Agent traffic participant behavior generation and evolution method and device for large-scale simulation, and related equipment

Abstract

The application provides an agent traffic participant behavior generation and evolution method and device for large-scale simulation and related equipment, and belongs to the technical field of traffic simulation. The method comprises the following steps: acquiring a road network map, a route coordinate file of a traffic participant and a generation plan of the traffic participant; generating the traffic participant on the road network map according to the route coordinate file and the generation plan at a generation time; performing traffic simulation according to individualized parameters of the traffic participant and first traffic information of the traffic participant within a neighborhood range in the road network map to obtain a first traffic simulation result; and adjusting the first traffic simulation result by using traffic control information to obtain a second traffic simulation result. The method can improve the simulation effect of traffic simulation.

Description

Technical Field

[0001] This application relates to the field of traffic simulation technology, and in particular to a method, apparatus and related equipment for generating and evolving the behavior of intelligent traffic participants for large-scale simulation. Background Technology

[0002] With urban road networks becoming increasingly saturated and traffic problems becoming more frequent, refined traffic management and channelization have become crucial requirements for road network governance. Traffic simulation is an important technical means to quantitatively assess the effectiveness of various control schemes in advance and to mitigate the risks of on-site modifications. Existing traffic simulation schemes mostly rely on purely historical traffic data or fixed behavioral rules to drive modeling. The behaviors of various traffic participants and the simulation results show obvious homogeneity, and it is difficult to incorporate external controls during the simulation process, resulting in poor traffic simulation effects. Summary of the Invention

[0003] This application provides a method, apparatus, and related equipment for generating and evolving the behavior of intelligent traffic participants in large-scale simulations, which can solve the technical problem of poor traffic simulation results.

[0004] In a first aspect, embodiments of this application provide a method for generating and evolving the behavior of intelligent traffic participants in large-scale simulations, the method comprising: The system acquires a road network map, a route coordinate file of traffic participants, and a generation plan for the traffic participants. The route coordinate file is used to constrain the running path of the traffic participants on the road network map, and the generation plan is used to determine the generation time and running status of the traffic participants on the road network map at the generation time. The traffic participants are preset traffic participation objects that make traffic behavior decisions through intelligent agents. The traffic participants are generated on the road network map at the generation time based on the route coordinate file and the generation plan. Based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants in the neighborhood range of the road network map, traffic simulation is performed using a neural network (NN) model to obtain intermediate simulation results. The intermediate simulation results are then adjusted using preset constraint rules to obtain the first traffic simulation result. The personalized parameters are used to characterize the traffic characteristics of the traffic participants. The first traffic simulation result is adjusted using traffic control information to obtain a second traffic simulation result. The traffic control information includes at least one of the following: operational status control information for the traffic participants, disturbance control information, safety rule constraints, and traffic light status prediction information.

[0005] Optionally, the route coordinate file includes a route index and the route trajectory data corresponding to the route index; The generation of traffic participants on the road network map at the generation time based on the route coordinate file and the generation plan includes: The route trajectory data is determined in the route coordinate file based on the route index; The route trajectory data is used to perform proximity matching on the roads in the road network map to obtain the first lane sequence; Perform topological feasibility filtering on the first lane sequence to obtain the running path of the traffic participant on the road network map; The traffic participant is generated at the starting point of the running path at the generation time.

[0006] Optionally, the NN model is a multi-agent motion generation network model based on the attention mechanism transformer; Based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants within their neighborhood in the road network map, a neural network (NN) model is used to perform traffic simulation, obtaining intermediate simulation results. These intermediate simulation results are then adjusted using preset constraint rules to obtain the first traffic simulation result, including: The personalized parameters of the traffic participants, the types of the traffic participants, and the historical simulation results of the traffic participants are spliced ​​together to obtain the spliced ​​result; The stitching result, the running path, and the first traffic information of the traffic participants within the neighborhood of the road network map are input into a transformer-based multi-agent motion generation network model for traffic simulation to obtain intermediate simulation results. The intermediate simulation results are then adjusted according to preset rule constraints to obtain the first traffic simulation result. The preset rule constraints include at least one of the following: constraints determined by the intelligent driver model (IDM) based on reference driving behavior; traffic operation constraints; traffic signal constraints; and route rule constraints.

[0007] Optionally, the transformer-based multi-agent motion generation network model is trained in the following manner: Obtain the real trajectory dataset of the traffic participants on the road network map; A first loss and a second loss are determined based on the real trajectory dataset and the second traffic simulation results. The first loss is the regression loss of the second traffic simulation results relative to the real trajectory dataset, and the second loss is the cross-entropy loss of the second traffic simulation results relative to the real trajectory dataset. Construct a loss function based on the first loss and the second loss; The transformer-based multi-agent motion generation network model is trained according to the loss function to obtain the trained transformer-based multi-agent motion generation network model.

[0008] Optionally, the generation time is within a first time range of the simulation cycle of the method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation, and the first time range is the time interval located at the beginning of the simulation cycle. The first traffic simulation result is obtained by inputting the stitching result, the running path, and the traffic participants' first traffic information within the neighborhood of the road network map into a transformer-based multi-agent motion generation network model for traffic simulation. The intermediate simulation results are then adjusted according to preset rule constraints to obtain the first traffic simulation result, including: Within the first time range, the stitching result, the running path, and the first traffic information of the traffic participants in the neighborhood of the road network map are input into a transformer-based multi-agent motion generation network model for traffic simulation, thereby obtaining multiple intermediate simulation results within the simulation period. According to preset rule constraints, the multiple intermediate simulation results within the simulation period are adjusted to obtain multiple first traffic simulation results within the simulation period. The step of adjusting the first traffic simulation result using traffic control information to obtain the second traffic simulation result includes: By using traffic control information, multiple first traffic simulation results within the simulation period are adjusted to obtain multiple second traffic simulation results within the simulation period.

[0009] Optionally, the method further includes: Remove the traffic participants according to preset conditions; The preset conditions include at least one of the following: The location of the traffic participant extends beyond the range of the road network map; The location of the traffic participant reaches the end of the operating path; The simulated trajectory of the traffic participant exceeds the preset range compared to the actual route.

[0010] Secondly, embodiments of this application provide a device for generating and evolving the behavior of intelligent traffic participants for large-scale simulation, the device comprising: The acquisition module is used to acquire a road network map, a route coordinate file of traffic participants, and a generation plan of the traffic participants. The route coordinate file is used to constrain the running path of the traffic participants on the road network map, and the generation plan is used to determine the generation time and running status of the traffic participants on the road network map at the generation time. The traffic participants are preset traffic participation objects that make traffic behavior decisions through intelligent agents. The first processing module is used to generate the traffic participants on the road network map at the generation time based on the route coordinate file and the generation plan, and to determine the running status of the traffic participants at the generation time; The second processing module is used to perform traffic simulation using a neural network (NN) model based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants in the neighborhood range of the road network map, to obtain intermediate simulation results, and to adjust the intermediate simulation results using preset constraint rules to obtain the first traffic simulation result. The personalized parameters are used to characterize the traffic characteristics of the traffic participants. The third processing module is used to adjust the first traffic simulation result using traffic control information to obtain the second traffic simulation result. The traffic control information includes at least one of the following: operational status control information of the traffic participants, disturbance control information, safety rule constraints, and traffic light status prediction information.

[0011] Thirdly, embodiments of this application provide an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein when the program is executed by the processor, it implements the steps of the method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation as described in the first aspect.

[0012] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation as described in the first aspect.

[0013] Fifthly, embodiments of this application provide a computer program product, including computer instructions that, when executed by a processor, implement the steps of the method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation as described in the first aspect.

