Lane line guided graph search heterogeneous interaction vehicle trajectory prediction method

Abstract

The invention relates to the technical field of automatic driving track prediction, in particular to a lane line guided graph search heterogeneous interaction vehicle track prediction method. The invention discloses a lane line guided graph search heterogeneous interaction vehicle trajectory prediction method. The method comprises the steps of graph model-oriented trajectory data preprocessing, human-vehicle-graph joint state coding, target lane guided path search strategy, node aggregation-based trajectory decoding prediction and multi-target loss function design. The method achieves the heterogeneous prediction of the interactive vehicle track through the map search guided by lane lines, achieves the building of a map attention convergence model through a vehicle-vehicle interaction module and a human-vehicle interaction module, effectively captures the dynamic interaction relation between different types of traffic participants, and improves the track prediction precision of a complex urban traffic scene.

Description

Technical Field

[0001] This application relates to the technical field of autonomous driving trajectory prediction, and particularly to a graph search heterogeneous interaction vehicle trajectory prediction method guided by lane lines. Background Art

[0002] Under the layout of the new generation of information technology, the vehicle intelligent system is gradually changing people's travel modes, and it is the strategic direction of the global automotive industry and a must-win area for the new round of scientific and technological industrial revolution. Intelligent vehicle trajectory prediction is a core and key technology emerging in the field of vehicle intelligence in recent years, which can reserve sufficient reaction time for vehicle decision-making and improve driving safety and traffic efficiency. In complex urban traffic scenarios, the dynamic interaction between heterogeneous traffic participants intensifies, the trajectories are complex and numerous, and traffic rules are difficult to describe, which poses a huge challenge to achieving accurate vehicle trajectory prediction.

[0003] First, the interaction between vehicles and surrounding traffic participants is the dominant factor leading to changes in vehicle trajectories. Most existing methods mainly consider vehicle-to-vehicle interaction and ignore the interaction between vehicles and pedestrians, the vulnerable group. Second, the current trajectory prediction methods have insufficient utilization of traffic rules and only use traffic rules as hard constraints for system boundaries, which is likely to produce unreasonable trajectory prediction results. Summary of the Invention

[0004] This application provides a graph search heterogeneous interaction vehicle trajectory prediction method guided by lane lines to solve the defect of incomplete consideration of interaction factors in the prior art, realize the construction of a graph attention aggregation model for vehicle-to-vehicle interaction module and vehicle-to-pedestrian interaction module, effectively capture the dynamic interaction relationship between heterogeneous traffic participants, and improve the trajectory prediction accuracy in complex urban traffic scenarios.

[0005] This application provides a graph search heterogeneous interaction vehicle trajectory prediction method guided by lane lines, including: Trajectory data preprocessing for the graph model: Combining the guiding and constraining effects of lane line information on vehicle driving trajectories, establish a directed graph model for vehicle trajectory prediction; Combining the motion states of the target vehicle and traffic participants, construct structured trajectory prediction input data; Joint state encoding of person-vehicle-graph: Combining the interaction between pedestrian-vehicle and vehicle-vehicle and traffic rule constraints, in the encoding stage, input different traffic participants and map information into the corresponding gated recurrent unit; Path search strategy guided by the target lane: The path search strategy first determines the possible lane segment positions that the vehicle may reach at the final time step within the prediction time domain; There are multiple feasible paths between the same starting points. Using the target node as the guide, combined with the map nodes near the current position of the target vehicle, perform forward rolling search within the prediction time domain to extract multiple drivable paths.

[0006] Trajectory decoding prediction based on node aggregation; Based on each drivable path and the different information provided by each step of node search, a trajectory decoder based on node aggregation is designed to make the predicted trajectory smooth.

[0007] Design of multi-objective loss function: The loss function is used for the training of the trajectory prediction model and is divided into two stages. Its fundamental purpose is to make the predicted trajectory track the true driving trajectory of the vehicle; In the first stage, it is divided into two sub-objectives. First, the target node is consistent with the end position of the vehicle's true driving. Second, the geometric error between the predicted trajectory and the vehicle's true driving trajectory is minimized; The second stage of training mainly focuses on whether the search path complies with traffic rules and the description of the interaction relationship between different types of traffic participants to associate the search path with the lane line segment.

