Behavior modeling method and system based on generative adversarial imitation and interaction representation

Abstract

The application discloses a kind of behavior modeling method and system based on generation confrontation imitation and interactive representation, wherein the method comprises: collecting the multi-modal driving and perception data of each traffic subject, and processing the data to obtain state input data in the form of uniform input sequence;State input data is input into behavior prediction model to obtain predicted driver driving behavior.The application establishes driver behavior prediction model, outputs implicit scene coding that fuses the historical motion state of ego vehicle, candidate trajectory and neighbor vehicle interaction information, converts scene information into output vector suitable for imitation learning, can highly summarize current traffic scene from decision-related perspective, thereby significantly reducing state space dimension and improving modeling stability;And implicit scene coding is input into generation confrontation imitation structure, optimized by minimizing the maximum value of sample discrimination loss function of discriminator network, improve the accuracy and reliability of driver behavior prediction.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving, and in particular to a behavior modeling method and system based on generative adversarial imitation and interactive representation. Background Technology

[0002] Driver behavior models play a bridging role in autonomous driving systems, serving as a crucial foundation for understanding human driving intentions and achieving safe and interpretable driving decisions. Current driver behavior modeling still faces several challenges, especially in human-machine co-driving scenarios. Models not only need to reproduce the driver's dynamic decision-making patterns but also need to appropriately respond to the collaborative relationships and control transfers between the driver and the automatic control system. Traditional rule-based models typically rely on manually set driving logic, which struggles to maintain stability and generalization performance in complex and dynamic traffic scenarios. In recent years, imitation learning methods have directly learned policy distributions from real driving data, reducing reliance on explicit rules and becoming an important direction for behavior modeling. To enhance model generalization, generative adversarial imitation learning has been introduced, but most research remains limited to behavior learning based on low-dimensional state features. These features are often insufficient to express the complex interactions between multiple traffic participants and the semantic information of the environment.

[0003] In existing technologies, rule-based methods based on manually designed behavioral guidelines or decision-making logic struggle to cope with the dynamic changes in traffic elements and the interactions between multiple stakeholders, failing to achieve flexible decision-making. Imitation learning methods, which directly fit expert behavior distributions, are easily limited by data distribution bias and sample sparsity, leading to decreased prediction performance in unknown scenarios. Existing driver modeling methods based on generative adversarial imitation learning typically rely only on low-dimensional state features (such as position, speed, and acceleration), making it difficult to effectively capture complex interactions between multiple stakeholders and high-order scene semantic information. Summary of the Invention

[0004] To address the challenges of existing technologies in handling dynamic changes in traffic elements, multi-agent interactions, and capturing complex interactive relationships and high-level scene semantic information, this invention provides a behavior modeling method and system based on generative adversarial imitation and interactive representation, thereby improving the accuracy and reliability of driver behavior prediction.

[0005] Therefore, the technical solution adopted by the present invention is as follows: A behavior modeling method based on generative adversarial imitation and interaction representation is provided, the method comprising: Multimodal driving and perception data of various traffic entities are collected and processed to obtain state input data in a unified input sequence format; the traffic entities include the vehicle, surrounding traffic participants, and road structure; the state input data specifically includes the vehicle's historical motion state and candidate trajectory points, and the neighboring vehicle's historical motion state and candidate trajectory points. Input the state input data into the behavior prediction model to obtain the predicted driver driving behavior; The construction of the behavior prediction model specifically includes: passing the state input data sequentially through a dynamic encoding layer and a cross-modal layer to establish the temporal dependency and spatial interaction relationship between the historical motion state of each vehicle and the candidate trajectory points; fusing the spatial interaction relationship of each vehicle through an aggregation layer; and adding the candidate trajectory points of the vehicle itself as a constraint condition for encoding in the output layer to obtain an implicit scene encoding that integrates the historical motion state of the vehicle, the interaction information of neighboring vehicles, and the candidate trajectory; inputting the implicit scene encoding into an adversarial imitation learning framework to obtain the initial predicted driving behavior; and optimizing it by minimizing the policy distribution probability of the policy network in the framework to obtain the final predicted driving behavior.

