Data-driven simulation method and system for complex interaction behavior of mixed traffic flow

Abstract

The application belongs to the field of computer software and transportation, and particularly relates to a data-driven simulation method and system for complex interaction behavior of mixed traffic flow, and the method comprises the following steps: obtaining trajectory data of real motor vehicles, non-motor vehicles and pedestrians, and performing group division on the non-motor vehicles and pedestrians to extract interaction scenes; constructing a mixed flow strategy simulation model, taking each traffic participant as an agent, constructing a hierarchical strategy network based on an upper strategy respectively constructed for the three types of participants to guide lower actions, and generating simulated trajectories; further based on real interaction data and simulated generated data, respectively constructing observation-action pairing samples, and realizing integrated and efficient simulation of complex interaction behavior of mixed traffic flow through an adversarial training mechanism. The application realizes integrated simulation modeling of mixed traffic flow of multiple types of traffic participants, can fully depict group travel characteristics of non-motor vehicles and pedestrians, and accurately describe a dynamic decision-making process of heterogeneous traffic participants.

Description

TECHNICAL FIELD

[0001] The present application relates to the fields of computer software and transportation, and particularly relates to a data-driven simulation method and system for complex interaction behaviors of mixed traffic flow. BACKGROUND

[0002] In urban road traffic systems, the actual interaction behaviors of human traffic participants do not completely follow traffic regulations, but to a large extent rely on an implicit "social interaction rule" (hereinafter referred to as "social rule"). Such behaviors guided by social rules can be referred to as social interaction behaviors. Mixed traffic environment, i.e. the situation where motor vehicles, non-motor vehicles and pedestrians use road space together, exists widely in urban road networks. Such roads often lack physical isolation facilities and also often have no clear traffic signs and markings, resulting in frequent complex interactions between different traffic subjects. In particular, non-motor vehicles and pedestrians, as key participants in the traffic system, exhibit strong randomness, uncertainty and conformity, and are not subject to strict lane rules, and are prone to form group movement characteristics.

[0003] Fine simulation of such complex interaction environment not only has important value for traffic planning and management, but also is a key support for promoting the safe deployment of autonomous vehicles in actual road environment. Therefore, it has important theoretical significance and application value to construct an integrated strategy simulation model that can uniformly and realistically simulate the social interaction behaviors of motor vehicles, non-motor vehicles and pedestrians.

[0004] There are significant differences among motor vehicles, non-motor vehicles and pedestrians in terms of physical characteristics, decision-making patterns and consequences of interaction failures. From the physical dimension, motor vehicles have larger body size and heavier weight, with lower maneuverability and longer braking distance; non-motor vehicles and pedestrians have smaller body size and lighter weight, with higher maneuverability and flexibility, which makes their motion characteristics and interaction modes in space significantly different from those of motor vehicles. In terms of decision-making patterns, the behavior of motor vehicles is usually sequential, with less parallel behavior in the same space, and its decision-making process relies more on the prediction of the behavior of interacting objects. In contrast, the behavior of non-motor vehicles and pedestrians is obviously herd-like, and they are easy to gather into groups under certain conditions and may quickly disperse into individuals. This behavior characteristic makes the decision-making of non-motor vehicles and pedestrians not only affected by individual judgment, but also significantly affected by group behavior. The complex social interaction dynamics make the behavior modeling of non-motor vehicles and pedestrians more challenging. In terms of consequences of interaction failures, motor vehicles have relatively low injury degree due to the relatively solid body protection (unless the vehicle speed is very high); while non-motor vehicles and pedestrians may lead to a higher degree of risk due to the lack of external protection. Therefore, when modeling the behavior of motor vehicles, non-motor vehicles and pedestrians, the same interaction logic cannot be simply adopted, and a heterogeneous strategy simulation model capable of simulating the interaction behavior of the three types of traffic participants should be constructed to more realistically simulate complex mixed traffic flow.

[0005] Existing micro-traffic flow simulation software (such as VISSIM, AIMSUN and TESS NG, etc.) can realize the behavior simulation of mixed traffic flow. However, on the one hand, such software mostly adopts rule-based behavior modeling method, simplifies the interaction rules, and is difficult to fully depict the complex and changeable social interaction behavior in reality; on the other hand, the modeling of non-motor vehicle and pedestrian group behavior is still insufficient, and it is difficult to effectively reflect the complex and dynamic interaction relationship among individuals in the group. In recent years, data-driven simulation models have gradually attracted attention, and they show great potential in the modeling of mixed flow behavior. However, the existing modeling framework of mixed flow mostly follows the behavior model of motor vehicles, only adjusts some parameters, ignores the differences between non-motor vehicles and pedestrians in interaction logic, especially lacks in-depth discussion on group behavior depiction. Overall, the simulation modeling of interaction behavior in mixed traffic flow still faces the following challenges: first, the interaction mechanism of non-motor vehicles and pedestrians needs to be deeply analyzed, especially the mechanism of individuals in self-organizing groups; second, a unified modeling framework needs to be constructed to integrate the group decision-making mechanism of non-motor vehicles and pedestrians, so as to realize the accurate simulation of social interaction behavior of different types of traffic participants. SUMMARY

[0006] The present application aims to overcome the defects of the prior art and provide a data-driven simulation method for complex interaction behavior of mixed traffic flow, which can construct an integrated strategy simulation model capable of comprehensively depicting the social interaction behavior of multiple types of heterogeneous traffic participants.

[0007] The object of the present application can be achieved by the following technical solutions:

[0008] The data-driven simulation method for complex interaction behavior of mixed traffic flow comprises the following steps:

[0009] Step S1, input stage. Real vehicle, non-motor vehicle and pedestrian trajectory data are obtained, the behavior of non-motor vehicle and pedestrian groups is analyzed, the state and action of each time slice are determined, each traffic participant is taken as an agent, and the agent state and action space are determined.

[0010] Step S2, building a mixed flow strategy simulation model and training. Taking each traffic participant as an agent, a hierarchical strategy network is built for three types of participants to guide the lower layer action, and a simulation trajectory is generated based on the hierarchical strategy network: motor vehicles mainly make individual decisions, while non-motor vehicles and pedestrians make individual decisions under the influence of group cooperation / competition tendency;

[0011] Based on real interaction data and simulation generated data, observation-action paired samples are constructed, and through an adversarial training mechanism, an integrated and efficient simulation of complex interaction behavior of mixed traffic flow is realized.

[0012] Step S3, interaction behavior simulation evaluation. It includes trajectory accuracy evaluation and interaction strategy rationality evaluation.