[0014] In this embodiment, a road network map, a route coordinate file of traffic participants, and a generation plan for the traffic participants are obtained. The route coordinate file is used to constrain the running path of the traffic participants on the road network map, and the generation plan is used to determine the generation time and running status of the traffic participants on the road network map at the generation time. The traffic participants are preset traffic participation objects for traffic behavior decision-making through intelligent agents. The traffic participants are generated on the road network map at the generation time according to the route coordinate file and the generation plan. Traffic simulation is performed based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants in the neighborhood range of the traffic participants in the road network map to obtain a first traffic simulation result. The personalized parameters are used to characterize the traffic characteristics of the traffic participants. The first traffic simulation result is adjusted using traffic control information to obtain a second traffic simulation result. The traffic control information includes at least one of the following: running status control information of the traffic participants, disturbance control information, safety rule constraints, and traffic light status prediction information.

[0015] Since the traffic participants are preset traffic participants who make traffic behavior decisions through intelligent agents, traffic simulation is performed based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants within their neighborhood in the road network map. This can distinguish the traffic preferences of different traffic participants, avoid the convergence of traffic behaviors among traffic participants, and improve the non-homogeneity of the first simulation results. Combining personalized parameters and the first traffic information, traffic simulation is performed using the neural network. The neural network completes the inference of complex micro-traffic behaviors to obtain intermediate simulation results. Then, the intermediate simulation results are adjusted by preset constraint rules to obtain accurate and reliable first traffic simulation results. The combination of the neural network and the preset rule constraints drives the precise evolution of the operating state of the traffic participants, significantly improving the realism and reliability of the first simulation results. Adjusting the first traffic simulation results, which have improved non-homogeneity, realism, and reliability, using traffic control information not only improves the non-homogeneity, realism, and reliability of the second simulation results, but also improves the control capability of the simulation process and the second simulation results through traffic control information, thereby effectively improving the simulation effect of traffic simulation. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is one of the flowcharts of a method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation provided in this application embodiment; Figure 2 This is the second flowchart of a method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation provided in this application embodiment; Figure 3 This is one of the flowcharts of a method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation provided in an embodiment of this application; Figure 4 This is a second part of the flowchart of a method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation provided in this application embodiment; Figure 5 This is the third part of a flowchart of a method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation provided in this application embodiment; Figure 6 This is the fourth part of a flowchart of a method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation provided in this application embodiment; Figure 7 This is a schematic diagram of the structure of an intelligent agent traffic participant behavior generation and evolution device for large-scale simulation provided in an embodiment of this application; Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] The terms "first," "second," etc., used in this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are generally of the same class, not limited in number; for example, the first object can be one or more. Furthermore, "and / or" in this application indicates at least one of the connected objects. For example, the scope of protection of "A and / or B" covers at least three scenarios: Scenario 1: including A but not B; Scenario 2: including B but not A; Scenario 3: including both A and B. Additionally, the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0020] See Figure 1 , Figure 1 This is one of the flowcharts for a method of generating and evolving the behavior of intelligent traffic participants for large-scale simulation provided in this application embodiment, such as... Figure 1 As shown, the method includes the following steps: Step 101: Obtain a road network map, a route coordinate file of traffic participants, and a generation plan for the traffic participants. The route coordinate file is used to constrain the running path of the traffic participants on the road network map. The generation plan is used to determine the generation time and running status of the traffic participants on the road network map at the generation time. The traffic participants are preset traffic participation objects that make traffic behavior decisions through intelligent agents. The road network map can be a structured road topology map generated through surveying and mapping, importing from an open-source geographic information platform, or on-site lidar scanning. The road network map can include lanes and traffic lights, as well as elements such as stop lines at intersections, traffic barriers, restricted areas, ramps, and curvature, and can correspond to a geographic coordinate system.

[0021] In this application, the generation and evolution of intelligent traffic participant behavior for large-scale simulation can be achieved by using traffic participant modeling units to simulate and deduce the dynamic traffic behavior of each traffic participant and output the corresponding traffic simulation results.

[0022] For example, see Figure 2 , Figure 2 This is the second flowchart of a method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation, as provided in the embodiments of this application. Figure 2 As shown: The road network map can be obtained by converting road network data from the Waymo Open dataset through the Common Navigation Data Environment (CNDE). The route coordinate file can be a type of structured route configuration file (RCF), such as including a route index (RCF_INDEX) and the route trajectory data corresponding to the route index; the route trajectory data can include road points arranged in sequence; in this application, RCF refers to the route coordinate file.

[0023] The traffic participants are preset traffic participants that make traffic behavior decisions through intelligent agents, such as traffic participants such as motor vehicles, pedestrians, or non-motor vehicles simulated by intelligent agents; the intelligent agent can be an artificial intelligence model with traffic behavior decision-making capabilities; the number of traffic participants can be one or more; in this application, since the types of traffic participants are not necessarily completely the same, they can also be referred to as heterogeneous traffic participants.

[0024] See Figure 3 , Figure 3 This is one of the flowcharts of a method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation provided in an embodiment of this application, such as... Figure 3 As shown: The generation plan may include multiple entries, each corresponding to a traffic participant, used to determine the generation of that traffic participant; each entry (spawn) in the generation plan includes the route index and generation time (index, step) of the corresponding traffic participant, as well as the running state of the corresponding traffic participant at the generation time, and may also include global parameters such as the simulation duration (which may include multiple simulation cycles, in this application one simulation cycle is also referred to as one simulation step) and the inference batch size (e.g., the number of traffic participants) for executing the intelligent agent traffic participant behavior generation and evolution method for large-scale simulation; the number of traffic participants can be adjusted by adding or deleting entries in the generation plan without adjusting the aforementioned intelligent agent or the multi-agent motion generation network model based on the attention mechanism transformer described later.

[0025] The motion state of the traffic participants may include position parameters, motion parameters, and orientation parameters, as detailed in the following first formula:

[0026] In the first formula, It can represent the i-th traffic participant; and They can be represented separately The corresponding x and y coordinates are used to determine Position parameters; It can represent the above. The orientation angle, which can be expressed in radians, is used to represent... The direction of operation; and They can be represented separately The horizontal and longitudinal components of velocity in motion parameters can be measured in meters per second. and It can represent The horizontal and longitudinal components of acceleration in motion parameters can be expressed in meters per second squared. For example, each entry can provide an index of a route, and in which simulation step the vehicle is generated, and can be accompanied by an initial velocity (spawn_speed) and a starting offset (offest), thus completing the behavior evolution along the route; The generation time can be the starting position of the simulation step (i.e., the simulation cycle) or other specified position; the generation plan can be a file in JavaScript Object Notation (JSON) format; In this step, a road network map, route coordinate files of traffic participants, and a generation plan for the traffic participants are obtained, providing a data foundation for generating the traffic participants described later and determining the first simulation result. Step 102: Generate the traffic participants on the road network map at the generation time based on the route coordinate file and the generation plan; For example, when the generation time of an entry in the generation plan is within the current simulation step, the starting point of the movement path on the road network map generates the corresponding traffic participant (i.e., agent) according to a given initial velocity and starting point offset; and the running path is registered as its predetermined route. The entire generation process is based solely on the plan and is completely deterministic, thus the result is reproducible. This is fundamentally different from random traffic flow generated by probability, facilitating scenario reproduction and control experiments.

[0027] The generation of the traffic participants can be achieved through artificial intelligence prediction models or multi-agent simulation engines. In this step, the generation process is carried out according to the production plan, thus offering the advantage of good reproducibility. Furthermore, the traffic participants are generated based on the route coordinate file and the generation plan, providing a data foundation for determining the first simulation results later.