[0006] According to the above scheme, multimodal driving and perception data include vehicle status information, surrounding traffic participants and road structure information, surrounding image data, point cloud data and BEV data; vehicle status information includes vehicle speed, acceleration, heading angle, steering wheel angle and accelerator and brake pedal status, and surrounding traffic participants include the dynamic characteristics of neighboring vehicles and pedestrians.

[0007] According to the above scheme, the implicit scene encoding is specifically obtained based on a multi-level Transformer structure, and is obtained in the following way: The time dependency between the historical motion state of a vehicle and the candidate trajectory points is modeled by a dynamic coding layer, and the historical motion dynamic embedding of each vehicle and the time embedding of each candidate trajectory point are output. Establish the interaction relationship between vehicle motion state and candidate trajectory points through cross-modal layers; A global scene representation of the vehicle is constructed through an aggregation layer, and the historical motion state of the vehicle and the cross-modal interaction results of neighboring vehicles are fused into a single decision vector. The output layer combines the fused single decision vector with the candidate trajectory points of the vehicle. Using the candidate trajectory of the vehicle as a constraint, the fused single decision vector is constrained and encoded to obtain an implicit scene code that integrates the historical motion state of the vehicle, the dynamic interaction information of neighboring vehicles, and the constraints of the candidate trajectory of the vehicle.

[0008] According to the above scheme, the dynamic coding layer also embeds and encodes the state input data, including encoding the geometry and feature semantics of the candidate paths of the vehicle and neighboring vehicles to form path feature vectors, and modeling the time dependency relationship between the vehicle's historical motion state and the candidate trajectory points based on the encoded path feature vectors.

[0009] According to the above scheme, implicit scene encoding is introduced into the generative adversarial learning framework in the following way: Implicit scene codes are input into the policy network and the discriminator network, respectively. The policy network outputs driving control actions or decision commands. The discriminator network receives state-action pairs composed of expert driving data and state-action pairs generated by the policy network, and determines the source of the sample data. The source of the sample data includes expert drivers and non-expert drivers. The state is specifically the implicit scene code, and the action is specifically the driving control action or decision command. The discriminator network is trained with the goal of maximizing the sample discrimination loss function, which is specifically calculated based on the expected values ​​of the state-action pairs formed by expert driver data and the state-action pairs generated by the policy network. The policy network is optimized and trained with the goal of minimizing the policy distribution probability of the policy network in the framework, which is specifically calculated based on the sample discrimination loss function.

[0010] According to the above scheme, the policy network is specifically input with the corresponding implicit scene code at any time, and the predicted driver continuous control quantity or discrete decision command in the scene state at that time is represented in the form of a parameterized probability distribution, wherein the continuous control quantity includes steering angle and acceleration.

[0011] According to the above scheme, the discriminator network specifically takes into account the implicit scene encoding and the continuous control quantity or discrete decision instruction predicted by the policy network, and outputs a scalar between 0 and 1 representing the probability that the current input comes from an expert driver; wherein, the closer the output scalar is to 1, the greater the probability that the input comes from an expert driver.

[0012] A behavior modeling system based on generative adversarial imitation and interactive representation is also provided, the system comprising: The data processing module is used to collect multimodal driving and perception data from various traffic entities, and process the data to obtain state input data in the form of an input sequence. The behavior prediction module is used to input state input data into the behavior prediction model to obtain the predicted driver driving behavior.

[0013] According to the above scheme, the behavior prediction module is specifically used to: model the temporal dependency between the vehicle's historical motion state and candidate trajectory points through a dynamic coding layer, and output the dynamic embedding of the historical motion of each vehicle and the temporal embedding of each candidate trajectory point; establish the interaction relationship between the vehicle's motion state and candidate trajectory points through a cross-modal layer; construct a global scene representation of the vehicle through an aggregation layer, and fuse the vehicle's historical motion state and the cross-modal interaction results of neighboring vehicles into a single decision vector; combine the fused single decision vector with the vehicle's candidate trajectory points through an output layer, and use the vehicle's candidate trajectory as a constraint condition to constrain and encode the fused single decision vector, thereby obtaining an implicit scene code that integrates the vehicle's historical motion state, the dynamic interaction information of neighboring vehicles, and the constraints of the vehicle's candidate trajectory.