[0013] The specific process of step S1 is as follows:

[0014] Step S101, interaction event extraction;

[0015] The basic data in the target interaction scene are collected, the position data are smoothed, and the speed and acceleration are recalculated based on the smoothed position data; based on the motion trajectory features and video analysis, the trajectory data and map data are spatially matched, the trajectory is time-aligned combined with the intersection signal control data, and the interaction scene is split and extracted through manual intervention; then the position translation and time truncation data enhancement method is used to expand the number of interaction scenes;

[0016] Step S102, group division;

[0017] The DBSCAN algorithm is used to divide the non-motor vehicles and pedestrians into groups, and the specific process is as follows:

[0018] First, the neighborhood radius and minimum density threshold minPts are determined based on domain knowledge;

[0019] Then, all sample points are traversed, and points in the neighborhood containing at least minPts samples are marked as core points;

[0020] Then, adjacent core points are merged by density reachability to form clusters, and boundary points (points located in the neighborhood of core points but not satisfying the core point conditions) are assigned to the corresponding clusters, while the remaining unclassified points are marked as noise.

[0021] Specifically, the mean of the inter-individual distance in the empirical data is used as the neighborhood radius, and the minimum density threshold minPts is set to 2 (i.e., at least two individuals are required to form a population).

[0022] Step S103: Determine the state and action space;

[0023] The observation space consists of its own state and the observation information of other objects. Its own state includes its own position, speed, distance from the destination, heading angle at the previous moment, speed at the previous moment, and the area occupied by the agent (if the agent is an individual, the area is the product of the vehicle's length and width; if the agent belongs to a group, the area is the group's area). The observation information of other objects includes the relative distance and speed difference with the interactive object, as well as the area of ​​the interactive object.

[0024] The action space of an agent is the rate of change of acceleration (excluding direction) and steering angle.

[0025] Specifically, in step S103,

[0026] Each traffic participant is an intelligent agent, and in each frame of observation information, at each time... intelligent agent observation vector Including its own driving status Information from observations of other objects :

[0027]

[0028] Its own driving status as follows:

[0029]

[0030] in, , Representing the intelligent agent in shaft and Position in the axial direction , Represents its shaft and Velocity in the axial direction; , They respectively represent the points at the endpoint and the destination. shaft and Distance along the axis , The positive or negative sign indicates the direction of movement; For intelligent agents in The heading angle at time (i.e., the angle between the agent and the target) (angle between axes) This represents the area occupied by the intelligent agent.

[0031] If the agent is an individual, the area The area is the product of the vehicle's length and width; if the intelligent agent belongs to a certain group, the area is... The area of ​​the group is the product of the maximum length and width of the geometric shape formed by all individuals in the group.

[0032] intelligent agent For other intelligent agents Observational information as follows:

[0033]

[0034] It consists of their relative distance and speed difference, as well as the area of ​​the interacting object. express and exist Relative distance along the axis, Indicates in Relative distance along the axis; They represent shaft and The velocity difference along the axis. The area of ​​the interactive object is calculated in the same way as... Consistency (if the agent) and intelligent agents Belonging to the same group, area The group area is the product of the maximum length and width of the geometry formed by all individuals within the group; if they belong to different groups, the length and width are the observed maximum length and width values ​​of the entire divided group. If the interaction object is an individual within the group, the parameters (including...) , , , ) represents the relative value of the agent's average parameter to the group's average parameter.

[0035] Given the limited field of view and the layout characteristics of the intersection, the agent cannot obtain the global environmental state, but can only obtain...

[0036] The information of the interactive objects within its visual field range is taken. The observation distance threshold is set for motor vehicles, non-motor vehicles and pedestrians respectively, and the 10 closest objects within the threshold are selected as the interactive objects. If there are non-motor vehicles or pedestrians in the interactive objects, it is judged whether they will form a group based on a clustering algorithm. If a group is formed, the group is regarded as the interactive object of the intelligent agent, and the kinematic characteristics of the group are the mean values of the characteristics of all individuals in the group.

[0037] The action space of the intelligent agent is modeled as:

[0038]

[0039] wherein, is the acceleration of the vehicle, is the change rate of the heading angle of each time step.

[0040] According to the observation space and the action space, each trajectory is converted into a series of observation-action pairs, represented as wherein, T is the total time step length of the vehicle trajectory, .

[0041] In step S2,

[0042] The hybrid flow policy simulation model comprises a generator and a discriminator, wherein the generator comprises a hierarchical policy network and a value estimator network, and the discriminator comprises a reward estimator and a potential function.

[0043] In the generator, a hierarchical policy network is constructed for three types of traffic participants to guide the lower layer action under the upper layer policy; an explicit closed-loop interaction process of “environment understanding-policy formation-action execution” is constructed in the generator to realize accurate interaction behavior modeling. Specifically as follows:

[0044] In the interactive environment, an individual dynamically adjusts its cooperation or competition tendency based on the understanding of the surrounding environment, and based on this tendency, a specific behavior direction (such as acceleration / deceleration, turning / rounding, etc.) is determined, and finally the corresponding action is executed to obtain the generated trajectory.

[0045] The value estimator network is constructed based on a multi-layer perceptron (MLP), and the main function is to estimate the value function of the state action.

[0046] When training the generator, the input is the observation data, the action data, the reward from the discriminator and the output of the value estimation network, and the output is the updated policy network parameters.

[0047] The hybrid flow policy simulation model is trained by using an adversarial training mechanism, and the process is as follows:

[0048] Based on the centralized training-decentralized execution (CTDE) framework, the discriminator network and the generator network are collaboratively trained in an alternating optimization manner. The discriminator network is committed to learning the feature distribution of real driving behaviors and effectively distinguishing real driving behaviors from algorithm-generated simulated behaviors through an adversarial training mechanism. At the same time, the generator network continuously optimizes its policy network based on the reward signal provided by the discriminator through the policy gradient method, thereby improving the rationality and authenticity of the action output.

[0049] As a preferred technical solution, the generator of the motor vehicle and the hierarchical generator construction method of the non-motor vehicle / pedestrian are as follows:

[0050] The generator of the motor vehicle adopting a hierarchical structure of "cooperation / competition tendency - behavior direction - action execution" is constructed. Specifically, it is divided into three layers:

[0051] The upper layer is based on the Transformer architecture, integrates spatio-temporal interaction information to generate cooperation / competition tendency, introduces the concept of Social Value Orientation (SVO) and constrains its output range within The closer to , the stronger the competition tendency; the closer to , the stronger the cooperation tendency.

[0052] The middle layer combines tendency and knowledge-driven methods to determine specific interaction behavior direction, which is the key connection between the upper layer and the lower layer, to ensure the consistency of tendency and action in the social interaction framework.

[0053] The lower layer learns specific actions based on the guidance of the upper layer and the middle layer. Special generator networks are constructed for motor vehicles of different driving directions to improve the relevance and accuracy of modeling.

[0054] The hierarchical generator of non-motor vehicles / pedestrians considering group behavior is constructed as follows:

[0055] The generator strategy network structure of non-motor vehicles and pedestrians follows the hierarchical architecture of motor vehicles, but there are differences in the interaction logic: motor vehicles mainly interact based on individual strategies, while non-motor vehicles and pedestrians are influenced by groups. For the case of group existence, individuals within the group will determine specific actions based on the similarity between their own behavior and the average behavior of the group under the guidance of the group cooperation / competition tendency. The final action is adjusted by the direction constraint module of the middle layer network, which combines domain knowledge to constrain the action space of the lower layer network, and constrains the acceleration range for different interaction tendencies to ensure a balance between group consistency and individual differences. Considering the differences in behavior characteristics and movement patterns between non-motor vehicles and pedestrians, special generator networks are constructed for non-motor vehicles and pedestrians.