[0028] Step 103: Based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants in the neighborhood range of the road network map, traffic simulation is performed using a neural network (NN) model to obtain intermediate simulation results. The intermediate simulation results are adjusted using preset constraint rules to obtain the first traffic simulation result. The personalized parameters are used to characterize the traffic characteristics of the traffic participants. Taking the traffic participant characterization of a motor vehicle as an example, the personalized parameters may include parameters used to characterize the traffic characteristics of the traffic participant, such as expected headway, maximum acceleration, and comfortable deceleration. The first traffic information within the neighborhood of the road network map may include the operating status of other traffic participants within the neighborhood, the status of traffic lights at the intersection, and other traffic-related information. The neighborhood range can be an area within a certain radius around the traffic participant; other traffic participants within the neighborhood range can be determined by the following second formula:

[0029] In the second formula, It can represent other traffic participants within the neighborhood of traffic participant i (also known as the i-th interactive agent or the set of neighborhood participants). It can represent the position of traffic participant i; It can represent the position of traffic participant j; The radius of the neighborhood can be represented; j can represent traffic participant j within the neighborhood of traffic participant i, provided that the positional distance between traffic participant j and traffic participant i is less than the radius of the neighborhood. ; The traffic light status at the intersection can be represented by a one-hot vector. It means that among them, " " indicates a green light" "Indicates yellow light, " "Indicates red light and " "Indicates an unknown or uncertain state.

[0030] The NN model can be a type of model that has the ability to autonomously mine deep features of data, fit complex nonlinear relationships, and perform generalized predictions on unknown samples; for example, the NN can be a Convolutional Neural Network (CNN) model, a Recurrent Neural Network (RNN) model, or a transformer-based multi-agent motion generation network model described later. The preset constraint rules include at least one of the following: constraint conditions determined by the intelligent driver model (IDM) based on reference driving behavior; traffic operation constraint conditions; traffic signal constraint conditions; and route rule constraint conditions.

[0031] In this step, on the one hand, personalized parameters can distinguish the traffic preferences of different traffic participants, avoid the convergence of traffic behaviors among traffic participants, and improve the non-homogeneity (also known as heterogeneity) of the first simulation results; on the other hand, by combining personalized parameters and the first traffic information, the neural network is used to perform traffic simulation. The neural network completes the inference of complex micro-traffic behaviors to obtain intermediate simulation results. Then, abnormal data is corrected by preset constraint rules to obtain accurate and compliant first traffic simulation results. By combining the neural network and the preset rule constraints, the operating state of the traffic participants can be accurately evolved, significantly improving the authenticity and reliability of the first simulation results.

[0032] Step 104: Adjust the first traffic simulation result using traffic control information to obtain the second traffic simulation result. The traffic control information includes at least one of the following: operational status control information for the traffic participants, disturbance control information, safety rule constraints, and traffic light status prediction information.

[0033] The operational status control information for the traffic participants can be the direct specification of the traffic participants' status and trajectory at a predetermined time or predetermined simulation step; the disturbance control information can be random or directional disturbances to the road network environment parameters or operational status of the traffic participants, thereby generating various types of simulation scenarios under different operating conditions (e.g., normal operating conditions, dangerous operating conditions, or extreme weather operating conditions); the safety rule constraints can include at least one of road traffic regulations constraints, vehicle dynamics safety threshold constraints, multi-vehicle collision avoidance constraints, intersection priority passage constraints, and extreme operating condition safety threshold constraints; the traffic light status prediction information can be a prediction of the future traffic light status to drive the traffic participants to take actions related to the signal.

[0034] The second simulation result may include the running trajectory and running status of the traffic participants, as well as the matching data between the running trajectory and the running path.

[0035] In this step, on the one hand, the first simulation result is perturbed based on the operational status control information and disturbance control information of the traffic participants, which can further reduce the homogenization problem of the second simulation result; on the other hand, the first simulation result is adjusted in combination with safety rule constraints and traffic light prediction information, so that the second simulation result is more in line with the traffic regulations of real traffic scenarios. Furthermore, the simulation process and the second simulation result can be adjusted through traffic control information, which can improve the control capability of the simulation process and simulation results, making the second simulation result more accurate and reliable, and achieving a better simulation effect.

[0036] In this embodiment, since the traffic participants are preset traffic participants who make traffic behavior decisions through intelligent agents, traffic simulation is performed based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants within the neighborhood of the road network map. This can distinguish the traffic preferences of different traffic participants, avoid the convergence of traffic behaviors among traffic participants, and improve the non-homogeneity of the first simulation results. Combining personalized parameters and the first traffic information, traffic simulation is performed using the neural network. The neural network completes the complex micro-traffic behavior deduction to obtain intermediate simulation results. Then, the intermediate simulation results are adjusted by preset constraint rules to obtain accurate and reliable first traffic simulation results. The combination of the neural network and the preset rule constraints drives the precise evolution of the traffic participants' operating states, significantly improving the realism and reliability of the first simulation results. Adjusting the first traffic simulation results, which have improved non-homogeneity, realism, and reliability, using traffic control information not only improves the non-homogeneity, realism, and reliability of the second simulation results, but also improves the control capability of the simulation process and simulation results through traffic control information, thereby effectively improving the simulation effect of traffic simulation.

[0037] In some implementations, the route coordinate file includes a route index and route trajectory data corresponding to the route index; The generation of traffic participants on the road network map at the generation time based on the route coordinate file and the generation plan includes: The route trajectory data is determined in the route coordinate file based on the route index; The route trajectory data is used to perform proximity matching on the roads in the road network map to obtain the first lane sequence; Perform topological feasibility filtering on the first lane sequence to obtain the running path of the traffic participant on the road network map; The traffic participant is generated at the starting point of the running path at the generation time.

[0038] Please see below. Figure 2 The step of determining the route trajectory data in the route coordinate file based on the route index can be achieved by searching the RCF using the route index to obtain the route trajectory data corresponding to the route index; the route trajectory data corresponding to the route index is then matched with the nearest available match (e.g., using a nearest-access matcher). Figure 2 The directSnapMatcher in the map maps the route trajectory data to a series of lanes on the road network map, resulting in the first lane sequence (i.e., Figure 2 (lane_sequence in the context of lane) Topological feasibility filtering of the first lane sequence can be used to remove disconnected or illegal lanes, thereby obtaining the running path of the traffic participants on the road network map; In this embodiment, the route trajectory data is used to perform proximity matching on roads in the road network map to obtain a first lane sequence, which improves the fit between the first lane sequence and the road network map. Topological feasibility filtering is then applied to the first lane sequence with improved fit to ensure that the running path conforms to the road network connectivity topology rules and improves the matching accuracy between the running path and the road network map. Based on the starting point of the running path that conforms to the road network connectivity topology rules and has improved matching accuracy with the road network map, at the generation time (e.g., Figure 2 When the simulation reaches the spawn step, the traffic participants (i.e., the corresponding agents) are generated, which can effectively improve the initial position accuracy and operational accuracy of the traffic participants, thereby improving the simulation accuracy and simulation effect of the first and second traffic simulation results described later.

[0039] In some implementations, the NN model is a multi-agent motion generation network model based on the attention mechanism transformer; Based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants within their neighborhood in the road network map, a neural network (NN) model is used to perform traffic simulation, obtaining intermediate simulation results. These intermediate simulation results are then adjusted using preset constraint rules to obtain the first traffic simulation result, including: The personalized parameters of the traffic participants, the types of the traffic participants, and the historical simulation results of the traffic participants are spliced ​​together to obtain the spliced ​​result; The stitching result, the running path, and the first traffic information of the traffic participants within the neighborhood of the road network map are input into a transformer-based multi-agent motion generation network model for traffic simulation to obtain intermediate simulation results. The intermediate simulation results are then adjusted according to preset rule constraints to obtain the first traffic simulation result. The preset rule constraints include at least one of the following: constraints determined by the intelligent driver model (IDM) based on reference driving behavior; traffic operation constraints; traffic signal constraints; and route rule constraints.

