vehicle mixed road environment, belong to the intelligent decision-making technical field of unmanned vehicle. Including: obtaining the historical data of different types of traffic participants;Establish the graph representation of multi-agent interactive environment, including node feature matrix, adjacency matrix and supplementary adjacency matrix;Traffic scene interaction feature extraction is carried out using the supplementary adjacency matrix, and is fused with the node feature matrix, to obtain the updated node feature matrix;The adjacency matrix and the updated node feature matrix are input into the decision model to complete the behavior decision-making.The application projects the edge feature and the edge type into the same feature space, splices to form the feature matrix containing the edge for traffic scene interaction feature extraction, and fuses with the node feature for information enhancement, to solve the problem that GAT cannot handle the edge feature in the interaction graph, help the understanding of vehicle behavior, and improve the prediction ability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent decision-making of unmanned vehicles, and particularly relates to a behavior decision-making method for unmanned vehicles in a mixed road environment of human and vehicles. BACKGROUND

[0002] The behavior decision-making model of an unmanned vehicle is a core link for realizing automatic driving, and its goal is to simulate the decision-making process of a human driver to adapt to a complex traffic environment under the premise of ensuring safety. In an actual traffic environment, there are various types of traffic participants including vehicles, non-motor vehicles, and pedestrians, and a human driver can comprehensively consider different characteristics of different types of traffic participants to make a safe and reasonable driving behavior decision. In order to realize similar or even superior decision-making ability to humans, an unmanned vehicle must consider different types of traffic participants in the traffic environment, and cannot only consider a single type.

[0003] As a kind of graph neural network, the graph attention network (GAT) provides strong information integration and decision support capability for unmanned vehicles, allowing the vehicle to effectively understand and cope with the complex traffic environment around it, thereby improving the performance and safety of the automatic driving system. When the unmanned vehicle makes a lane-changing decision, vehicles closer to the vehicle may be assigned a larger weight compared to vehicles farther away. In addition, a larger attention weight is assigned to traffic participants in front of the ego vehicle and overlapping with the ego vehicle in the longitudinal direction compared to traffic participants behind the ego vehicle. In GAT, this relative importance is automatically encoded as a weight, and the weight is learned from successful and unsuccessful training processes.

[0004] The input of GAT usually includes a node feature matrix and an adjacency matrix, which contain the original information of each vehicle such as speed, position, and intention, etc. The output of GAT is a node feature vector, which is obtained by fusing original data from different sources, including both local environmental information and more extensive global environmental information. The output node vector is mapped into the key information required by the unmanned vehicle to make driving decisions in the traffic environment it is in. The node information processed by GAT has better information integration and abstraction capability to more effectively support the formulation of intelligent driving decisions.

[0005] When autonomous vehicles make behavioral decisions, the interactions between traffic participants are a decisive factor influencing their driving behavior; therefore, the interaction information carried by edges cannot be ignored. However, the current GAT (Graph Aggregation and Analysis) method, during feature aggregation and updating of the graph, ignores the information carried by edges and only considers the relationships between nodes. This may cause the model to fail to fully utilize edge features, thus affecting its ability to understand and predict vehicle behavior. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention designs a heterogeneous edge-enhanced graph attention network (EGAT), which updates node features by aggregating information from the neighborhood, i.e., it gathers traffic participant information within the sensor range of the autonomous vehicle.

[0007] To achieve the above objectives, this invention proposes a method for decision-making regarding the behavior of unmanned vehicles in mixed human-vehicle road environments, comprising the following steps:

[0008] (1) Obtain historical data of different types of traffic participants, including motor vehicles, non-motor vehicles and pedestrians; the historical data of the participants includes the position, speed and heading angle of the traffic participants in the past time t.

[0009] (2) Establish a graph representation of the multi-agent interaction environment, including node feature matrix, adjacency matrix and supplementary adjacency matrix;

[0010] The historical data of the different types of traffic participants are encoded using a gated recurrent unit (GRU) to obtain dynamic features, and these dynamic features are added to the node feature vector to construct a node feature matrix.

[0011] The adjacency matrix is ​​constructed by using Gaussian process regression (GPR) to represent the Gaussian velocity field.

[0012] A supplementary adjacency matrix is ​​constructed using the edge feature matrix and edge type matrix of the relationships between the different types of traffic participants.