[0056] Further, a herd tendency quantification mechanism is introduced for non-motorized vehicles and pedestrians, i.e., the final action of an individual is determined by evaluating the individual's herd degree to the group. The motion similarity between an individual and the whole group is measured by using the cosine similarity (Cosine Similarity, ) as the core measurement index. The higher the similarity, the greater the influence of the group on the individual, and the higher the herd degree, the more likely the individual will take the group action; vice versa. Specifically as follows:

[0057] The cosine value of the angle between the vectors is calculated to capture the similarity of the direction consistency and the relative amplitude. In the specific implementation, key motion features are selected to construct feature vectors, including . The individual feature vector is composed of the step motion state sequence in the group maintenance period, and the group feature vector is obtained by calculating the mean value of all individual feature vectors at each step. Since the features such as , , and change relatively stably, , and dominate the change of motion mode, the cosine similarity can effectively reflect the direction consistency and relative amplitude consistency of the motion features (such as ratio and ratio). The cosine similarity ranges from . For easy comparison, the similarity result is normalized to map its value range to the interval. The closer the value is to 1, the higher the similarity between the individual and the group.

[0058] The specific similarity calculation formula is as follows:

[0059]

[0060] wherein, and are vectors containing dimensions, denotes the dot product, denotes the length of the vector.

[0061] As a preferred technical solution, each intelligent agent includes a binary classification discriminator. With the observation-action pair as input, it outputs the probability estimate derived from the real expert demonstration, which is used to distinguish whether the given input is an observation-action pair generated by the generator or a real observation-action pair. The discriminator includes a reward estimator and a potential function, both of which are double-layer MLP networks, and the reward estimator provides an optimization signal for the generator.

[0062] Step S3 is specifically as follows:

[0063] Interaction strategy rationality evaluation: evaluate from two aspects of overall decision accuracy (through conflict point order) and cooperation / competition tendency dynamic evolution rationality;

[0064] Trajectory accuracy evaluation: evaluate the similarity between the generated trajectory and the real trajectory at the micro level by calculating the average distance error of the position, and evaluate the similarity of the speed, heading angle and the average value distribution similarity between the generated trajectory and all interacting objects at the macro level.

[0065] Another aspect of the present application provides a data-driven simulation system for complex interaction behavior of mixed traffic flow, which is used to realize the aforementioned data-driven simulation method for complex interaction behavior of mixed traffic flow, and the simulation system comprises a model construction module and an application simulation module.

[0066] The model construction module comprises:

[0067] An expert data acquisition unit is responsible for defining the observation and action space in multi-agent imitation learning, acquiring real motor vehicle, non-motor vehicle and pedestrian traffic trajectory data, and constructing a feature matrix to support subsequent model training.

[0068] A strategy simulation training unit trains an agent to reproduce expert behavior logic based on a multi-agent imitation learning framework, and constructs an integrated strategy simulation model for mixed traffic flow.

[0069] A strategy simulation evaluation unit constructs a multi-dimensional evaluation system by combining strategy matching degree and trajectory accuracy to test the effectiveness and reliability of the simulation model in real interaction tasks.

[0070] The application simulation module comprises:

[0071] A scene initialization unit provides the initial state and background environment data of the mixed traffic flow required for generating a scene.

[0072] A dynamic scene generation unit generates diversified and high-complexity interaction scenes based on the trained strategy simulation model.

[0073] Advantages

[0074] (1) Integrated simulation modeling of mixed traffic flow multi-type traffic participants: the present application develops a data-driven mixed flow strategy simulation model based on a multi-agent adversarial inverse reinforcement learning framework, which can construct different interaction strategies for different vehicle types, and provides an integrated framework for mixed traffic flow modeling.

[0075] (2) Fully depict the group travel characteristics of non-motor vehicles and pedestrians: The application constructs an adaptive group division method through the DBSCAN algorithm to accurately represent group travel phenomena, and further measures the similarity of individual behavior and group behavior within the group based on cosine similarity, revealing the dynamic mechanism of individual decision-making in the group, and providing a theoretical basis for strategy simulation modeling.

[0076] (3) Accurately describe the dynamic decision-making process of heterogeneous traffic participants: The application constructs a hierarchical generator structure that can accurately describe the dynamic decision-making process of heterogeneous traffic participants: the upper layer network is the cooperation / competition tendency layer, the middle layer network is the behavior direction layer, and the lower layer network is the action execution layer, which can capture the behavior characteristics of individuals and groups, and accurately model the dynamic evolution process of cooperation / competition tendency and motion state of different traffic participants. BRIEF DESCRIPTION OF DRAWINGS

[0077] Figure 1 is the process flow chart for the method of example 1;

[0078] Figure 2 is the schematic diagram of the method framework of example 1;

[0079] Figure 3 is the schematic diagram of the mixed traffic flow interaction event extraction process in example 1;

[0080] Figure 4 is the schematic diagram of the motor vehicle generator structure in example 1;

[0081] Figure 5 is the schematic diagram of the non-motor vehicle / pedestrian generator structure in example 1;

[0082] Figure 6 is the trajectory comparison diagram of the typical interaction scene of mixed traffic flow in example 1;

[0083] Figure 7 is the analysis diagram of the cooperation / competition tendency of motor vehicle-non-motor vehicle interaction in the typical scene in example 1;

[0084] Figure 8 is the schematic diagram of the data-driven mixed traffic flow complex interaction behavior strategy simulation system in example 2;

[0085] Figure 9 is the process flow diagram of using the simulation system for simulation in example 2. DETAILED DESCRIPTION

[0086] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts should fall within the protection scope of the present application.

[0087] Embodiment 1

[0088] To solve the problems in the prior art, the embodiment provides a data-driven simulation method for complex interaction behaviors of mixed traffic flow. A mixed flow strategy simulation model MTSIM (Mixed Traffic Simulation) for complex interaction behaviors of motor vehicles, non-motor vehicles and pedestrians is developed based on multi-agent adversarial inverse reinforcement learning (MA-AIRL). The MTSIM model is composed of generators and discriminators for three types of traffic participants based on group division and individual interaction mechanism analysis within the group.

[0089] To comprehensively depict the social interaction behaviors of multiple types of traffic participants, the group travel mechanism of non-motor vehicles and pedestrians is analyzed in depth, a heterogeneous strategy network with a hierarchical structure is constructed, and a multi-dimensional evaluation system is developed to evaluate the effectiveness of the model in capturing strategies and generating trajectories.

[0090] Through training on real-world interaction trajectory data, the model can generate two main outputs: a strategy estimator and an action estimator.

[0091] The strategy estimator includes an upper layer network and a middle layer network, reflecting the cooperation or competition tendency of the subject in a specific situation and the specific behavior direction.