[0040] The types of traffic participants may include, but are not limited to, motor vehicles, non-motor vehicles and pedestrians. The types of traffic participants may also be further classified according to other preset conditions (such as speed, volume, safety factor and load capacity, etc.) and attribute parameters under the corresponding types may be assigned accordingly. The historical simulation results may include the first simulation results and the second simulation results of the traffic participants within a preset time range. For example, the historical simulation results are the operating status of the traffic participants in the first simulation results (e.g., multiple first simulation results corresponding to a preset number of recording steps before the first recording step, as described later) within the preset time range.

[0041] The transformer-based multi-agent motion generation network model can be an artificial intelligence model under a neural network framework. This model can use vectorized encoding to uniformly represent the first simulation result (e.g., the running trajectory of traffic participants in the first simulation result) and the road network map. Relying on the self-attention mechanism, it accurately models the interaction relationship between agents and the coupling constraints between agents and the road environment. It can output diverse simulation results in multi-agent concurrent scenarios, which is highly adapted to the core business needs of this application for multi-agent, large-scale simulation and generative scenario construction.

[0042] The transformer-based multi-agent motion generation network model may include: an input encoding layer, an interaction layer, and a decoding layer; The input encoding layer is used to perform polyline-level encoding on the lanes and edges of the road network map (which can be a sequence of points interpolated at intervals of 0.5 meters or other intervals) to obtain a map representation. The input encoding layer is also used to stitch together the historical simulation results of traffic participants (i.e., the operating status of the traffic participants within a preset time range), the type of traffic participants, and the personalized parameters (which can be the stitching together of the codes of these three) to obtain a stitched result (also known as an agent representation). For details, please refer to the following third formula:

[0043] in, It can represent the initial representation of traffic participants (intelligent agents); It can represent a historical trajectory encoder (such as a point-by-point multilayer perceptron with pooling or recurrent networks) used to encode the historical simulation results of the traffic participants; It can represent historical simulation results, such as the operational status of traffic participants as described in the most recent H frame; This can represent the type code of the traffic participant; This can represent the embedding of personalized parameters of the traffic participants; "Can represent a splicing operation; The interaction layer is used to stack multi-head attention, specifically including self-attention between traffic participants (i.e., between agents) (the visible range is determined by the aforementioned neighborhood range), attention between traffic participants and the road network map (between agents and the road network map), and extraction of interactions between traffic participants, between traffic participants and the road network map, and between traffic participants and traffic lights. The calculation of the attention can refer to the following fourth formula:

[0044] In the fourth formula, , , These are the query, key, and value matrices (obtained by linear projection from the agent / map representation); It can represent the dimension of the key vector; The decoding layer is used to decode the simulation intermediate results (or a sequence of simulation intermediate results composed of multiple simulation intermediate results) using the agent (i.e., the splicing result) as the query; it can also output multimodal simulation intermediate results (e.g., multiple candidate running trajectories and their probabilities), and take the optimal one or sample according to probability to further reduce the homogeneity of simulation results. The simulation intermediate results are constrained by the preset rules. Adjustments were made to obtain the first traffic simulation result that falls within the feasible region; The preset rule constraints It is a feasible domain constructed by IDM based on at least one of the constraints determined by reference driving behavior, traffic operation constraints, traffic signal constraints, and route rule constraints.

[0045] The IDM can be a car-following model used to simulate car-following behavior among traffic participants in traffic flow; the reference driving behavior can be the driving behavior of real traffic participants or preset compliant traffic driving behavior. The traffic signal constraints may include constraints corresponding to traffic lights (i.e., traffic light constraints), or other constraints (such as constraints corresponding to traffic signals issued by traffic enforcement officers).

[0046] In this embodiment, based on the personalized parameters of the traffic participants, the types of all traffic participants, the historical simulation results of the traffic participants, and the running paths and the first traffic information of the traffic participants within the neighborhood of the road network map, a transformer-based multi-agent motion generation network model is used for traffic simulation. This model can accurately predict the complex interaction relationships between the traffic participants, effectively reduce the prediction bias of the running states of the traffic participants, and improve the accuracy and reliability of the intermediate simulation results. Since the preset rules include one of the constraints determined by the IDM based on reference driving behavior, traffic operation constraints, traffic signal constraints, and route rule constraints, adjusting the intermediate simulation results, which improve both reliability and accuracy, according to the preset rule constraints not only improves the reliability and accuracy of the first simulation results but also makes the first traffic simulation results more adaptable to real traffic scenarios, thereby improving the simulation effect of traffic simulation.

[0047] In some implementations, the transformer-based multi-agent motion generation network model is trained in the following manner: Obtain the real trajectory dataset of the traffic participants on the road network map; A first loss and a second loss are determined based on the real trajectory dataset and the second traffic simulation results. The first loss is the regression loss of the second traffic simulation results relative to the real trajectory dataset, and the second loss is the cross-entropy loss of the second traffic simulation results relative to the real trajectory dataset. Construct a loss function based on the first loss and the second loss; The transformer-based multi-agent motion generation network model is trained according to the loss function to obtain the trained transformer-based multi-agent motion generation network model.

[0048] The real trajectory dataset is obtained by filtering the aforementioned Vimo public dataset. The trajectory data in the real trajectory dataset can be obtained by performing polyline interpolation and lane / edge / traffic light type mapping conversion on the data in the aforementioned Vimo public dataset. The loss function constructed based on the first loss and the second loss can refer to the following fifth formula:

[0049] In the fifth formula, The loss function can be represented as described above; This can represent the first loss (which can be either the absolute error regression loss (L1 regression loss) or the Huber regression loss (Huber regression loss)). This can represent a second loss, such as the pattern classification loss of the running trajectory in the second traffic simulation result relative to the running trajectory in the real trajectory dataset (specifically, the cross-entropy loss, used to determine the optimal running trajectory). Weights can be represented and can be set as needed by those skilled in the art.

[0050] Among them, the The determination can be made with reference to the following sixth formula:

[0051] In the sixth formula, This can represent the first loss in the fifth formula above; It can be expressed as in Time steps (e.g., as described later) The actual running trajectory of each recorded step; It can be expressed as in The predicted trajectory of the time step; C can represent the preset simulation duration (e.g., C recording steps). During the training of the transformer-based multi-agent motion generation network model according to the loss function, the model parameters can be updated using an Adaptive Moment Estimation (Adam) optimizer; the real trajectory dataset can also be divided into training and validation sets for batch training according to the scene, and the batch training size can be set as needed by those skilled in the art (e.g., less than or equal to 8 samples when memory is limited); training can be combined with rolling multi-step prediction to improve long-term evolution stability.

[0052] The hyperparameters of the transformer-based multi-agent motion generation network model (which may include, but are not limited to, the number of layers, the number of attention heads, the hidden dimension, the learning rate, the interaction radius r, the preset time range corresponding to the historical simulation results, and the number of generation steps C) can be tuned using a validation set. After training, the physical feasibility and statistical consistency of the generated trajectory (e.g., velocity / spacing distribution, and flow-density relationship) are verified on a reserved scene.

[0053] In this embodiment, regression loss and cross-entropy loss are introduced simultaneously to construct the loss function: regression loss accurately constrains continuous numerical errors such as trajectory coordinates and displacement, ensuring that the first simulation result fits the running path; cross-entropy loss constrains discrete driving modes and corrects behavior category deviations; the multi-agent motion generation network model based on transformer is trained through the above first and second losses, so that the model can achieve better simulation results when performing simulations.