[0013] (3) Use the supplementary adjacency matrix to extract traffic scene interaction features and fuse them with the node feature matrix to obtain the updated node feature matrix;

[0014] (4) Input the adjacency matrix and the updated node feature matrix into the graph attention neural network to extract interaction features, and complete the behavior decision through the decision model.

[0015] Furthermore, the node feature matrix is ​​specifically as follows:

[0016] The historical data of the traffic participant is encoded by using the GRU to extract the dynamic characteristics of the traffic participant

[0017]

[0018] Wherein: represents a gated recurrent unit GRU, represents the dynamic characteristics of the traffic participant at the moment t, t represents the historical data of the traffic participant at the moment t. i t i

[0019] The dynamic characteristics are added to the node feature vector to construct a node feature matrix.

[0020]

[0021] Wherein: is a node i at the moment t, and the vector contains the current speed information, position information, heading angle, acceleration, surrounding map information and dynamic characteristics of the traffic participant. t i i

[0022] Further, the adjacency matrix is specifically:

[0023] The probability distribution of the vehicle speed field is established by using the GPR to obtain a quantitative description of the uncertainty of the vehicle motion:

[0024]

[0025] Wherein: A is an amplification coefficient, is the position of the ego vehicle of the i-th traffic participant, i is the position of the i-th surrounding traffic participant, and σ is a scale constant. j The relative speed of the i-th traffic participant and the surrounding traffic participants at the position is used to describe the interaction between the vehicles, and therefore the calculation method of the elements in the adjacency matrix is as follows:

[0026] i j

[0027]

[0028]

[0029] Wherein:​​​​​​​​​​​​ The relative speed of the first traffic participant and the surrounding traffic participants i at the relative position, j , , , The speed of the first traffic participant. i, j

[0030] Further, the supplementary adjacency matrix comprises an edge feature matrix and an edge type matrix ;

[0031]

[0032]

[0033] wherein: and represent the edge feature and the edge type respectively from node j to node i ;

[0034] The edge feature comprises the relative speed, the relative position and the included angle between the two orientations of the nodes, i.e. ; the relative speed can be represented as , the relative position is represented by the Euclidean distance , and the orientation angle can be represented as ; , is the position of node i, , is the speed of node i in the x and y directions;

[0035] The edge type comprises: motor vehicle-motor vehicle, pedestrian-pedestrian, non-motor vehicle-non-motor vehicle, motor vehicle-pedestrian, pedestrian-motor vehicle, motor vehicle-non-motor vehicle, non-motor vehicle-motor vehicle, pedestrian-non-motor vehicle and non-motor vehicle-pedestrian.

[0036] Further, the step (3) is specifically:

[0037] (3.1) constructing a node conversion matrix , wherein is an edge feature conversion matrix, is an edge type conversion matrix, the feature dimension of the edge feature is , the feature dimension of the edge type is , the feature dimension of the target is , the conversion matrix dimension of the edge feature is , and the conversion matrix dimension of the edge type is​ ;

[0038] (3.2) Perform spatial transformation on the edge features and edge types:

[0039]

[0040]

[0041] in: and These represent the edge feature matrix and the edge type matrix, respectively.

[0042] (3.3) Concatenate the edge features and edge types after spatial transformation to form the edge feature matrix:

[0043]

[0044] in: These are the edge features after spatial transformation. The edge type is after spatial transformation, and || represents the concatenation operation;

[0045] (3.4) The feature matrix of the edge is fused with the feature matrix of the node to obtain a node feature matrix containing the edge features;

[0046]

[0047] in: Let j represent the feature vector of the traffic participant j.

[0048] (3.5) Obtain the node feature matrix containing edge features using the shared attention mechanism of a single-layer neural network. Attention coefficient obtained:

[0049]

[0050] in: To pay attention to the coefficients, It is the transpose of the parameterized attention mechanism. Indicates traffic participants i eigenvectors, This represents the feature matrix containing edge features, and || represents the concatenation operation;

[0051] (3.6) Update traffic participants based on attention coefficient j From the feature vectors, we obtain the updated node feature matrix;

[0052]

[0053]

[0054] wherein: σ is a sigmoid function, is an edge feature after spatial transformation, represents a feature vector of a traffic participant j , || represents a splicing operation, W is a linear transformation matrix.