[0092] The action estimator is a lower layer network responsible for generating specific actions and updating two-dimensional motion states. For agents that make up a group, the similarity of individuals and groups is considered to simulate the cooperation / competition tendency and motion state of different types of traffic participants in the mixed traffic flow interaction environment. Finally, the interaction strategy and vehicle trajectory of different groups of individuals over time are obtained.

[0093] Figure 1 The method processing flowchart of the embodiment is as follows, Figure 2 The method framework schematic diagram of the embodiment mainly includes input, training, output and simulation stages. Taking the interaction behaviors of left-turn motor vehicles and opposite straight motor vehicles, non-motor vehicles and pedestrians at an unprotected left-turn phase intersection as an example for illustration.

[0094] The data-driven simulation method for complex interaction behaviors of mixed traffic flow includes the following steps:

[0095] Step S1, input stage. Real vehicle, non-motor vehicle, pedestrian trajectory data is obtained, non-motor vehicle and pedestrian group behavior is analyzed, the state and action of each time slice are determined, each traffic participant is taken as an agent, and the agent state and action space are determined.

[0096] Step S1 specifically includes steps S101-S103:

[0097] Step S101, interaction event extraction.

[0098] The basic data in the target interaction scene is collected, the Savitzky-Golay (S-G) filter is used to smooth the position data, abnormal jitter is eliminated, and the speed and acceleration are recalculated based on the smoothed position data to ensure the accuracy of the motion parameters; based on the motion trajectory features and video analysis, the trajectory data and map data are spatially matched, the trajectory is time-aligned combined with the intersection signal control data, and the interaction scene is split and extracted through manual intervention, and then the position translation and time truncation data enhancement methods are used to expand the number of interaction scenes.

[0099] In this embodiment, the interaction event data is obtained through the independently collected Shanghai Xianxiajianhe intersection dataset (hereinafter referred to as the XXJH dataset), which includes intersection trajectory data, signal control data and map data, covers a variety of vehicle, non-motor vehicle and pedestrian interaction scenes, and the time resolution of the trajectory data is 0.12 s.

[0100] Taking the left-turn vehicle flow and the straight-ahead vehicle flow, non-motor vehicle flow and pedestrian flow as the research object, the data of the period in which the interaction behavior exists is extracted as the interaction event. The interaction event extraction process is as follows Figure 3The trajectory data and map data are first spatially matched after smoothing, and the left-turn vehicle trajectories of the west approach and the straight vehicle, non-motor vehicle and pedestrian trajectories of the east approach are screened. Then, combined with the intersection signal control data, the left-turn and straight trajectories are time-aligned according to the same timing cycle, and the trajectory data within the green light period is intercepted to ensure that the interaction behavior occurs within the effective signal phase. Considering that vehicles do not exist at every moment within a complete cycle, the interaction scenarios are reasonably split through manual intervention, that is, scenarios with obvious interaction behavior and no unreasonable behavior are selected, which ensures the integrity of the interaction subject trajectory while reducing the proportion of invalid time period. Finally, to ensure the continuity and analyzability of the interaction behavior, the total time of the scenario is further screened, and only scenarios with interaction time ≥ 2 s are retained to improve data quality. A total of 90 effective interaction scenarios are finally extracted, and data augmentation technology is used to expand to 630 scenarios, of which 540 scenarios are used for subsequent model training, and 90 scenarios are used for reasoning and verification. All scenarios are used for group behavior analysis together. It should be noted that this dataset is only a typical example, and similar datasets of other mixed-flow intersections can also be easily applied.

[0101] Step S102, group division. The DBSCAN algorithm without pre-specifying the number of clusters is adopted. This method can effectively identify clusters of any shape and accurately distinguish noise points based on density reachability principle. The neighborhood radius and the minimum density threshold minPts are determined based on domain knowledge; then all sample points are traversed, and points containing at least minPts samples in the neighborhood are marked as core points; then adjacent core points are merged to form clusters through density reachability, and boundary points are assigned to the corresponding cluster, and the remaining unclassified points are marked as noise.

[0102] In this embodiment, the mean value of individual distance is selected as the neighborhood radius, wherein the neighborhood radius of the non-motor vehicle group is set to 5.8 m, the neighborhood radius of the pedestrian is set to 3.6 m, and the minimum density threshold minPts is set to 2 (i.e. at least two individuals are required to form a group). In group division, the types of traffic participants are distinguished, non-motor vehicles only form groups with non-motor vehicles, pedestrians only form groups with pedestrians, and group division is performed based on location data at each moment. The parameter setting of this method has clear physical meaning and can dynamically adapt to changes in location data at different moments.

[0103] Step S103, state and action space determination.

[0104] The state and action space of the agent are determined by taking each traffic participant as an agent.

[0105] The observation space of the agent is composed of its own state and the observation information of other objects, the own state including its own position, speed, distance to the end point, heading angle at the last time, speed at the last time, area occupied by the agent (if the agent is an individual, the area is the product of the length and width of the vehicle body; if the agent belongs to a certain group, the area is the area of the group), and the observation information of other objects including the relative distance, speed difference and area of the interactive object;

[0106] The action space of the agent is the acceleration (without direction) and the change rate of the steering angle.

[0107] According to the observation space and the action space, each trajectory can be converted into a series of observation-action pairs.

[0108] Step S2, training of the hybrid flow strategy simulation model.

[0109] Specifically, steps S201-S203 are included.

[0110] Step S201, obtaining training data, i.e., the XXJH dataset interaction data obtained in step S1.

[0111] Step S202, generating an agent trajectory based on the pre-constructed generator.

[0112] Based on the analysis of group division and individual interaction mechanism within the group, a hybrid flow strategy simulation model is constructed, including generators and discriminators for three types of traffic participants.

[0113] The generator includes a policy network and a value estimator network. A hierarchical policy network is constructed for the three types of participants, with the upper layer guiding the lower layer action. The value estimator network is constructed based on a multi-layer perceptron (MLP). Each value estimator network is a four-layer MLP network.

[0114] The value function is used to estimate the expected total return of the policy network under the condition that the state of the agent is and other agents take actions , where the subscript i represents all agents except i . Further, the output of the value estimator will be used to calculate the advantage function A ( o, a ), which measures the additional return of taking action a compared to taking the average action according to the current policy. The advantage function will be used for parameter update in reinforcement learning. The advantage function is calculated as follows:

[0115]

[0116] in,

[0117]

[0118] in, Let i be the advantage function of agent i. Let i be the policy function of agent i. For value function network parameters, It is a discount factor. It is an intelligent agent i The reward signal k This refers to considering the time elapsed from the current moment when calculating the advantage function. k Long-term rewards after each time step. For intelligent agents i The observed function, To accumulate returns.