[0054] In some implementations, the generation time is within a first time range of the simulation cycle of the method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation, wherein the first time range is the time interval located at the beginning of the simulation cycle. The first traffic simulation result is obtained by inputting the stitching result, the running path, and the traffic participants' first traffic information within the neighborhood of the road network map into a transformer-based multi-agent motion generation network model for traffic simulation. The intermediate simulation results are then adjusted according to preset rule constraints to obtain the first traffic simulation result, including: Within the first time range, the stitching result, the running path, and the first traffic information of the traffic participants in the neighborhood of the road network map are input into a transformer-based multi-agent motion generation network model for traffic simulation, thereby obtaining multiple intermediate simulation results within the simulation period. According to preset rule constraints, the multiple intermediate simulation results within the simulation period are adjusted to obtain multiple first traffic simulation results within the simulation period. The step of adjusting the first traffic simulation result using traffic control information to obtain the second traffic simulation result includes: By using traffic control information, multiple first traffic simulation results within the simulation period are adjusted to obtain multiple second traffic simulation results within the simulation period.

[0055] The simulation cycle can be referred to as a simulation step; a simulation step can be divided into multiple time ranges (i.e., multiple recording steps, one time range corresponds to one recording step, and the durations of these time ranges can be equal or unequal); a recording step can also be denoted as control_step, and the number of each simulation step can be represented by C; Specifically, the conversion relationship between simulation steps and recording steps can be: ,in It can represent the duration of a simulation step. It can represent the duration of a recording step, and C can represent that a simulation step includes C recording steps; The total number of recorded steps satisfies ,in It can represent the number of pre-laid historical records; specifically, the parameter H corresponds to the number of historical records. , It can represent the total number of simulation steps during the entire simulation period; The historical simulation result can be the first simulation result corresponding to a preset number of time steps before the first recording step; For example, please see Figure 4 , Figure 4 This is a second part of a flowchart of a method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation, provided in an embodiment of this application. Figure 4 As shown: The simulation step (e.g., simulation step 0, simulation step 1, or simulation step 2) is 0.5 seconds, including 5 recording steps, each recording step lasting 0.1 seconds; taking the simulation step including five recording steps, with the first time being the first recording step (i.e., the first recording step) in the simulation step as an example: The generation time is located in the first recording step; in the first recording step, the traffic participants are generated according to the route coordinate file and the generation plan; In the first recording step, the stitching result, the running path, and the first traffic information of the traffic participants within the neighborhood of the road network map are input into a transformer-based multi-agent motion generation network model for traffic simulation. The intermediate simulation results are adjusted according to preset rule constraints to obtain the first traffic simulation results corresponding to the five recording steps in the simulation step. Specifically, please refer to the following seventh formula:

[0056] In the seventh formula, It can represent the first simulation result of each recording step in the simulation step (i.e., each simulation step includes C recording steps, from t+1 to t+C, a total of C recording steps); This can represent the aforementioned preset rule constraints; It can represent neural network (NN) models (such as the transformer-based multi-agent motion generation network model in this application). It can represent model parameters; It can represent historical simulation results; It can represent other traffic participants within the neighborhood. Both can represent the status of traffic lights, and together they represent the first traffic information; It is understandable that, according to the seventh formula above, the model only infers once in a simulation step, generating C first traffic simulation results corresponding to C future recording steps (e.g., five recording steps); it can not only balance inference efficiency and inference cost, but also generate multiple continuous micro-trajectories and second traffic simulation results. The preset rule constraint IDM may include multiple parameters (also referred to as IDM parameters in this application); the IDM parameters may include one or more of the following: feasible acceleration, desired speed limit, minimum stationary distance, and acceleration exponent, which may be specifically taken according to preset specifications or calibrated by historical data; For example, the IDM parameters provide feasible acceleration, and based on this, limit speed and following distance; simultaneously, impose constraints such as speed limits, red light stop lines, and driving along a predetermined route; specifically, refer to the following eighth formula:

[0057] In the eighth formula, This can represent the instantaneous acceleration of the traffic participant (intelligent driver model). This is the maximum acceleration; Current speed; The desired speed (based on the road speed limit); The acceleration index; This refers to the actual distance between the front of the vehicle; The desired spacing; This represents the minimum stationary spacing. The desired headway; To account for the speed difference with the vehicle in front; To improve comfort, neural networks reduce speed. They enable more realistic and diverse behaviors, while pre-defined constraints ensure that behaviors do not exceed physical and safety boundaries; the two complement each other.

[0058] By using the above-mentioned NN generation and preset rule constraints to drive the process, the first traffic simulation result in each recording step is finally obtained.

[0059] In this embodiment, traffic simulation is performed using a transformer-based multi-agent motion generation network model within a first time range at the start of the simulation cycle, resulting in multiple first traffic simulation results within the simulation cycle. This reduces the number of model inferences and improves inference efficiency. Furthermore, traffic control information is used to adjust the multiple first traffic simulation results within the simulation cycle, resulting in multiple second traffic simulation results within the simulation cycle. This increases the number of simulation results and enhances their reference value. Due to the reduced number of inferences, improved inference efficiency, and a sufficient number of more reliable simulation results, the effectiveness of traffic simulation is significantly improved.

[0060] In some implementations, please refer to Figure 2 This application provides a traffic simulation device, which is applied to the above-mentioned method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation, including: module 1, module 2, module 3 and module 4; Module 1 is used for route-driven multi-agent generation and injection (matching RCF and lane sequence). The module 2 is used to mix the above-mentioned NN with preset constraint rules to perform multi-resolution behavior evolution (for example, setting five recording steps in a simulation step can increase the number of simulation results in each simulation step, thereby improving the resolution of behavior evolution during the simulation process). Module 3 provides an external intervention interface (to obtain the aforementioned traffic control information); since the external intervention interface obtains traffic control information in each simulation step, it is also called a single-step intervention interface; Module 4 is used for lifecycle management of traffic participants and heterogeneous scenarios.

[0061] Furthermore, the simulation is iterated in each step: traffic participants are generated at the generation time (i.e., agents are generated upon expiration (spawn_due_agent)), traffic participants are deleted midway (remove_agents) / replenished (generate_agents), the first simulation result of each recording step is determined using the aforementioned NN and the aforementioned preset rule constraints, and the result is determined through control information (which can be used) Figure 2 Adjust the values ​​obtained from the external_step to obtain the second simulation result; Furthermore, the second simulation results are output (specifically, a binary file scenario.pkl containing sparse, record-by-record step-by-step trajectory and dynamic traffic light information, and a JSON-formatted traffic simulation file corresponding to the matching results of the actual running trajectory and the running path / the birth and death logs of traffic participants (which may include the matching data of the running path (i.e.)). Figure 2 The route matching and birth / death logs in manifest.json are used to achieve large-scale, reproducible, and controllable multi-agent behavior evolution.

[0062] In some implementations, please refer to Figure 5 , Figure 5 This is part three of the flowcharts for a method of generating and evolving the behavior of intelligent traffic participants for large-scale simulation provided in this application embodiment. Figure 5 As shown: The order of calling each simulation step is fixed as follows: propagation generation (step), external intervention (which can be used) Figure 5 The external step (represented by `external_step`) and write / update are the parameters of the external intervention interface, which are the future parameters generated in this simulation step. Each record step state and signal light state can be returned as is (observation only, no modification), or the states can be modified and returned while keeping the key set unchanged; see the ninth formula below for details:

[0063] In the ninth formula, It can represent an external interference quantifier, used to characterize the aforementioned traffic control information; The state after intervention, i.e., the second traffic simulation result; without damaging the data structure or affecting downstream writing and subsequent progress; X can represent the state before intervention; The function represents a function that reads all traffic participant identifiers from the status data, such as obtaining the number of each traffic participant; the status data can be stored in a key-value pair format of "traffic participant identifier - status value", where the traffic participant identifier is the key, and parameters such as position, speed, acceleration, and orientation angle can be the status values ​​corresponding to the key; This indicates that the traffic participant identifiers are completely identical before and after external intervention, meaning that external intervention cannot add or delete traffic participants, but can only modify the status values ​​of existing traffic participants.