[0014] A computer storage medium is also provided, which stores a computer program that can be executed by a processor, the computer program performing the behavior modeling method based on generative adversarial imitation and interactive representation described above.

[0015] The beneficial effects of this invention are as follows: By establishing a driver behavior prediction model, the state input data is sequentially passed through a dynamic encoding layer, a cross-modal layer, an aggregation layer, and an output layer to output implicit scene encoding. This transforms high-dimensional, structured scene information into an output vector suitable for imitation learning. Furthermore, this output vector integrates the vehicle's historical motion state, dynamic interaction information with neighboring vehicles, and future candidate trajectory constraints, enabling a high-level generalization of the current traffic scene from a decision-related perspective. This significantly reduces the dimensionality of the state space and improves modeling stability. Moreover, by optimizing the implicit scene encoding input to the adversarial imitation learning framework, the behavior distribution generated by the policy network approximates the behavior distribution of expert drivers, thereby improving the accuracy and reliability of driver behavior prediction.

[0016] Furthermore, this invention jointly optimizes and generates predicted driver simulation driving behavior through a policy network and a discriminator network. By imitating and learning from expert driving data, it generates driving behavior that closely resembles the logic of real behavior and follows the decision-making preferences of human drivers in complex interactive environments. This provides the convergence stability and generalization ability of imitation learning, thereby achieving the accuracy of the model's prediction of driving behavior. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of the behavioral modeling method based on generative adversarial imitation and interactive representation according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a multi-level Transformer structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the system structure of the behavioral modeling system based on generative adversarial imitation and interactive representation according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] To address the limitations of existing rule-based methods that rely on manually designed behavioral guidelines or decision logic in handling dynamic changes in traffic elements and multi-agent interactions, as well as effectively capturing complex inter-agent relationships and high-order scene semantic information, this invention provides a behavioral modeling method based on generative adversarial imitation and interactive representation. Figure 1As shown, the method includes: S1. Collect multimodal driving and perception data from various traffic entities, process the data, and obtain state input data in the form of a unified input sequence.

[0020] S2. Input the state input data into the behavior prediction model to obtain the predicted driver driving behavior.

[0021] Specifically, traffic participants include the driver's own vehicle, neighboring vehicles, and pedestrians; multimodal driving data includes the driver's own vehicle status information, surrounding traffic participant data, and road structure information; among these, the driver's own vehicle status information... Including vehicle speed acceleration Heading angle The data includes steering wheel angle and accelerator and brake pedal status, surrounding traffic participant data including vehicle and pedestrian dynamic characteristics, and road structure information from high-precision maps, used to reflect traffic constraints and topological relationships.

[0022] Specifically, the construction of the behavior prediction model includes: passing the state input data sequentially through a dynamic encoding layer and a cross-modal layer to establish the temporal dependency and spatial interaction relationship between the historical motion state of each vehicle and the candidate trajectory points; fusing the spatial interaction relationship of each vehicle through an aggregation layer; and adding the candidate trajectory points of the vehicle itself as a constraint condition for encoding in the output layer to obtain an implicit scene encoding that integrates the historical motion state of the vehicle, the interaction information of neighboring vehicles, and the candidate trajectory; inputting the implicit scene encoding into a generative adversarial imitation learning framework to obtain the initial predicted driving behavior; and optimizing it by minimizing the policy distribution probability of the policy network in the framework to obtain the final predicted driving behavior.

[0023] Specifically, the schematic diagram of the multi-level Transformer structure in this embodiment is as follows: Figure 2 As shown; the implicit scene encoding based on the multi-level Transformer structure is obtained in the following way: The time dependency between the historical motion state of a vehicle and the candidate trajectory points is modeled by a dynamic coding layer, and the historical motion dynamic embedding of each vehicle and the time embedding of each candidate trajectory point are output. Establish the interaction relationship between vehicle motion state and candidate trajectory points through cross-modal layers; A global scene representation of the vehicle is constructed through an aggregation layer, and the historical motion state of the vehicle and the cross-modal interaction results of neighboring vehicles are fused into a single decision vector. The output layer combines the fused single decision vector with the candidate trajectory points of the vehicle. Using the candidate trajectory of the vehicle as a constraint, the fused single decision vector is constrained and encoded to obtain the implicit encoding of the vehicle's historical motion state, the dynamic interaction information of neighboring vehicles, and the constraint of the candidate trajectory of the vehicle.