[0008] According to a lane-line-guided graph search heterogeneous interactive vehicle trajectory prediction method provided by the present application, in the step of preprocessing trajectory data for the graph model, the traffic participants include the target vehicle, participating vehicles, pedestrians, and the map.

[0009] According to a lane-line-guided graph search heterogeneous interactive vehicle trajectory prediction method provided by the present application, the historical motion state of the target vehicle is expressed as:

[0010] where is the historical motion state of the target vehicle, is the current moment; represents the backtracking time, that is, the total time step of the historical time domain, represents the state of the target vehicle at time corresponding to the longitudinal position of the target vehicle at time corresponding to the lateral position of the target vehicle at time corresponding to the speed of the target vehicle at time corresponding to the acceleration of the target vehicle at time corresponding to the yaw angle of the target vehicle at time, and the time can be expressed as ; Or, the historical motion states of the participating vehicles and the pedestrians are both expressed as:

[0011] where represents the to The historical state of a traffic participant, whose type is represented as , is the vehicle motion state, is the pedestrian motion state. The historical states of the traffic participating vehicles and pedestrians are consistent with those of the target vehicle, denoted as ; Or, the map is structured in the form of a graph model , where represents the set of nodes in the graph model, is the number of nodes, and each node corresponds to a section of the lane centerline:

[0012] where , and are respectively the starting position, ending position, and yaw angle of the lane centerline, corresponding to the attributes of the map nodes, including the presence of traffic signs, traffic lights, and ground marking restrictions.

[0013] According to a lane - line - guided graph - search heterogeneous interaction vehicle trajectory prediction method provided by the present application, the elements of the graph model include edges. The edges describe the connectivity and direction relationship between nodes. According to the information on whether there are connectable edges between nodes, an adjacency matrix is constructed as:

[0014] where represents the connectivity between the th and the th lane - segment nodes, represented by a boolean value, where represents node connectivity, represents node non - connectivity.

[0015] According to a lane - line - guided graph - search heterogeneous interaction vehicle trajectory prediction method provided by the present application, at the output level, since there are multiple feasible paths for the vehicle to reach the target position, the output is a multi - modal prediction of the vehicle trajectory:

[0016] where, is the th predicted trajectory, is the time step at which the specific coordinates of the predicted trajectory are located, is the prediction time domain.

[0017] A method for predicting the trajectory of a graph search heterogeneous interactive vehicle guided by lane lines provided by the present application. In the step of jointly encoding the states of the human-vehicle-graph, in the gated recurrent unit:

[0018] Where represents the motion encoding features of different types of traffic participants. The gated recurrent unit includes a reset gate and an update gate; The reset gate determines how to combine new input information with the historical state. The calculation process is expressed as:

[0019] Where is the weight matrix, is the input at the previous moment, is the hidden state at the previous moment, represents sigmoid, which is used to achieve non-linear transformation between states; The reset gate further calculates the hidden state :

[0020] Among them, is the weight matrix, represents element-wise multiplication, is the hyperbolic tangent function; The smaller the value, the corresponding The smaller the value, the more information from the previous moment is forgotten; conversely The larger the value; The update gate is used to control the degree of combination of the state at the previous moment and the state at the current moment. Its expression is:

[0021] Where is the weight matrix, which is used to perform a linear transformation on the matrix concatenated by and and calculate the value of the update gate to obtain the output of the entire GRU:

[0022] Among them, represents selective forgetting of the hidden state at the previous moment, represents further selective memory of the candidate hidden state; Based on this, traffic participants and map information at different times are encoded into state encodings containing historical information; According to the status encoding, the reasons for the change in the vehicle driving trajectory can be described as interactions; the motion state of the vehicle is aggregated to the lane segment nodes of the map through status propagation to describe traffic flow characteristics:

[0023] Among them is the traffic flow characteristics combining surrounding vehicles and map nodes, represents the splicing between matrices, 、 and respectively represent the query matrix, key matrix, and value matrix of cross-attention; specifically, is obtained by performing a linear transformation on ; is obtained; and are obtained through the linear transformation of the motion states of surrounding vehicles and pedestrians ; and are obtained; is a data normalization function; Further, through the propagation and aggregation of the status between lane segment nodes, vehicle-vehicle and person-vehicle interactions are described:

[0024] Among them represents the node features after the update of the lane line segment, 、 and represent another set of parameter matrices for cross-attention, and the adjacency matrix participates in the calculation process, enabling the lane line features to propagate between connected map nodes.