[0055] Further, the step (4) is specifically:

[0056] (4.1) input the updated node feature matrix and the adjacency matrix into a graph attention neural network to obtain a topological feature of the graph;

[0057] (4.2) input the topological feature of the graph into a decision algorithm module to select an action with the maximum value in an action space as an optimal action.

[0058] Further, the information of the action space is a combination of driving behaviors, as follows:

[0059]

[0060] wherein: CL, GS and CR respectively represent lane changing to the left lane, straight driving and lane changing to the right lane, and these actions are semantic level commands for guiding the motion of the ego vehicle.

[0061] Further, the decision model is realized by introducing a duel neural network on the basis of DQN.

[0062] The present application has the following beneficial effects:

[0063] The present application uses GRU to extract the future intention and behavior of the traffic participant as a dynamic feature input into a node feature vector to construct a node feature matrix. The edge feature and the edge type are projected into the same feature space, spliced and fused with the node feature for information enhancement, and an attention mechanism is introduced to score each node, so as to assign different importance to different nodes, i.e. to assign different attention to different traffic participants, to obtain an enhanced node feature matrix, so as to better model and understand the interaction relationship between different types of traffic participants, solve the problem that GAT cannot process edge features in an interactive graph, help to understand the vehicle behavior and improve the prediction ability. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 Fig. 1 is a flowchart of an autonomous vehicle behavior decision-making method for a human-vehicle mixed road environment according to an embodiment of the present application.

[0065] Figure 2 Fig. 2 is a flowchart of a construction process of a traffic scene interaction relationship graph representation according to an embodiment of the present application.

[0066] Figure 3 A traffic participant interaction relationship type schematic diagram for an embodiment of the present application.

[0067] Figure 4 A gated recurrent unit structure schematic diagram for an embodiment of the present application.

[0068] Figure 5 An attention mechanism with edge features added schematic diagram for an embodiment of the present application.

[0069] Figure 6 Three average reward value graphs based on DQN algorithm and rule-based algorithm for an embodiment of the present application.

[0070] Figure 7 A behavior decision algorithm reward value box plot for an embodiment of the present application. DETAILED DESCRIPTION

[0071] The present application will be further described below in conjunction with the drawings and embodiments.

[0072] As Figure 1 shown, an autonomous vehicle behavior decision method for a human-vehicle mixed road environment is provided, and a heterogeneous edge enhanced graph attention neural network is designed to aggregate information from the neighborhood to update node features, so as to make behavior decision of the autonomous vehicle. The method includes the following steps:

[0073] S101, historical data of different types of traffic participants is obtained, including motor vehicles, non-motor vehicles and pedestrians; the historical data of the participants includes the position, speed and heading angle of the traffic participants in the past t time.

[0074] In the actual traffic environment, there are various types of traffic participants including motor vehicles, non-motor vehicles and pedestrians, and human drivers can comprehensively consider the different characteristics of different types of traffic participants to make safe and reasonable driving behavior decisions. In order to realize similar or even superior decision-making ability to humans, the autonomous vehicle must consider all types of traffic participants in the traffic environment, and cannot only consider a single type.

[0075] Therefore, the historical data of different types of traffic participants needs to be obtained, including the position, speed and heading angle (driving direction of the vehicle, orientation of the pedestrian) of the traffic participants in the past t time.

[0076] S102, a graph representation of multi-agent interaction environment is established, including a node feature matrix, an adjacency matrix and a supplementary adjacency matrix.

[0077] As Figure 2 shown, the historical data of surrounding traffic participants is encoded using GRU to extract dynamic features of the traffic participants.

[0078]

[0079] in: Represents the GRU encoder, Indicates in t Traffic participants at all times i Extracted dynamic features, Indicates in t Traffic participants at all times i Historical data, including past traffic participants t Position, velocity, and heading angle over time.

[0080] Similar to standard recurrent neural networks, GRU maintains a hidden state vector, which is updated at each time step. The hidden state captures the sequence information encountered, and the updated hidden state is calculated as a combination of the previous hidden state and the candidate hidden state, such as... Figure 4 As shown.

[0081]

[0082]

[0083] h Indicates the hidden state of the cell. r This indicates the degree of forgetting in the hidden state, with a value of 0 indicating "completely forgotten" and 1 indicating "retained". z о represents the degree of update of the hidden state, and о represents the element-wise multiplication of the matrix.