[0024] Specifically, the vehicle's historical motion state includes the historical motion state sequence of the vehicle itself and neighboring vehicles. It is used to describe the dynamic characteristics of a vehicle, such as its position, speed, and heading, over a period of time; the candidate trajectory points are specifically the set of candidate trajectory points generated for each vehicle. This characterizes the future driving path that the vehicle may take in the current traffic environment; the vehicle itself is denoted as ego, and the other vehicles are denoted as neighboring vehicles. .

[0025] Specifically, the dynamic coding layer models the temporal dependency between the vehicle's historical motion and candidate trajectory points. Without introducing multi-vehicle interaction, it extracts high-quality temporal dynamic representations to characterize the vehicle's overall motion trend and behavioral characteristics at the current moment, providing basic features for subsequent interaction modeling.

[0026] In addition, the dynamic coding layer also embeds and encodes the state input data. Specifically, it encodes the geometry and semantic features of the candidate paths of the vehicle and neighboring vehicles, including the start point, end point, curvature, length, and lane line coordinates, to form a path feature vector. Based on the encoded path feature vector, it models the time dependency between the vehicle's historical motion state and the candidate trajectory points.

[0027] Based on the vehicle's historical motion state Build-time Transformer encoder Its calculation form is:

[0028] Where X=Q=K=V represents the input historical motion sequence; MHA represents the multi-head self-attention mechanism; MaxPool is used to compress the time dimension and enhance feature robustness; MLP is used for non-linear feature mapping.

[0029] To enhance the temporal structure representation of candidate trajectory points, candidate points are linked to learnable embedding vectors. Perform the stitching to construct a trajectory-time Transformer:

[0030] This structure is used to distinguish the category features between different vehicles and different candidate trajectories, and to provide a consistent feature space for subsequent cross-level queries.

[0031] The output of the dynamic coding layer can be represented as: Historical motion dynamics of each vehicle embedded and the temporal embedding of each candidate trajectory point ,in, Indicates the first The first car Candidate trajectories.

[0032] Specifically, a cross-modal Transformer can be represented as:

[0033] Among them, query Derived from vehicle dynamics features; key and value The features are derived from candidate trajectories; M is the corresponding attention mask.

[0034] For the The cross-modal dynamics of a neighboring vehicle can be represented as:

[0035] This layer structure is based on the assumption of real-world traffic scenarios, namely that there is no explicit cooperative behavior between vehicles, and the future driving path of a vehicle is mainly determined by its own historical dynamics and available trajectories. Therefore, cross-modal modeling focuses on characterizing the correlation between the motion state features of an individual vehicle and its candidate trajectory features. By constructing a cross-modal Transformer encoder, using the vehicle's historical dynamic embedding as the query vector and the corresponding candidate trajectory embedding as the key and value, selective attention to future trajectory information by the vehicle's state is achieved. Simultaneously, a residual connection structure is introduced, ensuring the effective transmission of the vehicle's original dynamic features even when candidate trajectory information is missing or unreliable, thereby enhancing the model's robustness.

[0036] Specifically, the aggregation layer is used to construct a global scene representation for vehicle decision-making. This layer uses the vehicle's historical dynamic embedding as the core query vector, jointly encoding the vehicle's own dynamic features with the cross-modal interaction results of all neighboring vehicles. Through a self-attention mechanism, the vehicle can proactively select the neighbor information most relevant to its current decision, while suppressing traffic participants that are irrelevant or have little impact on the decision, thus achieving scene information aggregation centered on the vehicle. This aggregation result forms a high-level scene representation used to describe the dynamic interaction states in the current traffic environment that are highly relevant to the vehicle's decision. Its structure can be represented as follows:

[0037] Specifically, the output layer further incorporates candidate trajectory information from the vehicle to enhance the encoded aggregated scene representation, explicitly modeling the vehicle's future intentions and its potential interactions with the environment. By constructing a cross-attention Transformer, the vehicle's aggregated features are used as query vectors to focus on the embeddings of the vehicle's candidate trajectories, thus integrating information about the future action space into the final representation. This stage also employs a residual connection structure to ensure the stability of the information flow and the continuity of expression. The final output implicit scene representation comprehensively considers historical dynamics, neighbor interactions, and future trajectory constraints, and can be directly used as the input state for reinforcement learning policy networks or autonomous driving decision-making modules. Its structure can be represented as follows:

[0038] Furthermore, implicit scene encoding is introduced into the Generative Adversarial Imitation Learning (GAIL) framework in the following way: Implicit scene encoding is input into the policy network and the discriminator network, respectively. The policy network outputs driving control actions or decision commands. The discriminator network receives state-action pairs composed of expert driving data and state-action pairs generated by the policy network, and determines the source of the sample data. The policy network is optimized and trained with the goal of minimizing the policy distribution probability of the policy network in the framework. The specific policy distribution probability is calculated based on the sample discrimination loss function.

[0039] Specifically, the policy network uses implicit scenario encoding. As input, it outputs corresponding driving control actions or high-level decision commands to generate simulated driving behavior; the discriminator network receives state-action pairs composed of expert driving data and state-action pairs generated by the policy network, and determines their source based on an adversarial training mechanism. Because It explicitly includes multi-vehicle interaction relationships and future intention constraints. The discriminator can distinguish between human driving behavior and policy generation behavior at a higher semantic level, thereby guiding the policy network to learn behavioral patterns that are more in line with the decision-making logic of real drivers.

[0040] because This continuous implicit representation, obtained through a multi-stage Transformer structure, possesses excellent temporal consistency and robustness. Therefore, during GAIL sequence training, it effectively mitigates training instability issues caused by sensor noise, variations in the number of traffic participants, or uncertainty in candidate trajectories. Furthermore, this representation avoids explicitly modeling complex traffic interaction rules during the imitation learning phase, making driver behavior modeling more data-driven and exhibiting stronger generalization capabilities.

[0041] Specifically, the implicit scene encoding at input time t of the policy network. The output is the corresponding predicted driver actions. The policy network can be represented as a parameterized probability distribution. in, This represents the set of parameters for the policy network. It can be a continuous control quantity, such as steering angle and acceleration, or a discrete decision command; through this strategy distribution, the behavior of a simulated driver can be generated under given scenario coding conditions.

[0042] Corresponding to the policy network, the discriminator network distinguishes whether the current state-action pair originates from real driver data or data generated by the policy network. The discriminator uses implicit scene encoding. Corresponding actions The discriminator network takes a combination of inputs as input and outputs a scalar between 0 and 1, representing the probability of that state-action pair coming from an expert driver. This discriminator network can be represented as: ,in, The set of parameters for the discriminator network is denoted by 1. The closer the output value is to 1, the more likely the sample is to be derived from real driver behavior.

[0043] During training, the policy network and the discriminator network are jointly updated through an adversarial optimization objective. The discriminator's optimization objective is to maximize its ability to distinguish between real driver samples and policy-generated samples, and its loss function can be expressed as:

[0044] The optimization goal of the policy network is to generate driving behaviors that can "deceive" the discriminator, making the generated state-action pairs indistinguishable from expert driving behavior in the discriminator's eyes. To this end, the policy network constructs reward signals from the discriminator's output and updates them through reinforcement learning. The corresponding policy optimization objective can be expressed as:

[0045] The optimization objective of generative adversarial learning can then be expressed as:

[0046] Due to implicit scene state By integrating the vehicle's historical dynamics, neighbor vehicle interaction information, and future candidate trajectory constraints, the policy network can implicitly follow the decision-making preferences of human drivers in complex interactive environments when generating actions. Simultaneously, the discriminator distinguishes driving behaviors at a higher semantic level, making the adversarial training process no longer dependent on low-level physical features, thereby improving the convergence stability and generalization ability of imitation learning.

[0047] Furthermore, this invention also provides a behavior modeling system based on generative adversarial imitation and interactive representation, used to implement the behavior modeling method based on generative adversarial imitation and interactive representation described in this invention, such as... Figure 2 As shown, the system includes: The data processing module is used to collect multimodal driving and perception data from various traffic entities, and process the data to obtain state input data in the form of an input sequence. The behavior prediction module is used to input state input data into the behavior prediction model to obtain the predicted driver driving behavior.