[0025] According to a lane line-guided graph search heterogeneous interaction vehicle trajectory prediction method provided by the present application, in the steps of the path search strategy guided by the target lane, specifically including: The path search strategy first determines the possible lane segment positions that the vehicle may reach at the final time step within the prediction time domain, which are called target nodes:

[0026] Among them represents the possibility that the predicted end point of the target vehicle trajectory falls on different map nodes, is a non-linear activation function, is a fully connected layer network; by comparing the element values within , the target node index with the highest probability can be determined, so as to retrieve its encoded features:

[0027] Among them denotes the maximum value element in denotes the index of the target node, denotes the retrieval feature of the target node; Considering that the vehicle trajectory is multimodal, that is, there are multiple feasible paths between the same starting points, using the target node as a guide, combined with the map nodes near the current position of the target vehicle, a forward rolling search within the prediction time domain is performed to extract multiple drivable paths, each path containing sampling points; Taking a single-step search as an example, the process of the target vehicle searching from node to the next node is expressed as:

[0028] Among them , is a One-Hot vector, denoting the connection relationship between node and ; denotes the attention weight of node to node in the connected node set , expressed in the form of probability; when traversing all nodes in, the corresponding attention weights are stored in ; guided by the target node, the index is located according to the highest probability principle, so as to retrieve the encoded feature of the subsequent node; When , node searches for the longitudinal successor node , indicating that the vehicle is driving in a straight line; when , node searches for the lateral successor node , indicating that the vehicle is turning or changing lanes; it should be noted that in order to represent behaviors such as vehicle deceleration or stopping, a search strategy is designed to describe the self-loop mode, that is, the next node searched is still the current node.

[0029] According to a lane-line-guided graph search heterogeneous interaction vehicle trajectory prediction method provided by the present application, in the step of trajectory decoding prediction based on node aggregation, it specifically includes: The path search strategy provides lane segment node features for the potential driving paths of the target vehicle, and its feature set is expressed as ; For a path, the information provided by each step of node search is different. Based on the node-aggregated trajectory decoder, the predicted trajectory is made smoother:

[0030] where represents all the path features after aggregation, 、 and are self-attention parameter matrices, which are linear transformations of the target vehicle's motion state and path features respectively; Through this operation, the path obtained by map node search and the vehicle motion information are further combined to form the basis of the fully connected layer decoder. The decoding process is as follows:

[0031] where is the th predicted trajectory of the target vehicle.

[0032] According to a lane-line-guided graph search heterogeneous interaction vehicle trajectory prediction method provided by the present application, in the steps of designing the multi-objective loss function, it specifically includes: The target node is consistent with the end position of the vehicle's actual driving, denoted as ; The geometric error between the predicted trajectory and the vehicle's actual driving trajectory is as small as possible, denoted as ; The first-stage training can be expressed as:

[0033] where represents the position of the center line of the lane of the target node, is the end point of the vehicle's actual driving trajectory, represents the Euclidean distance norm; The second-stage training mainly focuses on whether the search path complies with traffic rules and the description of the interaction relationship between different types of traffic participants, which can be expressed as:

[0034] where represents the possibility of the vehicle continuously accessing adjacent map nodes in the form of negative logarithmic probability, enhancing the correlation between the search path and the lane segment.