[0084] Input this dynamic feature into the node feature vector to construct the node feature matrix.

[0085]

[0086] in: For nodes i (Traffic participants) i )exist t The node feature vector at time point, which contains traffic participants. i Current speed information, location information, heading angle (vehicle direction of travel, pedestrian orientation), acceleration, surrounding map information, and traffic participants. i The dynamic characteristics.

[0087] In the process of constructing the adjacency matrix, Gaussian Process Regression (GPR) is used to construct the Gaussian velocity field. The application of the Gaussian velocity field allows the spatial and temporal interactions between the ego vehicle and other surrounding traffic participants to be modeled within a specified spatial range. Compared to the standard method of using only relative position or velocity data to complete the adjacency matrix modeling, the introduction of the Gaussian velocity field has significant advantages. GPR is a non-parametric regression method that allows the probability distribution of an unknown function to be modeled without prior assumptions about the specific form of the function. When an autonomous vehicle is driving in a complex traffic environment, this unknown function represents the vehicle's velocity field, which varies over time and space. GPR can capture the interaction effects in space and time, not just relative position or velocity data. This allows it to better adapt to changing traffic environments, such as cooperative behavior between vehicles and traffic flow, thereby enabling more accurate quantification of interactions between vehicles.

[0088] The probability distribution of the velocity field is established using GPR to obtain a quantitative description of the uncertainty of motion:

[0089]

[0090] where: A is the amplification coefficient, is the position of the ego vehicle, i is the position of the th surrounding traffic participant, and j is the scale constant, K represents the probability distribution of the velocity field.

[0091] The correlation between two agents (people and cars in the scene) decays with distance. A smaller σ means faster decay, i.e., reducing the correlation of agents at a greater distance. Agents within the ego vehicle's range will influence each other. To consider key interactive behaviors while saving computational resources, the scale constant is set to 4m in the embodiments of the present invention.

[0092] The relative velocity of the i th traffic participant and its surrounding traffic participants j at the position is used to describe the interaction between vehicles, so the calculation method of the elements in the adjacency matrix is as follows:

[0093]

[0094]

[0095] where: is the relative velocity of the i th traffic participant and its surrounding traffic participantsj In the relative speed of the position, , the speed of the first i, j traffic participant.

[0096] The relationship between traffic participants is represented by a supplementary adjacency matrix, which contains an edge feature matrix and an edge type matrix .

[0097]

[0098]

[0099] wherein: and represent the edge feature and the edge type from node j to node i , respectively.

[0100] The edge feature includes the relative speed, the relative position, and the included angle between the two orientations, i.e. ; the relative speed can be represented as , the relative position is represented by the Euclidean distance , and the orientation angle can be represented as ; , is the position of node i , , is the speed of node i in the x and y directions.

[0101] The interactions between different types of agents have different types. In order to model and analyze such interaction types, the previous graph only has one type of edge, which cannot meet the design needs. The type of edge also needs to be reflected in the graph. Different types of edges are shown in Figure 3 , each arrowed line represents one edge type, and there are a total of 9 types: motor vehicle-motor vehicle, pedestrian-pedestrian, non-motor vehicle-non-motor vehicle, motor vehicle-pedestrian, pedestrian-motor vehicle, motor vehicle-non-motor vehicle, non-motor vehicle-motor vehicle, pedestrian-non-motor vehicle, and non-motor vehicle-pedestrian.

[0102] S103, using the supplementary adjacency matrix to extract the interaction features of the traffic scene, and fusing with the node feature matrix to obtain an updated node feature matrix.

[0103] The edge features in the supplementary adjacency matrix include relative measurements such as relative distances and relative speeds between traffic participants, and the edge type is a discrete quantity composed of one-hot encoding or other identification methods. The spatial transformation of the edge information and the edge type is performed, and the node feature matrix is fused with the node feature matrix to obtain an updated node feature matrix corresponding to the relative measurements in the formula, specifically as follows: Figure 2

[0104] (1) Construct a node transformation matrix wherein is an edge feature transformation matrix, is an edge type transformation matrix, the feature dimension of the edge feature is , the feature dimension of the edge type is , the feature dimension of the target is , the transformation matrix dimension of the edge feature is , and the transformation matrix dimension of the edge type is .