[0048] The various modules or mechanisms of the system are mainly used to implement the various steps of the above method embodiments, and will not be described in detail here.

[0049] This invention establishes a driver behavior prediction model, passing state input data sequentially through a dynamic encoding layer, a cross-modal layer, an aggregation layer, and an output layer to output implicit scene encoding. This transforms high-dimensional, structured scene information into an output vector suitable for imitation learning. Furthermore, this output vector incorporates the vehicle's historical motion state, dynamic interaction information with neighboring vehicles, and future candidate trajectory constraints, enabling a high-level generalization of the current traffic scenario from a decision-related perspective. This significantly reduces the dimensionality of the state space and improves modeling stability. Additionally, the implicit scene encoding input to the adversarial imitation learning framework is optimized, making the behavior distribution generated by the policy network approximate the behavior distribution of expert drivers, thus improving the accuracy and reliability of driver behavior prediction.

[0050] Furthermore, in this embodiment of the invention, the driver simulation behavior is jointly optimized and generated by a policy network and a discriminator network. By imitating and learning from expert driving data, the driving behavior that closely resembles the real driving logic and follows the decision-making preferences of human drivers in complex interactive environments is generated. This provides the convergence stability and generalization ability of imitation learning, and enables the model to predict the accuracy of driving behavior.

[0051] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0052] The order of the steps in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0053] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A behavior modeling method based on generative adversarial imitation and interactive representation, characterized in that, The method includes: Multimodal driving and perception data of various traffic entities are collected and processed to obtain state input data in a unified input sequence format; the traffic entities include the vehicle, surrounding traffic participants, and road structure; the state input data specifically includes the vehicle's historical motion state and candidate trajectory points, and the neighboring vehicle's historical motion state and candidate trajectory points. Input the state input data into the behavior prediction model to obtain the predicted driver driving behavior; The construction of the behavior prediction model specifically includes: passing the state input data sequentially through a dynamic encoding layer and a cross-modal layer to establish the temporal dependency and spatial interaction relationship between the historical motion state of each vehicle and the candidate trajectory points; fusing the spatial interaction relationship of each vehicle through an aggregation layer; and adding the candidate trajectory points of the vehicle itself as a constraint condition for encoding in the output layer to obtain an implicit scene encoding that integrates the historical motion state of the vehicle, the interaction information of neighboring vehicles, and the candidate trajectory; inputting the implicit scene encoding into a generative adversarial imitation learning framework to obtain the initial predicted driving behavior; and optimizing it by minimizing the policy distribution probability of the policy network in the framework to obtain the final predicted driving behavior.

2. The behavior modeling method based on generative adversarial imitation and interactive representation according to claim 1, characterized in that, Multimodal driving and perception data includes vehicle status information, surrounding traffic participants and road structure information, surrounding image data, point cloud data and BEV data; vehicle status information includes vehicle speed, acceleration, heading angle, steering wheel angle and accelerator and brake pedal status, and surrounding traffic participants include the dynamic characteristics of neighboring vehicles and pedestrians.

3. The behavior modeling method based on generative adversarial imitation and interactive representation according to claim 1, characterized in that, Implicit scene encoding is specifically obtained based on a multi-level Transformer structure, and is achieved through the following methods: The time dependency between the historical motion state of a vehicle and the candidate trajectory points is modeled by a dynamic coding layer, and the historical motion dynamic embedding of each vehicle and the time embedding of each candidate trajectory point are output. Establish the interaction relationship between vehicle motion state and candidate trajectory points through cross-modal layers; A global scene representation of the vehicle is constructed through an aggregation layer, and the historical motion state of the vehicle and the cross-modal interaction results of neighboring vehicles are fused into a single decision vector. The output layer combines the fused single decision vector with the candidate trajectory points of the vehicle. Using the candidate trajectory of the vehicle as a constraint, the fused single decision vector is constrained and encoded to obtain an implicit scene code that integrates the historical motion state of the vehicle, the dynamic interaction information of neighboring vehicles, and the constraints of the candidate trajectory of the vehicle.

4. The behavior modeling method based on generative adversarial imitation and interactive representation according to claim 3, characterized in that, The dynamic coding layer also embeds and encodes the state input data, including encoding the geometry and semantic features of the candidate paths of the vehicle and neighboring vehicles to form path feature vectors, and models the time dependency between the vehicle's historical motion state and the candidate trajectory points based on the encoded path feature vectors.