[0119] The policy network for the motor vehicle generator is divided into three layers, such as... Figure 4 As shown:

[0120] The upper layer, based on the Transformer architecture, integrates spatiotemporal interaction information to generate cooperation / competition tendencies. It introduces the concept of Social Value Orientation (SVO) to quantify the agent's cooperative or competitive tendencies. The network input consists of time-series observation data containing historical information, and the network output is considered as the tendency at each moment. The output range is constrained within... Within the range, the closer to This indicates a stronger competitive tendency;

[0121] The middle layer combines tendency and knowledge-driven methods to determine specific interaction behavior directions, serving as a key connection between the upper and lower layers to ensure consistency between tendencies and actions within the social interaction framework. Based on the cooperative / competitive tendencies learned by the upper layer, domain knowledge is used to constrain the action range of the lower layer. Specific constraints are shown in Table 1, where acc is the magnitude of the vehicle's acceleration. It is the rate of change of heading angle at each time step;

[0122] The lower level learns specific actions based on the guidance from the upper and middle levels.

[0123] Table 1. Motor Vehicle Mid-Level Behavioral Direction Constraints

[0124]

[0125] The generator strategy network structure for non-motorized vehicles and pedestrians follows the hierarchical architecture of motorized vehicles, but differs in its interaction logic, such as... Figure 5As shown: for the case of the existence of the group, the individuals in the group will determine the specific action according to the similarity between their own behavior and the average behavior of the group under the guidance of the group cooperation / competition tendency, and the interaction logic can be regarded as that the individuals execute specific actions based on their own characteristics (similarity) under the guidance of the group strategy. Considering that there are also differences between non-motor vehicles and pedestrians in behavior characteristics and motion patterns, special generator networks are constructed for non-motor vehicles and pedestrians respectively. The upper network takes the observations of all individuals in the group at the current time and historical time as the input of the upper network, and outputs the group cooperation / competition tendency at each time, wherein the observation data of the historical time of the group only contains the time period during which the individuals in the group remain unchanged. The middle network combines the field knowledge to constrain the action space of the lower network based on the cooperation / competition tendency learned by the upper layer, to ensure that the individual behavior balances between group consistency and individual difference, and the specific constraint rules are shown in Table 2. The interaction tendency is divided into three intervals, which correspond to different acceleration constraints. The lower network obtains specific actions based on the upper cooperation / competition tendency and the constraint of the middle behavior direction, and in the design of the lower network of non-motor vehicles and pedestrians, a conformity tendency quantization mechanism is introduced to determine the final executed action by evaluating the conformity degree of the individual to the group.

[0126] Table 2 Non-motor vehicle / pedestrian middle layer behavior direction constraint

[0127]

[0128] Among them, the motion similarity between the individuals in the group and the group as a whole adopts Cosine Similarity as the core measurement index, which captures the similarity of direction consistency and relative amplitude by calculating the cosine value of the angle between vectors.

[0129] Select key motion features to construct feature vectors, including Among them, the individual feature vector is composed of the m-step motion state sequence in the group maintenance period, and the group feature vector is obtained by calculating the mean value of all individual feature vectors at each step. The cosine similarity range is [-1, 1], for easy comparison, the similarity result is normalized, and its value range is uniformly mapped to the interval [0, 1], the closer the value is to 1, the higher the similarity between the individual and the group. The specific similarity SI The calculation formula is as follows:

[0130]

[0131] Among them, A and B are vectors containing n dimensions, denotes the dot product, denotes the length of the vector.

[0132] The higher the similarity of the behavior characteristics of the individual and the group during the existence of the group, the more likely the individual's action at the current time is similar to the overall action of the group; on the contrary, when the similarity is low, the individual's action is more likely to deviate from the group as a whole. Therefore, in the modeling of interactive behavior, the group can be regarded as a whole, the overall cooperation / competition tendency is first learned, then the similarity of the individual and the group historical behavior characteristics is measured, and the action of the individual is adjusted based on the similarity. The interaction logic can be regarded as that the individual performs specific actions based on its own characteristics (similarity) under the guidance of the group strategy. On this basis, an action correction coefficient is introduced to dynamically adjust the standard deviation of the group action. The higher the behavior similarity of the individual and the group, the smaller the adjustment of the standard deviation, indicating that the individual is more inclined to follow the action under the guidance of the group strategy; on the contrary, the greater the difference between the individual and the group, the greater the adjustment of the standard deviation, indicating that the individual is less subordinate to the group, and its action is different from the overall action of the group.

[0133] Step S203, model training. Based on the centralized training-distributed execution (CTDE) framework, the discriminator network and the generator network are collaboratively trained by alternating optimization.

[0134] Training the discriminator: the input is the observation-action pair sampled from the generated data and the expert data and ; the reward estimator parameters , the potential function parameters are updated using Adam Optimizer , the following target is maximized:

[0135]

[0136] wherein, is the reward function, including the reward estimator and the potential function , both of which are double-layer MLP networks. The reward estimator provides an optimization signal for the generator. is an adaptive sampler with parameters , indicating the probability of the generator selecting action under the given observation . ,

[0137] Training the generator: the input is the observation data, the action data, the reward from the discriminator, and the value estimator network output; the generator network parameters are updated using the MACK algorithm, and the target function is as follows:

[0138]

[0139] By repeatedly optimizing the parameters in the generator and discriminator for different observation-action pairing samples, the learned policy network and reward function are obtained, which serve as the output of the entire training algorithm.

[0140] Step S3, Interactive Behavior Simulation Evaluation. Specifically, this step includes steps S301-S302.

[0141] Step S301, trajectory accuracy evaluation. The similarity between the generated trajectory and the real trajectory is evaluated from two levels: micro-trajectory accuracy and macro-feature distribution authenticity.

[0142] At the micro-trajectory level, the average distance error of the position is used ( This metric is used to quantify accuracy.

[0143] Calculate the positional deviation between the generated trajectory and the real trajectory at each moment, for a single scene. The calculation method is as follows:

[0144]

[0145] in, 、 yes x-y Intelligent agents in coordinate system exist Real-time location 、 yes x-y Intelligent agents in coordinate system exist The position where the simulation model is generated at any given time. It is the number of intelligent agents. This is the total step length in a traffic scenario. It is based on the average of all test scenarios. This is used to comprehensively measure the simulation accuracy of the model at the micro-kinematic level; the lower the value, the higher the trajectory matching degree.

[0146] At the macro-feature distribution level, the similarity between generated and expert scenarios in key feature distribution is analyzed from a statistical perspective. Specifically, this involves... divergence As evaluation indicators, characteristics include speed ( ), heading angle ( ) and average ( ).in, For an individual and all the objects they interact with The mean. The closer the divergence is to 0, the more similar the two distributions are, and the better the simulation effect of the model on macroscopic features.

[0147] The model of the application (MTSIM model) and each benchmark model are evaluated on the test data set without distinguishing between motor vehicles, non-motor vehicles and pedestrians, and the specific results are shown in Table 3. The imitation learning model selects the Social-MAIL model and the MA-GAIL model. The Social-MAIL model does not consider the group travel characteristics of non-motor vehicles and pedestrians, and regards all agents and interactive objects as individuals. The basic MA-GAIL adopts a standard adversarial imitation learning framework and is not optimized for mixed traffic flow. The MTSIM-GAIL introduces the group behavior modeling method and hierarchical strategy network proposed by the application based on the MA-GAIL, to explore whether it can adapt to the intersection mixed traffic flow simulation modeling. The reinforcement learning model selects the MACK model. The model combines the policy gradient method and value function learning, and performs well in processing large-scale continuous action space and high-dimensional state space. The deep learning model selects the Social-LSTM model. The model does not consider the group behavior of non-motor vehicles and pedestrians, and regards all agents and interactive objects as individuals, and predicts the trajectory based on the observation data. The mechanism-driven model selects the classical one-dimensional simulation model IDM, which assumes that all vehicles strictly travel along the ideal lane center line and does not consider lateral movement or complex interaction.