[0105] (2) Perform spatial transformation on the edge features and the edge type:

[0106]

[0107]

[0108] (3) Concatenate the spatially transformed edge features and the edge type to form an edge feature matrix:

[0109]

[0110] wherein is the spatially transformed edge feature, is the spatially transformed edge type, and || represents a concatenation operation.

[0111] (4) Fuse the edge feature matrix with the node feature matrix to obtain a node feature matrix containing edge features:

[0112]

[0113] wherein represents a feature vector of the traffic participant j.

[0114] (5) Process the node feature matrix containing edge features using a shared attention mechanism of a single-layer neural network to obtain an attention coefficient as shown in the formula. Figure 5

[0115]

[0116] wherein ​​is a coefficient of attention, is a transpose of the parameterized attention mechanism, represents a feature vector of a traffic participant i, represents a node feature matrix containing edge features, and || represents a splicing operation.

[0117] (6) updating the traffic participant j feature vector according to the coefficient of attention to obtain an updated node feature matrix.

[0118]

[0119] wherein σ is a sigmoid function, is an edge feature after spatial transformation, represents a feature vector of a traffic participant j, and || represents a splicing operation, W is a linear transformation matrix.

[0120] S104, inputting the adjacency matrix and the updated node feature matrix into a graph attention neural network to extract interaction features, and completing behavior decision-making through a decision algorithm model.

[0121] (1) inputting the updated node feature matrix and the adjacency matrix into the graph attention neural network to obtain topological features of the graph.

[0122] (2) inputting the topological features of the graph into the decision model to select an action with the maximum value in an action space as an optimal action.

[0123] In order to better focus on the state with high training value and the action corresponding thereto, the decision model of the embodiment of the application is realized by introducing a duel neural network on the basis of DQN (Dueling DQN).

[0124] The information of the action space is a combination of driving behaviors, as follows:

[0125]

[0126] wherein CL, GS and CR respectively represent lane changing to a left lane, straight driving and lane changing to a right lane, and these actions are semantic-level commands for guiding the motion of the ego vehicle.

[0127] The reward function is composed of rewards and penalties: average speed and intention rewards, lane changing and collision penalties. The reward function is expressed as:

[0128]

[0129] wherein: represents the weight of each item in the reward function.

[0130] The speed reward is represented as encouraging the unmanned vehicle to drive as close to the maximum speed limit as possible to improve the efficiency of the entire traffic system, and the calculation formula is:

[0131]

[0132] Wherein: and respectively represent the speed of the unmanned vehicle and the speed limit on the current road.

[0133] The term represents that the algorithm expects the unmanned vehicle to complete the driving task with the minimum number of lane changes, that is, the lane change penalty term:

[0134]

[0135] Wherein: is the number of lane changes of the autonomous vehicle at the current time step.

[0136] The collision penalty is to ensure the safety of the system:

[0137]

[0138] Wherein: is the number of collisions at the current time step. In order to ensure strict punishment for collision behavior, the is set to so as to ensure that the reward value is negative after the collision occurs.

[0139] In order to further verify the advantages of the embodiment of the present application, the following are compared and analyzed using the rule-based method (Rule-Based), the graph convolutional neural network combined with the DQN method (GCN-DQN), the EGAT fusion DQN method and the EGAT-DuelingDQN method (the embodiment of the present application). Among them, GCN represents the graph convolutional neural network, and EGAT represents the graph attention neural network using the enhanced feature matrix in the embodiment of the present application.

[0140] The training parameter details are shown in Table 1.

[0141] Table 1

[0142]

[0143] After all the training work is completed, Figure 6The training process of each driving behavior decision algorithm is shown, with the x-axis representing the number of training rounds and the y-axis representing the average reward value obtained in each round during training. The results show that EGAT-DQN and EGAT-DuelingDQN both converge to a similar performance level, with reward values converging to around 3300. GCN-DQN also achieves convergence, but the average reward is slightly lower than the first two algorithms. In contrast, the rule-based method's average reward curve fluctuates constantly and does not have a clear convergence point, and it is still fluctuating until the end of training.