5. The behavior modeling method based on generative adversarial imitation and interactive representation according to claim 1, characterized in that, The implicit scene encoding is introduced into the generative adversarial learning framework in the following way: Implicit scene codes are input into the policy network and the discriminator network, respectively. The policy network outputs driving control actions or decision commands. The discriminator network receives state-action pairs composed of expert driving data and state-action pairs generated by the policy network, and determines the source of the sample data. The source of the sample data includes expert drivers and non-expert drivers. The state is specifically the implicit scene code, and the action is specifically the driving control action or decision command. The discriminator network is trained with the goal of maximizing the sample discrimination loss function, which is specifically calculated based on the expected values ​​of the state-action pairs formed by expert driver data and the state-action pairs generated by the policy network. The policy network is optimized and trained with the goal of minimizing the policy distribution probability of the policy network in the framework, which is specifically calculated based on the sample discrimination loss function.

6. The behavior modeling method based on generative adversarial imitation and interactive representation according to claim 5, characterized in that, The policy network takes the corresponding implicit scene code at any given time as input and represents the predicted continuous control quantity or discrete decision command of the driver in the scene state at that time in the form of a parameterized probability distribution. The continuous control quantity includes steering angle and acceleration.

7. The behavior modeling method based on generative adversarial imitation and interactive representation according to claim 6, characterized in that, The discriminator network takes as input the implicit scene code and the continuous control quantity or discrete decision instruction predicted by the policy network, and outputs a scalar between 0 and 1 representing the probability that the current input comes from an expert driver; the closer the output scalar is to 1, the greater the probability that the input comes from an expert driver.

8. A behavior modeling system based on generative adversarial imitation and interactive representation, characterized in that, The system includes: The data processing module collects multimodal driving and perception data from various traffic entities and processes the data to obtain state input data in a unified input sequence format. The traffic entities include the vehicle, surrounding traffic participants, and road structure. The state input data specifically includes the vehicle's historical motion state and candidate trajectory points, and the neighboring vehicle's historical motion state and candidate trajectory points. The behavior prediction module is used to input state input data into the behavior prediction model to obtain the predicted driver driving behavior; The construction of the behavior prediction model specifically includes: passing the state input data sequentially through a dynamic encoding layer and a cross-modal layer to establish the temporal dependency and spatial interaction relationship between the historical motion state of each vehicle and the candidate trajectory points; fusing the spatial interaction relationship of each vehicle through an aggregation layer; and adding the candidate trajectory points of the vehicle itself as a constraint condition for encoding in the output layer to obtain an implicit scene encoding that integrates the historical motion state of the vehicle, the interaction information of neighboring vehicles, and the candidate trajectory; inputting the implicit scene encoding into a generative adversarial imitation learning framework to obtain the initial predicted driving behavior; and optimizing it by minimizing the policy distribution probability of the policy network in the framework to obtain the final predicted driving behavior.

9. The behavior modeling system based on generative adversarial imitation and interactive representation according to claim 8, characterized in that, The behavior prediction module is specifically used for: modeling the temporal dependency between the vehicle's historical motion state and candidate trajectory points through a dynamic coding layer, outputting the dynamic embedding of the historical motion of each vehicle and the temporal embedding of each candidate trajectory point; establishing the interaction relationship between the vehicle's motion state and candidate trajectory points through a cross-modal layer; constructing a global scene representation of the vehicle through an aggregation layer, fusing the vehicle's historical motion state and the cross-modal interaction results of neighboring vehicles into a single decision vector; and combining the fused single decision vector with the vehicle's candidate trajectory points through an output layer, using the vehicle's candidate trajectory as a constraint condition to constrain and encode the fused single decision vector, thereby obtaining an implicit scene code that integrates the vehicle's historical motion state, the dynamic interaction information of neighboring vehicles, and the constraints of the vehicle's candidate trajectory.

10. A computer storage medium, characterized in that, It contains a computer program that can be executed by a processor, which performs the behavior modeling method based on generative adversarial imitation and interactive representation as described in any one of claims 1-7.

Images