[0148] Overall, the MTSIM model is significantly higher than other benchmark models in position accuracy, while 、 and The key traffic flow feature distribution is also very similar to the expert scene. It is fully proved that the model can reproduce the real characteristics of the mixed traffic flow at the intersection.

[0149] Table 3 Comparison results of MTSIM model and benchmark model

[0150]

[0151] Step S302, interaction strategy rationality evaluation. The evaluation is performed from two dimensions of overall decision accuracy and cooperation / competition tendency dynamic evolution rationality.

[0152] The overall decision accuracy is used to measure the authenticity of the final interaction result, and the evaluation method is to compare whether the sequence of traffic participants passing through the conflict point in the generated and real scenes is consistent.

[0153] In this embodiment, the overall decision-making accuracy of the model reaches 87%, demonstrating the effectiveness of the MTSIM model in capturing basic interactive behaviors. For motor vehicles and non-motorized vehicles and pedestrians not belonging to a group, the model learns its own cooperation / competition tendencies; for non-motorized vehicles and pedestrians belonging to a group, it learns the overall tendency of the group. This embodiment selects a typical scenario to analyze in detail the dynamic decision-making process of mixed-flow interactive behaviors. The scenario simultaneously includes left-turning motor vehicles, straight-going motor vehicles, individual / group non-motorized vehicles, and individual / group pedestrians. By plotting the trajectories of traffic participants in the generated and expert scenarios, and by analyzing the positions, cooperation / competition tendencies, similarity of individuals within a group, and scatter plots of actions of traffic participants at key moments in the generated scenario, the rationality of the dynamic evolution of tendencies is evaluated. Figure 6 , 7 As shown.

[0154] like Figure 6 As shown, this scenario involves a complex interaction process involving two left-turning motor vehicles (L0, L1), two straight-going motor vehicles (S0, S1), four non-motorized vehicles (B0, B1, B2, B3), and two pedestrians (P0, P1). Trajectory comparison reveals a high degree of consistency between the generated trajectory (blue) and the actual trajectory (green), with all traffic participants showing an average... The value was 2.94 m, verifying the accuracy of the model in trajectory generation.

[0155] Figure 7 This analysis examines the cooperative / competitive tendencies between motor vehicles and non-motor vehicles in typical scenarios, reflecting changes in interaction patterns over time. In the initial stage ( =3.0s), the left-turning motor vehicle L0 mainly interacts with the non-motorized vehicle B0. At this time, L0 adopts a clear cooperative strategy, actively decelerating during the turn; while B0 shows a cooperative tendency, the degree is low; at the same time, B1 has not yet formed a strong interaction with the vehicle, so it adopts a weak cooperative tendency, tentatively accelerating forward.

[0156] As time progressed to At 5.3 seconds, the interaction pattern changes significantly. At this point, the straight-going vehicle S0 has entered the intersection, and L0 has completed its interaction with B0 (B0 passed the conflict point first). The focus shifts to S0 and B1. Notably, both L0 and S0 show a cooperative tendency, adopting a slowing-down probing strategy, but S0's deceleration is consistently less than L0's. This phenomenon may stem from the traffic rule constraint of yielding to straight-going vehicles when turning left. Meanwhile, B1's cooperative tendency weakens further during this period, even intensifying its acceleration behavior, with its speed continuously increasing.

[0157] exist At 7.3s, the interaction pattern evolved again. L0 had completed its interaction with S0 (S0 passed the conflict point first), and its attention shifted to the newly entered non-motorized vehicles B2 and B3. At this point, B2 and B3 had formed a clear non-motorized vehicle group. L0 and this group both adopted a cooperative strategy, but the degree of cooperation shown by the group was relatively low. This phenomenon is consistent with the expectation that non-motorized vehicles perceive increased safety within a group. B2 and B3 showed a high degree of similarity to the overall behavior of the group, so their actions were almost identical, with both performing deceleration operations, but to a lesser degree than motorized vehicles. At the same time, S0 began to interact with the newly entered left-turning vehicle L1. Facing L1, S0 showed a tendency to cooperate, but did not significantly decelerate. This strategy likely stemmed from its assessment of the current traffic situation and a comprehensive consideration of historical interaction experience: on the one hand, S0 was close to the intersection exit, and continuing to pass would not substantially interfere with L1's turning; on the other hand, S0 had shown a cooperative tendency multiple times in previous interactions, and tended to maintain its pace of passage in the absence of conflict. Therefore, its choice to maintain speed within the safety boundary reflects a limited cooperative strategy that dynamically balances safety and efficiency. Correspondingly, L1 also exhibits a cooperative tendency when facing S0, which is approaching the intersection exit, decelerating more than S0, reflecting its proactive response to potential conflict risks. Furthermore, B1, located to the right of S0, does not have a direct interaction with L1. Based on the current situation, B1's strategy exhibits a competitive tendency, accelerating forward, reflecting its preference for prioritizing its own passage objective when there is no potential interference.

[0158] exist At 10.0s, L0 moves closer to B2 and B3, at which point the overall cooperative tendency of the group weakens compared to before. Specifically, B2 has already passed the conflict point, while B3 is still in a state of not having passed. This asynchronous state prompts L0 to strengthen its cooperative tendency, choosing to slow down and wait for the group to complete its passage, reflecting its consideration for the non-motorized vehicle group as a whole. It is worth noting that B2 and B3 maintain a highly consistent behavior pattern throughout the process, with their trajectory similarity and speed matching degree remaining at a high level, fully demonstrating the cooperative behavior characteristics of the non-motorized vehicle group. At the same time, S0 has completed its interaction with L1 and then accelerates away from the intersection area. L1's interaction objects then become the newly entered straight-going vehicle S1 and the non-motorized vehicle group. Faced with the simultaneous presence of heterogeneous traffic participants, L1's cooperative tendency is relatively weaker. At 7.3s, the effect was further enhanced. S1, recognizing that L1 had acquired the right-of-way, adopted a more explicit cooperative strategy, actively slowing down and yielding to ensure intersection safety. This series of changes in interaction patterns not only reflects the degree to which traffic participants comply with right-of-way rules, but also reveals the continuous adjustment process of strategies in multi-vehicle interaction environments, influenced by changes in the behavior of other participants and historical interaction experiences.

[0159] In general, the MTSIM model can effectively model the social interaction behavior of different types of traffic participants in mixed traffic flow, dynamically depict the continuous evolution process of cooperation / competition tendency, finely reflect the group characteristics of non-motor vehicles and pedestrians, embody the understanding and response ability of individuals to traffic situations, and successfully learn diversified interaction behavior patterns.