[0144] The results show that introducing GCN and EGAT into the decision-making process can effectively consider the interactions between various types of road traffic participants, enabling the decision algorithm to converge and have a high reward value. In addition, compared with GCN-DQN, EGAT-DQN achieves a higher reward. Based on the above comparative analysis, the decision-making method that integrates EGAT can more comprehensively consider the interactions between traffic participants and make more intelligent decisions.

[0145] To compare the robustness of the rule-based method with other GCN-DQN-based methods. Figure 7 The box plot of the rewards of the four algorithms is shown. The upper horizontal line (A) represents the maximum value, and the lower horizontal line (E) represents the minimum value. The line in the middle of the box (C) is the median, the line above the box (B) is the 75th percentile, and the line below the box (D) is the 25th percentile. The median of the rule-based method is lower than that of the other three algorithms, indicating that the rewards of the rule-based method are mostly low. In addition, the median of the GCN-based method is also lower than that of the two EGAT-based algorithms.

[0146] The range of the box represents the interquartile range (IQR), which is the range of the middle 50% of the data. The smaller the IQR, the smaller the variability of the data, which can be used as an indicator of robustness. In this example, the interquartile range of the three reinforcement learning-based algorithms is smaller than that of the rule-based method, indicating that the rewards of the GCN-DQN-based methods are more concentrated around the median and have stronger robustness. In addition, the same EGAT is used to process the interaction graph, and the interaction features obtained are used as the input of the reinforcement learning module. The median of DuelingDQN is higher than that of DQN, and the optimized DuelingDQN method in this paper shows better decision-making ability.

[0147] Table 2 shows the average test rewards in the test phase. Replacing GCN with EGAT can increase the reward by about 6%, indicating that EGAT can more effectively extract and understand the interaction between vehicles than GCN. The reason why GCN cannot achieve the effect of EGAT may be that it lacks the processing of edge information. In addition, the results show that DuelingDQN performs best in testing. It can be concluded that the combination of EGAT and DuelingDQN has good effect in interactive scenarios.

[0148] Table 2

[0149]

[0150] After comparing various algorithms, it is proved that the decision effect of the EGAT-DuelingDQN method is the best, and the average reward value is 6% higher than that of the GCN-DQN method. The decision method can realize more safe and reasonable behavior decision in the interactive scene.

[0151] The specific embodiments of the application are described above with reference to the drawings. It should be understood that the application is not limited to the specific implementation described above, and those skilled in the art can make various modifications or changes within the scope of the claims, which does not affect the essence of the application.

Claims

1. A method for autonomous vehicle behavior decision making for mixed human-vehicle road environment, characterized in that, The method comprises the following steps: (1) Obtain historical data of different types of traffic participants, including motor vehicles, non-motor vehicles, and pedestrians; the historical data of the participants includes the positions, speeds, and heading angles of the traffic participants in the past t time (2) establishing a graph representation of the multi-agent interaction environment, including a node feature matrix, an adjacency matrix and a supplementary adjacency matrix; The historical data of the different types of traffic participants are encoded by using a gated recurrent unit (GRU) to obtain dynamic features, and the dynamic features are added to the node feature vector to construct a node feature matrix; The adjacency matrix is constructed by using a Gaussian process regression (GPR) to represent a Gaussian velocity field; The supplementary adjacency matrix is constructed by using an edge feature matrix and an edge type matrix of the relationships between the different types of traffic participants; (3) traffic scene interaction features are extracted by using the supplementary adjacency matrix, and are fused with the node feature matrix to obtain an updated node feature matrix; (4) the adjacency matrix and the updated node feature matrix are input into a graph attention neural network to extract interaction features, and a decision model is used to complete behavior decision-making.