[0064] The external intervention interface in module 3 allows for four types of operations: first, external control, directly specifying the future state / trajectory of a vehicle; second, disturbance injection, artificially adding disturbances to construct test scenarios; third, safety constraints, imposing additional safety restrictions on the generated state; and fourth, traffic light prediction, inputting predictions of future traffic lights to drive signal-related behaviors. This interface enables generative simulation to possess standardized controllability and editability.

[0065] In some embodiments, the method further includes: Remove the traffic participants according to preset conditions; The preset conditions include at least one of the following: The location of the traffic participant extends beyond the range of the road network map; The location of the traffic participant reaches the end of the operating path; The simulated trajectory of the traffic participant exceeds the preset range compared to the actual route.

[0066] For details, please see Figure 6 , Figure 6 This is a partial flowchart of a method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation, provided in an embodiment of this application. Figure 6 As shown: Agents can be generated (by plan or probability) at the end of the running path or at the boundary of the road network; the removal end removes agents that have reached the end, have significantly deviated from the predetermined route, or have left the road network according to a unified rule; for details, please refer to the following formula ten:

[0067] In the tenth formula, The lateral deviation of the agent from its predetermined route (i.e., the aforementioned operating path); This is a deviation from the threshold; This represents the remaining drivable distance along the route. To reach the judgment threshold; This is a logical OR operation. The agent is removed if any condition is met, ensuring consistency in the dynamic entry and exit of agents and their routes in large-scale scenarios. Among them, deviation from threshold The value can be determined by those skilled in the art through statistical analysis of the lateral deviation distance of traffic participants relative to the predetermined route in real trajectories, combined with lane width, lateral dimensions of traffic participants, and map matching error; furthermore, multiple candidate values ​​can be tested in validation data, and the optimal value or optimal range can be determined based on the lowest erroneous removal rate of normal traffic participants and the lowest omission removal rate of traffic participants deviating from the route; specifically, the deviation threshold The preferred range is 0.5 meters to 2.0 meters; more specifically, the deviation from the threshold... The preferred value is 1.5 meters; Determination threshold The value can be determined by those skilled in the art based on the route trajectory data at sampling intervals of a set length (e.g., approximately 0.5 meters), a set recording step duration (e.g., 0.1 seconds), the travel distance of traffic participants within one set recording step, and the positional error. Furthermore, multiple candidate values ​​can be tested by replaying real trajectories with endpoint arrival markers. The criterion is to determine the optimal range or optimal value, with early or delayed removal not exceeding one set recording step and the lowest omission rate. Specifically, the threshold value is determined... The preferred range is 0.5 meters to 3.0 meters; more specifically, the determination threshold... The preferred value is 2.0 meters.

[0068] In this embodiment, the traffic participants are removed according to preset conditions to complete the lifecycle management of traffic participants; this avoids invalid traffic participants from continuously participating in the simulation process, ensures that the set of intelligent agents in the road network map matches the real traffic operation status, improves the reliability and accuracy of the simulation, and thus improves the simulation effect.

[0069] In some implementations, different types of traffic participants, such as motor vehicles, pedestrians, and non-motor vehicles, uniformly use the state parameters in the aforementioned first formula (i.e. Figure 6 The 7-dimensional state of the dynamic traffic lights is represented by a type code, and different rule parameters are applied to different types (e.g., reducing the maximum acceleration and increasing the safety distance for non-motorized vehicles); the unique state of the dynamic traffic lights is updated step by step.

[0070] In some implementations, please refer to Figure 6 Meanwhile, traffic participants are generated according to the generation plan or boundary probability at the generation time to achieve better traffic participant joining patterns and traffic simulation effects. In some implementations, the second traffic simulation result may further include a scene file serialized and saved in a sparse manner according to the recording step index (i.e. Figure 6 The `scenario.pkl` file contains: step-by-step traffic participant status, dynamic traffic light status, vehicle static dimensions, road network and routes, and lane sequence and operation summary for each agent under route-driven conditions; it can also control the number of traffic participants participating in inference in batches (e.g., no more than 8 per batch) to prevent memory overflow (OOM).

[0071] In some implementations, the simulation focuses on the morning rush hour in the central urban area of ​​a city. The road network map contains dozens of signal-controlled intersections and hundreds of road segments, aiming to reproduce the behavior and traffic flow evolution of tens of thousands of heterogeneous traffic participants.

[0072] Simulation parameters: Simulation step duration 0.5 seconds, recording step duration 0.1 seconds, number of recorded steps per simulation step C=5, initial history H=11, total simulation steps The time period is set; the neural network inference batch size is ≤8 when there are many vehicles to prevent memory overflow. The behavior generation network adopts a Transformer-based multi-agent motion generation network model, with historical steps H=11, generated steps C=5, and the interaction radius set to tens of meters according to the scene. It is trained offline on a real trajectory dataset (such as maps and trajectories converted from Waymo Open Dataset) with trajectory regression loss supervision and Adam optimization; IDM and other rule parameters are set according to specifications or calibrated by historical data.

[0073] The input consists of a road network map, a route coordinate file (RCF), and a generation plan (JSON). The map is obtained from a publicly available dataset scene through polyline interpolation (approximately 0.5 meters) and type mapping. The traffic flow scale is adjusted by adding or deleting entries in the generation plan (examples of 1 / 10 / 50 / 200 vehicles are provided). Each entry specifies the route index, generation simulation step, initial velocity, and starting point offset. The runtime engine loops according to the simulation steps: agents arriving at a point are injected with the nearest matched lane sequence (this lane sequence can be the first lane sequence mentioned above, or it can be a second lane sequence obtained by adjusting the first lane sequence (e.g., adjusting the topology)); at each step, second traffic information is constructed for each agent (i.e., the traffic participants and traffic light status information within the neighborhood in this embodiment, which can be obtained by adjusting the first traffic information). The behavior generation network model generates the state of the next 5 recording steps at once, and the rules are projected according to the IDM parameters, speed limit, safety distance, red light stop line, and route. Subsequently, the scenario is written via a single-step intervention interface (which can inject control / disturbance / traffic light prediction). Agents that reach the destination, deviate from the route by more than a threshold, or leave the road network are automatically removed and recorded in the operation summary. At the end of the simulation, a sparse scenario file is output (the seven-dimensional state of vehicles, dynamic traffic light states, static dimensions, road network and routes, and lane sequences of vehicles on each route) along with matching and birth / death logs. Large-scale scenarios of different sizes and routes can be reproduced by modifying only the generation plan, and external control or disturbance injection can be applied to specified vehicles using the single-step intervention interface to construct control experiments.

[0074] It should be noted that the above-mentioned method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation can be executed by an electronic device, that is, all the steps included in the above method are executed by the electronic device, which can be an electronic device such as a server, computer or mobile phone.

[0075] See Figure 7 , Figure 7 This is a schematic diagram of the structure of an intelligent agent traffic participant behavior generation and evolution device for large-scale simulation provided in an embodiment of this application, as shown below. Figure 7 As shown, the intelligent agent traffic participant behavior generation and evolution device 700 for large-scale simulation includes: The acquisition module 701 is used to acquire a road network map, a route coordinate file of traffic participants, and a generation plan of the traffic participants. The route coordinate file is used to constrain the running path of the traffic participants on the road network map, and the generation plan is used to determine the generation time and running status of the traffic participants on the road network map at the generation time. The traffic participants are preset traffic participation objects that make traffic behavior decisions through intelligent agents. The first processing module 702 is used to generate the traffic participants on the road network map at the generation time based on the route coordinate file and the generation plan, and to determine the running status of the traffic participants at the generation time; The second processing module 703 is used to perform traffic simulation using a neural network (NN) model based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants in the neighborhood range of the road network map, to obtain intermediate simulation results, and to adjust the intermediate simulation results using preset constraint rules to obtain the first traffic simulation result. The personalized parameters are used to characterize the traffic characteristics of the traffic participants. The third processing module 704 is used to adjust the first traffic simulation result using traffic control information to obtain a second traffic simulation result. The traffic control information includes at least one of the following: operational status control information for the traffic participants, disturbance control information, safety rule constraints, and traffic light status prediction information.