[0160] Embodiment 2

[0161] Based on the foregoing embodiments, referring to Figure 8 The embodiment provides a data-driven simulation system for complex interaction behavior of mixed traffic flow, which is used for implementing the data-driven simulation method for complex interaction behavior of mixed traffic flow in the embodiment 1. The simulation system comprises a model construction module and an application simulation module.

[0162] The simulation system comprises a model construction module and an application simulation module.

[0163] The model construction module comprises:

[0164] An expert data acquisition unit is responsible for determining the observation and action space in multi-agent imitation learning, acquiring real motor vehicle, non-motor vehicle and pedestrian traffic trajectory data, and constructing a feature matrix to support subsequent model training.

[0165] A strategy simulation training unit trains an agent to reproduce expert behavior logic based on a multi-agent imitation learning framework, and constructs an integrated strategy simulation model of mixed traffic flow.

[0166] A strategy simulation evaluation unit constructs a multi-dimensional evaluation system by fusing strategy matching degree and trajectory accuracy, and tests the effectiveness and reliability of the simulation model in real interaction tasks.

[0167] The application simulation module comprises:

[0168] A scene initialization unit provides initial state and background environment data of mixed traffic flow required for generating a scene.

[0169] A dynamic scene generation unit generates diversified and high-complexity interaction scenes based on the trained strategy simulation model.

[0170] Referring to Figure 9 For the flowchart of the simulation process using the established simulation system, first, the model construction module is used to determine the observation and action space of imitation learning, acquire and process expert data, and construct a multi-agent imitation learning framework. The mixed traffic flow integrated strategy simulation model is trained, a multi-dimensional evaluation system is constructed, and the effect of the strategy simulation model is evaluated. Finally, the application simulation module is used to generate complex and diverse strong interaction scenes based on the obtained simulation model.

[0171] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A data-driven simulation method for complex interaction behavior of mixed traffic flow, characterized in that, Comprise the following steps: Step S1, input stage; Real motor vehicle, non-motor vehicle, pedestrian trajectory data is obtained, the behavior of non-motor vehicle and pedestrian group is analyzed, the state and action of each time slice are determined, each traffic participant is taken as an agent, the state and action space of the agent are determined; Step S2, build hybrid flow strategy simulation model and train; each traffic participant is taken as an agent, a hierarchical strategy network is respectively built for the three types of participants, the upper layer network of the hierarchical strategy network is the cooperation / competition tendency layer, the middle layer network is the behavior direction layer, and the lower layer network is the action execution layer; based on the hierarchical strategy network, the simulation trajectory is generated: the motor vehicle is mainly individual decision, while the non-motor vehicle and the pedestrian make individual decision under the influence of group cooperation / competition tendency; Based on real interaction data and simulation data, observation-action paired samples are respectively constructed, and through the adversarial training mechanism, the integration and efficient simulation of complex interaction behavior of hybrid traffic flow is realized; Step S3, interaction behavior simulation evaluation, including trajectory accuracy evaluation and interaction strategy rationality evaluation.

2. The data-driven simulation method for mixed traffic flow complex interaction behavior according to claim 1, characterized in that, The specific process of step S1 is as follows: Step S101, interaction event extraction; Collect the basic data in the target interaction scene, smooth the position data, and recalculate the speed and acceleration based on the smoothed position data; Based on the motion trajectory characteristics and video analysis, the trajectory data and map data are spatially matched, the trajectory is time-aligned combined with the intersection signal control data, and the interaction scene is split and extracted through manual intervention; then the position translation and time truncation data enhancement method is used to expand the number of interaction scenes; Step S102, group division; The DBSCAN algorithm is used to divide the non-motor vehicle and pedestrian groups, and the specific process is as follows: First, determine the neighborhood radius and minimum density threshold minPts based on domain knowledge; Then, traverse all sample points, and mark the points containing at least minPts samples in the neighborhood as core points; Then, adjacent core points are merged through density reachability to form clusters, and boundary points are classified into corresponding clusters, and the remaining unclassified points are marked as noise; Step S103, determine the state and action space; The observation space is composed of the state of itself and the observation information of other objects, including the position, speed, distance to the end point, heading angle at the last time, speed at the last time, and the area occupied by the agent, and the observation information of other objects includes the relative distance, speed difference and area of the interactive object; The action space of the agent is the change rate of acceleration and steering angle.

3. The data-driven simulation method for mixed traffic flow complex interaction behavior according to claim 2, characterized in that, In step S103, Each traffic participant is an agent, and the observation information of each frame is Agent The observation vector of the agent includes the driving state of itself and the observation information of other objects : The driving state of the vehicle itself As follows: in, , Representing the intelligent agent in shaft and Position in the axial direction , Represents its shaft and Velocity in the axial direction; , They respectively represent the points at the endpoint and the destination. shaft and Distance along the axis , The positive or negative sign indicates the direction of movement; For intelligent agents in The heading angle at any given moment is the angle between the agent and the target. The included angle of the axis; Represents the area occupied by the intelligent agent; If the agent is an individual, the area is the product of the length and width of the body; if the agent belongs to a group, the area is the area of the group, which is the product of the maximum length and width of the geometric body composed of all the individuals in the group; Intelligent agent Observation information of other intelligent agents As follows: consists of its relative distance and velocity difference in the axis direction and the area of the interactive object, wherein, represents and in the axis direction, the relative distance in the axis direction, represents the relative distance in the axis direction; respectively represent the velocity difference in the axis direction and the axis direction; the axis and the axis direction; is the area of the interactive object, which is calculated in the same way as ; if the interactive object is an individual in the group, the parameter is the relative value of the individual and the average parameter of the group; The observation distance threshold is set for motor vehicles, non-motor vehicles and pedestrians, and the nearest 10 individuals within the threshold are selected as interactive objects; if there are non-motor vehicles or pedestrians in the interactive objects, determine whether they will form a group based on the clustering algorithm; if a group is formed, the group is regarded as the interactive object of the agent, and the group kinematics characteristics are the mean value of all individual characteristics in the group; The action space of the agent is modeled as: wherein, is the magnitude of the acceleration of the vehicle, is the rate of change of the heading angle for each time step; According to the observation space and the action space, each trajectory is converted into a series of observation-action pairs, denoted as wherein T is the total time step of the vehicle trajectory, .

4. The data-driven simulation method for mixed traffic flow complex interaction behavior according to claim 1, characterized in that, In step S2, The mixed flow strategy simulation model comprises a generator and a discriminator, wherein the generator comprises a hierarchical strategy network and a value estimator network, and the discriminator comprises a reward estimator and a potential function; In the generator, the hierarchical strategy network is constructed for three types of traffic participants; an explicit closed-loop interaction process of "environment understanding-strategy formation-action execution" is constructed in the generator to achieve accurate interaction behavior modeling; specifically as follows: in the interaction environment, an individual dynamically adjusts its cooperation or competition tendency based on its understanding of the surrounding environment, and determines a specific behavior direction based on the tendency, and finally executes the corresponding action; The value estimator network is constructed based on a multi-layer perceptron (MLP), and the role is to estimate the value function of the state action; during training of the generator, the input is observation data, action data, reward from the discriminator, and output of the value estimation network, and the output is the updated strategy network parameters; The mixed flow strategy simulation model is trained using an adversarial training mechanism, and the process is as follows: Based on the centralized training-distributed execution (CTDE) framework, the discriminator network and the generator network are collaboratively trained through alternating optimization; the discriminator network is committed to learning the feature distribution of real driving behavior, and effectively distinguishes real driving behavior from simulated behavior generated by the algorithm through the adversarial training mechanism; at the same time, the generator network optimizes its strategy network based on the reward signal provided by the discriminator through the policy gradient method, thereby improving the rationality and authenticity of the action output.