2. The method for behavior decision of an autonomous vehicle in a mixed traffic environment according to claim 1, wherein, The node feature matrix is specifically as follows: The historical data of the traffic participants are encoded by using a GRU to extract dynamic features of the traffic participants: in: Represents the gated recurrent unit (GRU). Indicates in t Traffic participants at all times i The dynamic characteristics, Indicates in t Traffic participants at all times i Historical data; The dynamic features are added to the node feature vector to construct a node feature matrix; wherein: is a node i At t a time instant, the vector contains traffic participants i current speed information, position information, heading angle, acceleration, surrounding map information, and dynamic characteristics of traffic participants i ​ 3. The method for behavior decision of an autonomous vehicle in a mixed traffic environment according to claim 1, wherein, The adjacency matrix is specifically as follows: A probability distribution of a vehicle velocity field is established by using a GPR to obtain a quantitative description of the uncertainty of vehicle motion: in: A This is the magnification factor. For the first i The location of each traffic participant from their vehicle For the first j The positions of the surrounding traffic participants, where σ is a scale constant; The first i j The relative speed of the traffic participants at their surrounding traffic participants is used to describe the interaction between vehicles, so the calculation method of the elements in the adjacency matrix is as follows:​​ wherein: is the relative speed of the i-th traffic participant with respect to the traffic participant i-1, i is the relative speed of the i-th traffic participant with respect to the traffic participant i+1, j is the relative speed of the i-th traffic participant with respect to the traffic participant i-1, is the relative speed of the i-th traffic participant with respect to the traffic participant i+1, , is the relative speed of the i-th traffic participant with respect to the traffic participant i-1, i, j is the relative 4. The method for behavior decision of a driverless vehicle in a mixed human- vehicle road environment according to claim 1, wherein: The supplementary adjacency matrix contains an edge feature matrix and an edge type matrix ; wherein: and respectively represent edge features and edge types of edges pointing from node j to node i ; The edge features include the relative velocity between nodes, the relative position, and the included angle between the two heading directions, i.e. ; the relative velocity can be expressed as , the relative position is expressed by the Euclidean distance , and the heading angle can be expressed as ; 、 is the position of node i, 、 is the velocity of node i in the x and y directions; The edge types include motor vehicle-motor vehicle, pedestrian-pedestrian, non-motor vehicle-non-motor vehicle, motor vehicle-pedestrian, pedestrian-motor vehicle, motor vehicle-non-motor vehicle, non-motor vehicle-motor vehicle, pedestrian-non-motor vehicle and non-motor vehicle-pedestrian.

5. The method for behavior decision of a driverless vehicle in a mixed human- vehicle road environment according to claim 1, wherein, The step (3) is specifically as follows: (3.1) Constructing node conversion matrix wherein is an edge feature transformation matrix, is an edge type transformation matrix, the feature dimension of the edge feature is , the feature dimension of the edge type is , the feature dimension of the target is , the conversion matrix dimension of the edge feature is , the conversion matrix dimension of the edge type is ; (3.2) the edge features and the edge types are subjected to spatial transformation: wherein: and respectively denote edge feature and edge type matrices; (3.3) the spatially transformed edge features and edge types are spliced together to form an edge feature matrix: wherein: is a spatially transformed edge feature, is a spatially transformed edge type, and || denotes a concatenation operation; (3.4) the edge feature matrix is fused with the node feature matrix to obtain a node feature matrix containing edge features; wherein: represents a feature vector of the traffic participant j ; (3.5) obtaining a node feature matrix containing edge features by using a shared attention mechanism of a single-layer neural network obtaining attention coefficients: in: To pay attention to the coefficients, It is the transpose of the parameterized attention mechanism. Indicates traffic participants i eigenvectors, This represents the feature matrix containing edge features, and || represents the concatenation operation; (3.6) updating the traffic participant according to the attention coefficient j a feature vector, to obtain an updated node feature matrix; where: σ is a sigmoid function, is a spatially transformed edge feature, denotes the feature vector of traffic actor j, || denotes the concatenation operation, W is a linear transformation matrix.

6. The method for behavior decision of a driverless vehicle in a mixed human-vehicle road environment according to claim 1, wherein, The step (4) is specifically as follows: (4.1) the updated node feature matrix and the adjacency matrix are input into a graph attention neural network to obtain a graph topology feature; (4.2) the graph topology feature is input into a decision algorithm module to select an action with the maximum value in an action space as an optimal action.

7. The method for behavior decision of an autonomous vehicle in a mixed traffic environment according to claim 6, wherein, The information of the action space is a combination of driving behaviors, and is as follows: Wherein, CL, GS and CR represent lane changing to the left lane, straight driving and lane changing to the right lane, respectively. These actions are semantic-level commands for guiding the motion of the ego vehicle.

8. The method for behavior decision of a driverless vehicle in a mixed human-vehicle road environment according to claim 1, wherein: The decision model is realized by introducing a duel neural network on the basis of a deep Q network (DQN).

Images