[0076] Optionally, the route coordinate file includes a route index and the route trajectory data corresponding to the route index; The generation of traffic participants on the road network map at the generation time based on the route coordinate file and the generation plan includes: The route trajectory data is determined in the route coordinate file based on the route index; The route trajectory data is used to perform proximity matching on the roads in the road network map to obtain the first lane sequence; Perform topological feasibility filtering on the first lane sequence to obtain the running path of the traffic participant on the road network map; The traffic participant is generated at the starting point of the running path at the generation time.

[0077] Optionally, the NN model is a multi-agent motion generation network model based on the attention mechanism transformer; Based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants within their neighborhood in the road network map, a neural network (NN) model is used to perform traffic simulation, obtaining intermediate simulation results. These intermediate simulation results are then adjusted using preset constraint rules to obtain the first traffic simulation result, including: The personalized parameters of the traffic participants, the types of the traffic participants, and the historical simulation results of the traffic participants are spliced ​​together to obtain the spliced ​​result; The stitching result, the running path, and the first traffic information of the traffic participants within the neighborhood of the road network map are input into a transformer-based multi-agent motion generation network model for traffic simulation to obtain intermediate simulation results. The intermediate simulation results are then adjusted according to preset rule constraints to obtain the first traffic simulation result. The preset rule constraints include at least one of the following: constraints determined by the intelligent driver model (IDM) based on reference driving behavior; traffic operation constraints; traffic signal constraints; and route rule constraints.

[0078] Optionally, the transformer-based multi-agent motion generation network model is trained in the following manner: Obtain the real trajectory dataset of the traffic participants on the road network map; A first loss and a second loss are determined based on the real trajectory dataset and the second traffic simulation results. The first loss is the regression loss of the second traffic simulation results relative to the real trajectory dataset, and the second loss is the cross-entropy loss of the second traffic simulation results relative to the real trajectory dataset. Construct a loss function based on the first loss and the second loss; The transformer-based multi-agent motion generation network model is trained according to the loss function to obtain the trained transformer-based multi-agent motion generation network model.

[0079] Optionally, the generation time is within a first time range of the simulation cycle of the method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation, and the first time range is the time interval located at the beginning of the simulation cycle. The first traffic simulation result is obtained by inputting the stitching result, the running path, and the traffic participants' first traffic information within the neighborhood of the road network map into a transformer-based multi-agent motion generation network model for traffic simulation. The intermediate simulation results are then adjusted according to preset rule constraints to obtain the first traffic simulation result, including: Within the first time range, the stitching result, the running path, and the first traffic information of the traffic participants in the neighborhood of the road network map are input into a transformer-based multi-agent motion generation network model for traffic simulation, thereby obtaining multiple intermediate simulation results within the simulation period. According to preset rule constraints, the multiple intermediate simulation results within the simulation period are adjusted to obtain multiple first traffic simulation results within the simulation period. The step of adjusting the first traffic simulation result using traffic control information to obtain the second traffic simulation result includes: By using traffic control information, multiple first traffic simulation results within the simulation period are adjusted to obtain multiple second traffic simulation results within the simulation period.

[0080] Optionally, the intelligent agent traffic participant behavior generation and evolution device 700 for large-scale simulation also includes a fourth processing module; The fourth processing module is used to remove the traffic participants according to preset conditions; The preset conditions include at least one of the following: The location of the traffic participant extends beyond the range of the road network map; The location of the traffic participant reaches the end of the operating path; The simulated trajectory of the traffic participant exceeds the preset range compared to the actual route.

[0081] The intelligent agent traffic participant behavior generation and evolution device 700 for large-scale simulation can realize each process of the above-mentioned intelligent agent traffic participant behavior generation and evolution method for large-scale simulation. The technical features are one-to-one and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0082] This application also provides an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor. When the program is executed by the processor, it implements the various processes of the above-described embodiment of the intelligent agent traffic participant behavior generation and evolution method for large-scale simulation, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0083] For details, see Figure 8 This application also provides an electronic device, including a bus 801, a transceiver 802, an antenna 803, a bus interface 804, a processor 805, and a memory 806.

[0084] The transceiver 802 is used to acquire a road network map, a route coordinate file of traffic participants, and a generation plan of the traffic participants. The route coordinate file is used to constrain the running path of the traffic participants on the road network map, and the generation plan is used to determine the generation time of the traffic participants on the road network map and their running status at the generation time. The traffic participants are preset traffic participation objects that make traffic behavior decisions through intelligent agents. The processor 805 is configured to generate the traffic participants on the road network map at the generation time based on the route coordinate file and the generation plan; Based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants in the neighborhood range of the road network map, traffic simulation is performed using a neural network (NN) model to obtain intermediate simulation results. The intermediate simulation results are then adjusted using preset constraint rules to obtain the first traffic simulation result. The personalized parameters are used to characterize the traffic characteristics of the traffic participants. The first traffic simulation result is adjusted using traffic control information to obtain a second traffic simulation result. The traffic control information includes at least one of the following: operational status control information for the traffic participants, disturbance control information, safety rule constraints, and traffic light status prediction information.

[0085] Optionally, the route coordinate file includes a route index and the route trajectory data corresponding to the route index; The generation of traffic participants on the road network map at the generation time based on the route coordinate file and the generation plan includes: The route trajectory data is determined in the route coordinate file based on the route index; The route trajectory data is used to perform proximity matching on the roads in the road network map to obtain the first lane sequence; Perform topological feasibility filtering on the first lane sequence to obtain the running path of the traffic participant on the road network map; The traffic participant is generated at the starting point of the running path at the generation time.

[0086] Optionally, the NN model is a multi-agent motion generation network model based on the attention mechanism transformer; Based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants within their neighborhood in the road network map, a neural network (NN) model is used to perform traffic simulation, obtaining intermediate simulation results. These intermediate simulation results are then adjusted using preset constraint rules to obtain the first traffic simulation result, including: The personalized parameters of the traffic participants, the types of the traffic participants, and the historical simulation results of the traffic participants are spliced ​​together to obtain the spliced ​​result; The stitching result, the running path, and the first traffic information of the traffic participants within the neighborhood of the road network map are input into a transformer-based multi-agent motion generation network model for traffic simulation to obtain intermediate simulation results. The intermediate simulation results are then adjusted according to preset rule constraints to obtain the first traffic simulation result. The preset rule constraints include at least one of the following: constraints determined by the intelligent driver model (IDM) based on reference driving behavior; traffic operation constraints; traffic signal constraints; and route rule constraints.

[0087] Optionally, the transformer-based multi-agent motion generation network model is trained in the following manner: Obtain the real trajectory dataset of the traffic participants on the road network map; A first loss and a second loss are determined based on the real trajectory dataset and the second traffic simulation results. The first loss is the regression loss of the second traffic simulation results relative to the real trajectory dataset, and the second loss is the cross-entropy loss of the second traffic simulation results relative to the real trajectory dataset. Construct a loss function based on the first loss and the second loss; The transformer-based multi-agent motion generation network model is trained according to the loss function to obtain the trained transformer-based multi-agent motion generation network model.