5. The data-driven simulation method for mixed traffic flow complex interaction behavior according to claim 4, characterized in that, The construction methods of the generator of motor vehicles and the hierarchical generator of non-motor vehicles and pedestrians are as follows: The generator of motor vehicles adopts a hierarchical structure of "cooperation / competition tendency-behavior direction-action execution", and is specifically divided into three layers: The upper layer is based on the Transformer architecture, integrates spatio-temporal interaction information to generate cooperation / competition tendency, introduces the concept of Social Value Orientation and restricts its output range to ; the closer to , the stronger the competition tendency; the closer to , the stronger the cooperation tendency; The middle layer combines the tendency and knowledge-driven method to determine the specific interaction behavior direction, which is the key connection between the upper layer and the lower layer, to ensure the consistency of the tendency and the action in the social interaction framework; The lower layer learns specific actions based on the guidance of the upper layer and the middle layer, and a dedicated generator network is constructed for motor vehicles in different driving directions to improve the pertinence and accuracy of modeling; The hierarchical generator of non-motor vehicles and pedestrians considering group behavior is constructed as follows: The strategy network structure of the generator of non-motor vehicles and pedestrians follows the hierarchical architecture of motor vehicles, but there are differences in the interaction logic: motor vehicles mainly interact based on individual strategy, while non-motor vehicles and pedestrians are influenced by groups; for groups, individuals within the group will determine specific actions based on the similarity between their own behavior and the average behavior of the group under the guidance of the group cooperation / competition tendency; the final action is adjusted by the direction constraint module of the middle layer network, which combines domain knowledge to constrain the action space of the lower layer network, and constrains the acceleration range for different interaction tendencies to ensure a balance between group consistency and individual difference; considering the differences in behavior characteristics and movement patterns between non-motor vehicles and pedestrians, dedicated generator networks are constructed for non-motor vehicles and pedestrians.

6. The data-driven simulation method for mixed traffic flow complex interaction behavior according to claim 5, characterized in that, A flocking tendency quantization mechanism is introduced for non-motor vehicles and pedestrians, that is, the final execution action of an individual is determined by evaluating the flocking degree of the individual to the group; the motion similarity between individuals in the group and the whole group is measured by cosine similarity As a core measurement index, the higher the similarity, the greater the influence of the group on the individual, the higher the flocking degree, and the more inclined to take the group action; vice versa; as follows: By calculating the cosine of the angle between vectors, both directional consistency and relative amplitude similarity are captured. In the specific implementation, key motion features are selected to construct feature vectors, including... ,in, , Representing the intelligent agent in shaft and Position in the axial direction , Represents its shaft and Velocity in the axial direction; , They respectively represent the points at the endpoint and the destination. shaft and Distance along the axial direction; The heading angle of the agent, i.e., the angle between the agent and the target. The angle between the axes; where the individual feature vector is formed by the population maintenance period. The sequence of motion states is composed of steps, and the group feature vector is obtained by calculating the mean of the feature vectors of all individuals at each step; due to the real-world scenario... , , and The characteristics change relatively stably. , and Features dominate changes in motion patterns, and cosine similarity can effectively reflect the directional consistency and relative amplitude consistency of motion features; the range of cosine similarity is... To facilitate comparison, the similarity results were normalized, mapping their value range to a uniform standard. The closer the value is to 1, the higher the similarity between the individual and the group. similarity The calculation formula is as follows: wherein and is a vector comprising dimensions, denotes the dot product, denotes the length of the vector.

7. The data-driven simulation method for mixed traffic flow complex interaction behavior according to claim 4, characterized in that, Each agent comprises a binary classification discriminator; The observation-action pair is input, and the probability estimation derived from the real expert demonstration is output, which is used to distinguish whether the given input is from the observation-action pair generated by the generator or from the real observation-action pair; the discriminator includes a reward estimator and a potential function, both of which are double-layer MLP networks, and the reward estimator provides an optimization signal for the generator.

8. The data-driven simulation method for mixed traffic flow complex interaction behavior according to claim 4, characterized in that, Based on the centralized training-distributed execution CTDE framework, the discriminator network and the generator network are collaboratively trained by alternating optimization, and the specific process is as follows: Train discriminator: input is observation-action pairs sampled from generated data and expert data and ; update reward estimator parameters using AdamOptimizer , potential function parameters , maximize the following objective: where, is the discriminator output, representing the probability that the sample comes from the expert trajectory, is the reward function, containing a reward estimator and a potential function , both are double-layer MLP networks; the reward estimator provides the optimization signal for the generator; is an adaptive sampler with parameters , representing the probability that the generator selects action given observation , and is the discount factor; Training the generator: input is observation data, action data, reward from the discriminator, and value estimator network output; update the generator network parameters using the MACK algorithm, and the objective function is as follows: By repeatedly optimizing the parameters in the generator and the discriminator for different observation-action pair samples, the learned strategy network and reward function are obtained as the output of the entire training algorithm.

9. The data-driven simulation method for mixed traffic flow complex interaction behavior according to claim 1, wherein, Step S3 is specifically: Interaction strategy rationality evaluation: evaluate from two aspects of overall decision accuracy and cooperation / competition tendency dynamic evolution rationality; Trajectory accuracy evaluation: the similarity between the generated trajectory and the ground truth is evaluated by calculating the average distance error of the position at the micro level, and the distribution similarity of the velocity, heading angle and the average value of all interactions with objects at the macro level. value at the macro level.

10. A data-driven simulation system for complex interaction behavior of mixed traffic flows for implementing the method according to any one of claims 1 to 9, characterized in that The model construction module and the application simulation module are included. The model construction module includes: An expert data acquisition unit is responsible for defining the observation and action space in multi-agent imitation learning, collecting real motor vehicle, non-motor vehicle and pedestrian traffic trajectory data, and constructing a feature matrix to support subsequent model training; A strategy simulation training unit trains an agent to reproduce expert behavior logic based on a multi-agent imitation learning framework, and constructs a hybrid traffic flow integrated strategy simulation model; A strategy simulation evaluation unit fuses strategy matching degree and trajectory accuracy to construct a multi-dimensional evaluation system to test the effectiveness and reliability of the simulation model in real interaction tasks; The application simulation module includes: A scene initialization unit provides initial state and background environment data required for generating a scene; A dynamic scene generation unit generates diversified and high-complexity interaction scenes based on the trained strategy simulation model.

Images