[0035] The lane-line-guided graph search heterogeneous interaction vehicle trajectory prediction method provided by the present application realizes the construction of a graph attention aggregation model for vehicle-vehicle interaction module and vehicle-human interaction module by graph search heterogeneous prediction of interaction vehicle trajectories, effectively captures the dynamic interaction relationship between heterogeneous traffic participants, and improves the trajectory prediction accuracy in complex urban traffic scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0037] Figure 1 It is a flowchart of a lane-line-guided graph search heterogeneous interaction vehicle trajectory prediction method according to an embodiment of the present application; Figure 2 It is a schematic diagram of lane-line node search according to an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] To make the objectives, technical solutions and advantages of the present application clearer, the following will clearly and completely describe the technical solutions in the present application with reference to the drawings in the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present application belong to the scope of protection of the present application.

[0039] In the related art, the interaction between a vehicle and surrounding traffic participants is the dominant factor causing changes in the vehicle trajectory. Most existing methods mainly consider the interaction between vehicles and ignore the interaction between the vehicle and pedestrians, a vulnerable group. To address this issue, the present patent application fully considers the trajectory characteristics of heterogeneous traffic participants, designs a corresponding coding structure, and describes the dynamic interaction between them to improve the prediction accuracy. In addition, the current trajectory prediction methods have insufficient utilization of traffic rules and only use traffic rules as a hard constraint for the system boundary, which is likely to produce unreasonable trajectory prediction results. To address this issue, the present patent application establishes traffic rules as a directed graph model, gradually searches for feasible trajectories along the center line of the lane, and improves the traffic rule compliance rate of the prediction results while giving full play to the guiding and constraining role of traffic rules.

[0040] In a dynamic and complex urban traffic environment, the intertwined lane-line information is difficult to be effectively utilized by the target vehicle. The center line of the lane in the high-definition (HD) map is discretized, the end position of the vehicle within the prediction time domain is constrained to the center line segment of the lane, and the prediction process is divided into a search process guided by the map information around the vehicle. Based on the graph search method, it is ensured that the prediction results comply with the guidance and constraints of traffic rules, and a trajectory smoothing processing mechanism is introduced to make the prediction results tend to the trend of the vehicle's true driving trajectory.

[0041] The following refers to the attachedFigure 1 - Figure 2 Describe in detail the lane-line-guided graph search heterogeneous interaction vehicle trajectory prediction method according to the embodiments of the present application. It should be understood that the following description is only an exemplary illustration and not a specific limitation of the present application.

[0042] As Figure 1 and Figure 2 shown, the lane-line-guided graph search heterogeneous interaction vehicle trajectory prediction method according to the embodiments of the present application specifically includes the following steps: Step 1: Trajectory data preprocessing for the graph model: The input information for vehicle trajectory prediction involves the historical trajectory of the vehicle, the interaction between vehicles, and the high-precision map lane-line information to infer the driving trajectory of the vehicle within a certain prediction time domain; considering the guiding and constraining effects of lane-line information on the vehicle driving trajectory, integrating the dynamic heterogeneous interactions between vehicle-vehicle and person-vehicle, establishing a directed graph model for vehicle trajectory prediction, and considering the motion states of the target vehicle and traffic participants to construct structured trajectory prediction input data; For the vehicle to be predicted, which is called the target vehicle, its historical motion state is expressed as:

[0043] In the formula is the historical motion state of the target vehicle, is the current moment; represents the backtracking time, that is, the total time step of the historical time domain, represents the state of the target vehicle at the moment, where corresponds to the longitudinal position of the target vehicle at the moment, corresponds to the lateral position of the target vehicle at the moment, corresponds to the speed of the target vehicle at the moment, corresponds to the acceleration of the target vehicle at the moment, corresponds to the yaw angle of the target vehicle at the moment, and the moment can be expressed as ; As an important object participating in vehicle dynamic interaction, the surrounding neighboring other traffic participating vehicles and pedestrians of the target vehicle have a significant impact on the trajectory change, and their historical motion states are expressed as:

[0044] In the formula represents the historical state of the th to th traffic participants, and its type is expressed as , is the vehicle motion state, is the pedestrian motion state. The historical states of traffic participating vehicles and pedestrians are the same as those of the target vehicle, denoted as ; In addition to the motion states of agents such as vehicles and pedestrians, high-definition (HD) map-derived advanced semantic information is used as an important input to the prediction model. To combine traffic constraints and rules in the HD map with motion states, a map vectorization method based on uniform spatial sampling is designed; High-precision map is structured in the form of a graph model , where represents the set of nodes in the graph model, is the number of nodes. Each node corresponds to a section of the lane centerline:

[0045] where , and are the starting position, ending position, and yaw angle of the lane centerline respectively, corresponding to the attributes of the map nodes, including the existence of traffic signs, traffic lights, ground marking restrictions, etc.; Another key element of the graph model is the edge, which describes the connectivity and directional relationship between nodes. According to the information on whether there is a connectable edge between nodes, an adjacency matrix is constructed as:

[0046] where represents the connectivity between the -th and -th lane segment nodes, represented by a boolean value, where indicates connectivity between nodes, indicates non-connectivity between nodes; At the output level, since there are multiple feasible paths for the vehicle to reach the target position, the output is a multi-modal prediction of the vehicle trajectory:

[0047] where, is the -th predicted trajectory, is the time step at the specific coordinates of the predicted trajectory, is the prediction time domain; Step 2: Joint state encoding of human-vehicle-graph: Considering the interactions between pedestrians and vehicles, vehicles and vehicles, as well as traffic rule constraints, during the encoding stage, different traffic participants and map information are input into the corresponding Gate Recurrent Unit (GRU):

[0048] Among them represents the motion encoding features of different types of traffic participants. GRU is a typical recurrent neural network, which includes a reset gate and an update gate; the reset gate determines how to combine new input information with the historical state, and the calculation process is expressed as:

[0049] Among them is the weight matrix, is the input at the previous moment, is the hidden state at the previous moment, represents sigmoid, which is used to achieve the non-linear transformation between states; The reset gate further calculates the hidden state :

[0050] Among them, is the weight matrix, represents element-wise multiplication, is the hyperbolic tangent function; The smaller the value of is, the smaller the corresponding value is, indicating that more information from the previous moment is forgotten; on the contrary The larger the value of

[0051] Among them is the weight matrix, which is used to linearly transform the matrix concatenated by and and calculate the value of the update gate , and then obtain the output of the entire GRU:

[0052] Among them, represents selective forgetting of the hidden state at the previous moment, represents further selective memory of the candidate hidden state; based on this, traffic participants and map information at different moments are encoded into state encodings containing historical information; According to the status encoding, the reasons for the change in the vehicle driving trajectory can be described as interactions; specifically, the motion state of the vehicle is aggregated to the lane segment nodes of the map through status propagation, thereby describing the traffic flow characteristics:

[0053] Among them is the traffic flow characteristics combining the surrounding vehicles and map nodes, represents the splicing between matrices, 、 and respectively represent the query matrix, key matrix, and value matrix of cross-attention; specifically, is obtained by performing a linear transformation on ; is obtained; and are obtained through the linear transformations of the motion states of the surrounding vehicles and pedestrians and ; is the data normalization function; In addition, the vehicle-vehicle and person-vehicle interactions are further described through the propagation and aggregation of the status between the lane segment nodes:

[0054] Among them represents the node features after the update of the lane line segment, 、 and represent another set of parameter matrices for cross-attention, and the adjacency matrix participates in the calculation process, enabling the lane line features to propagate only between connected map nodes; Step 3. Path search strategy for target lane guidance: The encoded lane segment nodes not only carry traffic information but also implicitly represent the road topology and traffic rules through their connectivity relationships; therefore, the path search strategy first determines the possible lane segment positions that the vehicle may reach at the final time step within the prediction time domain, which are called target nodes:

[0055] Among them represents the possibility that the predicted end point of the target vehicle trajectory falls on different map nodes, is the non-linear activation function, is the fully connected layer network; by comparing the element values within , the target node index with the highest probability can be determined, and thus its encoded features can be retrieved:

[0056] Among them represents the maximum value element in indicating the index of the target node representing the retrieval feature of the target node; Considering that vehicle trajectories are multimodal, that is, there are multiple feasible paths between the same starting points, the target node is used as a guide, combined with the map nodes near the current position of the target vehicle, and a forward rolling search within the prediction time domain is performed to extract multiple drivable paths, each path containing sampling points; Specifically, taking a single-step search as an example, the process of the target vehicle searching from node to the next node is expressed as:

[0057] Among them , is a One-Hot vector, representing the connectivity relationship between nodes and ; represents the attention weight of node to node in the connected node set , expressed in the form of probability; when traversing all nodes in , the corresponding attention weights are stored in ; guided by the target node, the index is located according to the principle of the highest probability, so as to retrieve the encoded features of the subsequent nodes; Specifically, when , node searches for the longitudinal successor node , indicating that the vehicle is driving in a straight line; when , node searches for the lateral successor node , indicating that the vehicle is turning or changing lanes; it should be noted that in order to represent behaviors such as vehicle deceleration or stopping, a search strategy is designed to describe the self-loop mode, that is, the next node searched is still the current node; Step 4. Trajectory decoding prediction based on node aggregation: The path search strategy provides lane segment node features for the potential driving paths of the target vehicle. Considering 10 paths, the feature set is expressed as ; for a path, the information provided by each step of node search is different. Based on this, a trajectory decoder based on node aggregation is designed to make the predicted trajectory smoother:

[0058] Among them represents all the path features after aggregation , and are self-attention parameter matrices, which are linear transformations of the target vehicle's motion state and path features respectively Through this operation, the path obtained by map node search and vehicle motion information are further combined to form the basis of the fully connected layer decoder. The decoding process is as follows

[0059] Among them is the th predicted trajectory of the target vehicle Step 5. Design of multi-objective loss function The loss function is used for training the trajectory prediction model and is divided into two stages. Its fundamental purpose is to make the predicted trajectory track the vehicle's true driving trajectory. In the first stage, it is divided into two sub-goals. First, the target node is consistent with the end position of the vehicle's true driving, denoted as ; Secondly, the geometric error between the predicted trajectory and the vehicle's true driving trajectory is as small as possible, denoted as ; The training in the first stage can be expressed as

[0060] Among them represents the position of the center line of the lane of the target node is the end point of the vehicle's true driving trajectory represents the Euclidean distance norm The training in the second stage mainly focuses on whether the search path complies with traffic rules and the description of the interaction relationship between different types of traffic participants, which can be expressed as

[0061] Among them represents the possibility of the vehicle continuously accessing adjacent map nodes in the form of negative logarithmic probability, enhancing the correlation between the search path and the lane segment

[0062] The lane-line-guided graph search heterogeneous interactive vehicle trajectory prediction method according to the embodiments of the present application has the following effects 1. Based on the vehicle-vehicle interaction module and the vehicle-pedestrian interaction module, a graph attention aggregation model is constructed to effectively capture the dynamic interaction relationship between heterogeneous traffic participants and improve the trajectory prediction accuracy in complex urban traffic scenarios 2. Design the target node query mechanism, execute the forward path search strategy based on the target lane segment, and make full use of the guiding and restrictive effects of traffic rules to improve the compliance rate of traffic rules in the trajectory prediction results; 3. The designed two-stage loss function optimization strategy ensures the accuracy of the model in identifying the target nodes of vehicle driving in the initial stage of training, further enhances the consistency between the trajectory and traffic rules in the later stage of training, and takes into account both prediction accuracy and rationality.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A lane-guided graph search heterogeneous interactive vehicle trajectory prediction method, characterized in that: include: Trajectory data preprocessing for graph models: Combine the guidance and constraint of lane line information on vehicle trajectory to establish a directed graph model for vehicle trajectory prediction; Combining the motion states of target vehicles and traffic participants, constructing structured trajectory prediction input data; Joint state encoding of pedestrian-vehicle-map: combining pedestrian-vehicle, vehicle-vehicle interactions and traffic rules constraints, different traffic participants and map information are input into the corresponding gated recurrent units during the encoding stage; Path search strategy for target lane guidance: The path search strategy first determines the lane segment position that the vehicle reaches at the final time step in the prediction time domain; there are multiple feasible paths between the same starting point, and the target node is used as a guide. Combined with the map nodes near the current position of the target vehicle, a forward rolling search is performed in the prediction time domain to extract multiple drivable paths; Trajectory decoding prediction based on node aggregation; Based on each drivable path and the different information provided by each node search step, a trajectory decoder based on node aggregation is designed to make the predicted trajectory smooth; Multi-objective loss function design: The loss function is used to train the trajectory prediction model, which is divided into two stages. Its fundamental purpose is to make the predicted trajectory track the actual driving trajectory of the vehicle; The first stage is divided into two sub-goals: first, the target node is consistent with the actual end point of the vehicle, and second, the geometric error between the predicted trajectory and the actual vehicle trajectory is as small as possible; The second stage of training focuses on whether the search path complies with traffic regulations and the description of the interaction between different types of traffic participants in order to associate the search path with lane segments.