[0088] Optionally, the generation time is within the first time range of the simulation cycle of the method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation. The first traffic simulation result is obtained by inputting the stitching result, the running path, and the traffic participants' first traffic information within the neighborhood of the road network map into a transformer-based multi-agent motion generation network model for traffic simulation. The intermediate simulation results are then adjusted according to preset rule constraints to obtain the first traffic simulation result, including: Within the first time range, the stitching result, the running path, and the first traffic information of the traffic participants in the neighborhood of the road network map are input into a transformer-based multi-agent motion generation network model for traffic simulation. The intermediate simulation results are adjusted according to preset rule constraints to obtain multiple first traffic simulation results within the simulation period. The step of adjusting the first traffic simulation result using traffic control information to obtain the second traffic simulation result includes: By using traffic control information, multiple first traffic simulation results within the simulation period are adjusted to obtain multiple second traffic simulation results within the simulation period.

[0089] Optionally, the processor 805 is further configured to remove the traffic participant according to preset conditions; The preset conditions include at least one of the following: The location of the traffic participant extends beyond the range of the road network map; The location of the traffic participant reaches the end of the operating path; The simulated trajectory of the traffic participant exceeds the preset range compared to the actual route.

[0090] exist Figure 8 In this document, a bus architecture (represented by bus 801) is used. Bus 801 can include any number of interconnected buses and bridges, linking various circuits including one or more processors represented by processor 805 and memory represented by memory 806. Bus 801 can also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 804 provides an interface between bus 801 and transceiver 802. Transceiver 802 can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 805 is transmitted over a wireless medium via antenna 803, which further receives data and transmits data to processor 805.

[0091] The processor 805 manages the bus 801 and handles general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. The memory 806 can be used to store data used by the processor 805 during operation.

[0092] Optionally, the processor 805 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a complex programmable logic device (CPLD).

[0093] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, this computer program implements the various processes of the above-described embodiment of the intelligent agent traffic participant behavior generation and evolution method for large-scale simulation, and achieves the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0094] This application also provides a computer program product, including computer instructions. When executed by a processor, these computer instructions implement the various processes of the above-described embodiment of the method for generating and evolving the behavior of intelligent traffic participants for large-scale simulation, and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0095] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0096] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0097] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for generating and evolving the behavior of intelligent traffic participants in large-scale simulation, characterized in that, The method includes: The system acquires a road network map, a route coordinate file of traffic participants, and a generation plan for the traffic participants. The route coordinate file is used to constrain the running path of the traffic participants on the road network map, and the generation plan is used to determine the generation time and running status of the traffic participants on the road network map at the generation time. The traffic participants are preset traffic participation objects that make traffic behavior decisions through intelligent agents. The traffic participants are generated on the road network map at the generation time based on the route coordinate file and the generation plan. Based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants in the neighborhood range of the road network map, traffic simulation is performed using a neural network (NN) model to obtain intermediate simulation results. The intermediate simulation results are then adjusted using preset constraint rules to obtain the first traffic simulation result. The personalized parameters are used to characterize the traffic characteristics of the traffic participants. The first traffic simulation result is adjusted using traffic control information to obtain a second traffic simulation result. The traffic control information includes at least one of the following: operational status control information for the traffic participants, disturbance control information, safety rule constraints, and traffic light status prediction information.

2. The method according to claim 1, characterized in that, The route coordinate file includes a route index and the route trajectory data corresponding to the route index; The generation of traffic participants on the road network map at the generation time based on the route coordinate file and the generation plan includes: The route trajectory data is determined in the route coordinate file based on the route index; The route trajectory data is used to perform proximity matching on the roads in the road network map to obtain the first lane sequence; Perform topological feasibility filtering on the first lane sequence to obtain the running path of the traffic participant on the road network map; The traffic participant is generated at the starting point of the running path at the generation time.

3. The method according to claim 1, characterized in that, The NN model is a multi-agent motion generation network model based on the attention mechanism transformer. Based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants within their neighborhood in the road network map, a neural network (NN) model is used to perform traffic simulation, obtaining intermediate simulation results. These intermediate simulation results are then adjusted using preset constraint rules to obtain the first traffic simulation result, including: The personalized parameters of the traffic participants, the types of the traffic participants, and the historical simulation results of the traffic participants are spliced ​​together to obtain the spliced ​​result; The stitching result, the running path, and the first traffic information of the traffic participants within the neighborhood of the road network map are input into a transformer-based multi-agent motion generation network model for traffic simulation to obtain intermediate simulation results. The intermediate simulation results are then adjusted according to preset rule constraints to obtain the first traffic simulation result. The preset rule constraints include at least one of the following: constraints determined by the intelligent driver model (IDM) based on reference driving behavior; traffic operation constraints; traffic signal constraints; and route rule constraints.

4. The method according to claim 3, characterized in that, The transformer-based multi-agent motion generation network model is trained in the following manner: Obtain the real trajectory dataset of the traffic participants on the road network map; A first loss and a second loss are determined based on the real trajectory dataset and the second traffic simulation results. The first loss is the regression loss of the second traffic simulation results relative to the real trajectory dataset, and the second loss is the cross-entropy loss of the second traffic simulation results relative to the real trajectory dataset. Construct a loss function based on the first loss and the second loss; The transformer-based multi-agent motion generation network model is trained according to the loss function to obtain the trained transformer-based multi-agent motion generation network model.

5. The method according to claim 3 or 4, characterized in that, The generation time is within the first time range of the simulation cycle of the intelligent agent traffic participant behavior generation and evolution method for large-scale simulation, and the first time range is the time interval located at the beginning of the simulation cycle. The first traffic simulation result is obtained by inputting the stitching result, the running path, and the traffic participants' first traffic information within the neighborhood of the road network map into a transformer-based multi-agent motion generation network model for traffic simulation. The intermediate simulation results are then adjusted according to preset rule constraints to obtain the first traffic simulation result, including: Within the first time range, the stitching result, the running path, and the first traffic information of the traffic participants in the neighborhood of the road network map are input into a transformer-based multi-agent motion generation network model for traffic simulation, thereby obtaining multiple intermediate simulation results within the simulation period. According to preset rule constraints, the multiple intermediate simulation results within the simulation period are adjusted to obtain multiple first traffic simulation results within the simulation period. The step of adjusting the first traffic simulation result using traffic control information to obtain the second traffic simulation result includes: By using traffic control information, multiple first traffic simulation results within the simulation period are adjusted to obtain multiple second traffic simulation results within the simulation period.

6. The method according to any one of claims 1 to 4, characterized in that, The method further includes: Remove the traffic participants according to preset conditions; The preset conditions include at least one of the following: The location of the traffic participant extends beyond the range of the road network map; The location of the traffic participant reaches the end of the operating path; The simulated trajectory of the traffic participant exceeds the preset range compared to the actual route.

7. A device for generating and evolving the behavior of intelligent traffic participants for large-scale simulation, characterized in that, The device includes: The acquisition module is used to acquire a road network map, a route coordinate file of traffic participants, and a generation plan of the traffic participants. The route coordinate file is used to constrain the running path of the traffic participants on the road network map, and the generation plan is used to determine the generation time and running status of the traffic participants on the road network map at the generation time. The traffic participants are preset traffic participation objects that make traffic behavior decisions through intelligent agents. The first processing module is used to generate the traffic participants on the road network map at the generation time based on the route coordinate file and the generation plan, and to determine the running status of the traffic participants at the generation time; The second processing module is used to perform traffic simulation based on the personalized parameters of the traffic participants and the first traffic information of the traffic participants in the neighborhood range of the road network map, and obtain the first traffic simulation result. The personalized parameters are used to characterize the traffic characteristics of the traffic participants. The third processing module is used to adjust the first traffic simulation result using traffic control information to obtain the second traffic simulation result. The traffic control information includes at least one of the following: operational status control information of the traffic participants, disturbance control information, safety rule constraints, and traffic light status prediction information.

8. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 6.

Images