2. The method for predicting heterogeneous interactive vehicle trajectories based on graph search guided by lane lines according to claim 1, characterized in that: In the step of preprocessing trajectory data for the graphical model, the traffic participants include a target vehicle, participating vehicles, pedestrians and a map.

3. The method for predicting heterogeneous interactive vehicle trajectories based on graph search guided by lane lines according to claim 2, characterized in that: The historical motion state of the target vehicle is expressed as: In the formula is the historical motion state of the target vehicle, for the current moment; represents the lookback time, that is, the total time step in the historical time domain, represent The state of the target vehicle at the moment, where correspond The longitudinal position of the target vehicle at time instant, correspond The lateral position of the target vehicle at each moment, correspond The speed of the target vehicle at any moment, correspond The acceleration of the target vehicle at time correspond The yaw angle of the target vehicle at time , expressed as ; or, The historical motion states of the participating vehicles and the pedestrians are expressed as: In the formula Indicates arrive The historical status of a traffic participant is represented by its type: , is the vehicle motion state, is the movement state of the pedestrian, and the historical state of the traffic participating vehicles and pedestrians is consistent with the target vehicle, recorded as ; or, map Structured as a graph model of the form, in which Represents a set of nodes in a graph model. is the number of nodes, each node Corresponding to a lane centerline: in , and are the starting position, ending position and yaw angle of the lane centerline, The properties of the corresponding map nodes, including whether there are traffic signs, traffic lights, and ground marking restrictions.

4. The method for predicting heterogeneous interactive vehicle trajectories based on graph search guided by lane lines according to claim 1, characterized in that: The elements of the graph model include edges, which describe the connectivity and direction relationship between nodes. According to the information whether there are connectable edges between nodes, the adjacency matrix is ​​constructed as follows: in Indicates Article and The connectivity between lane segment nodes is represented by a Boolean value, where Indicates the connectivity between nodes. Indicates that the nodes are not connected.

5. The method for predicting heterogeneous interactive vehicle trajectories based on graph search guided by lane lines according to claim 1, characterized in that: At the output level, since there are multiple feasible paths for the vehicle to reach the target location, the output is a multimodal prediction of the vehicle trajectory: in, For the Prediction trajectories, is the time step The specific coordinates of the predicted trajectory, For the prediction time domain.

6. The method for predicting heterogeneous interactive vehicle trajectories based on graph search guided by lane lines according to claim 1, characterized in that: In the step of encoding the joint state of the person-vehicle-image, in the gated recurrent unit: in Representing motion coding features of different types of traffic participants, the gated recurrent unit comprises a reset gate and an update gate; The reset gate determines how to combine new input information with the historical state. The calculation process is expressed as: in is the weight matrix, is the input at the previous moment, is the hidden state at the last moment, Represents sigmoid, which is used to realize nonlinear transformation between states; Reset the gate to further calculate the hidden state : in, is the weight matrix, represents the element-wise product, is the hyperbolic tangent function; The smaller the value, the The smaller the value, the more information about the previous moment is forgotten; The larger the value; The update gate is used to control the degree of integration between the state at the previous moment and the state at the current moment, and its expression is: in is the weight matrix used to and The concatenated matrix is ​​linearly transformed and the update gate is calculated The value of , and then get the output of the entire GRU: in, Indicates the selective forgetting of the hidden state at the previous moment. Represents further selective memory of candidate hidden states; based on this, traffic participants and map information at different times are encoded as state encodings containing historical information; According to the state encoding, the reasons for the changes in vehicle trajectories are described as interactions; the vehicle's motion state is aggregated to the lane segment nodes of the map through state propagation to describe the traffic flow characteristics: in It is a combination of traffic flow characteristics of surrounding vehicles and map nodes. represents the concatenation between matrices, , and denote the query matrix, key matrix, and value matrix of cross attention respectively; specifically, Through Perform linear transformation get; and Based on the movement status of surrounding vehicles and pedestrians Linear transformation of and get; is the data normalization function; The vehicle-vehicle and human-vehicle interactions are further described through the propagation and aggregation of states between lane segment nodes: in represents the node features after the lane segment is updated, , and Represents another set of parameter matrices for cross attention, the adjacency matrix Participate in the calculation process so that lane line features are propagated between connected map nodes.

7. The method for predicting heterogeneous interactive vehicle trajectories guided by graph search using lane lines according to claim 1, characterized in that: The steps of the target lane guidance path search strategy specifically include: The path search strategy first determines the lane segment position that the vehicle reaches at the final time step in the prediction time domain, which is called the target node: in Indicates the probability that the predicted end point of the target vehicle trajectory falls on different map nodes, is a nonlinear activation function, is a fully connected layer network; by comparison The element value in determines the target node index with the highest probability, thereby retrieving its encoding feature: in express The maximum value element in represents the index of the target node, Represents the retrieval features of the target node; Considering that the vehicle trajectory is multimodal, that is, there are multiple feasible paths between the same starting point, the target node is used as a guide, combined with the map nodes near the current position of the target vehicle, a forward rolling search is performed in the prediction time domain to extract multiple drivable paths, each of which contains sampling points; Taking a single-step search as an example, the target vehicle starts from node Search to next node The process is expressed as: in , It is a One-Hot vector, representing a node and The connectivity relationship between them; Representation Node For the connected node set Midpoint The attention weight is expressed in the form of probability; when traversing For all nodes in Medium; guided by the target node, locate the index according to the principle of highest probability , thereby retrieving the encoding features of subsequent nodes ; when When the node Searching for vertical successor nodes , indicating that the vehicle is traveling in a straight line; when When the node Search for lateral successor nodes , indicating that the vehicle is turning or switching to driving; it should be noted that in order to represent the vehicle deceleration or parking behavior mode, the search strategy is designed Describes a self-loop mode, where the next node to search is still the current node.

8. The method for predicting heterogeneous interactive vehicle trajectories guided by graph search using lane lines according to claim 1, characterized in that: The step of trajectory decoding prediction based on node aggregation specifically includes: The path search strategy provides lane segment node features for the potential driving path of the target vehicle, and its feature set is expressed as ; For a path, the information provided by each node search step is different. The trajectory decoder based on node aggregation makes the predicted trajectory smoother: in represents all path features after aggregation, , and is the self-attention parameter matrix, which are the linear transformations of the target vehicle’s motion state and path characteristics; After this operation, the path and vehicle motion information obtained by the map node search are further combined to form the basis of the fully connected layer decoder. The decoding process is: in The target vehicle Prediction trajectories.

9. The method for predicting heterogeneous interactive vehicle trajectories using graph search guided by lane lines according to claim 1, characterized in that: The steps of designing the multi-objective loss function specifically include: The target node is consistent with the actual destination position of the vehicle, denoted as ; The geometric error between the predicted trajectory and the actual driving trajectory of the vehicle is as small as possible, denoted as ; The first stage of training is expressed as: in represents the lane centerline position of the target node, is the end point of the vehicle’s actual driving trajectory, Represents the Euclidean distance norm; The second stage of training focuses on whether the search path complies with traffic regulations and the description of the interaction relationship between different types of traffic participants, which is expressed as: in The possibility of a vehicle continuously visiting adjacent map nodes is expressed in the form of negative logarithmic probability, which enhances the correlation between the search path and lane segments.

Images