Automatic driving simulation microscopic behavior interaction method and system based on AI intelligent agent

Abstract

The invention discloses an AI agent-based automatic driving simulation microscopic behavior interaction method and system, and relates to the field of automatic driving simulation. In order to solve the problems of poor simulation trueness, inaccurate prediction and low calculation efficiency in the prior art, the method comprises the following steps: firstly, preprocessing multi-source data; then, through a fusion embedding step, the data is deeply fused and embedded into a unified feature space, and a multi-modal feature vector is generated; then, through a Sim Transform coding step, converting the multi-modal feature vector into a high-dimensional representation of a global interaction feature; in the dense network prediction step, generating prediction information by using global interaction characteristics; then, a specific behavior decision is output through a Sim Transform decoding step in combination with Query; and finally, the agent behavior output step adjusts the actual operation of the agent according to the specific behavior decision. Accurate prediction and adjustment of the agent behavior decision are realized.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving simulation, and in particular to an AI agent-based autonomous driving simulation micro-behavior interaction method. Background Art

[0002] As autonomous driving technology gradually moves towards large-scale commercialization, simulation testing has become a key tool to verify its effectiveness. However, existing autonomous driving simulation technology is often limited to specific data sets or scenarios in simulation testing, and is unable to fully combine multi-source data (such as open source data, simulation data, and roadside data) to build complex and diverse micro-interactions. At the same time, existing behavior prediction based on rules or traditional machine learning methods is more difficult to extract effective features from high-dimensional complex data, resulting in poor performance in complex traffic scenarios and difficulty in achieving accurate prediction of AI agents under various conditions. Finally, existing simulation technologies are inefficient in processing dynamic environmental data such as location information and map information, and the computing resource consumption in large-scale scenarios is high, resulting in insufficient real-time performance of the simulation process and waste of computing resources.

[0003] Through searching the prior art literature, it is found that the invention application "A Simulation Test Method for Autonomous Driving" with the publication number CN115032912A adopts a deep reinforcement learning-based technology to generate and optimize traffic flow strategies. By accumulating a large amount of traffic flow sample data and training the deep reinforcement learning model, a single agent can autonomously learn and form an activity trajectory that conforms to the actual situation. By combining and arranging multiple single agents, a multi-agent traffic flow is formed, which is used to simulate the autonomous driving vehicle (main vehicle). The scene editor is used to configure the interaction between the multi-agent and the main vehicle, so as to simulate more complex traffic scenes and close to the traffic flow strategy of the real world. In addition, the deep reinforcement learning model optimizes the behavior of the single agent by constantly interacting with the environment, so that it can respond to the actions of the main vehicle more intelligently and improve the accuracy and reliability of the simulation test. However, the deep reinforcement learning model used in the invention has insufficient generalization ability. The model may perform well in a specific training environment, but due to its strong dependence on training data, the generalization ability of the model is often insufficient when facing unseen scenes. This may cause the model to be unable to accurately respond to emerging traffic conditions in actual autonomous driving simulations, thereby affecting the reliability of the simulation test. At the same time, the training of the deep reinforcement learning model in the invention requires a lot of computing resources and time, especially when dealing with complex traffic flow scenarios, this resource consumption problem is more obvious. This not only increases the development cost, but also limits the real-time application capability of the model. Finally, the invention is difficult to handle complex interactions between multiple agents. Although deep reinforcement learning can optimize the behavior of a single agent through training, in complex multi-agent interaction scenarios, the model may be difficult to fully capture the subtle interactions between agents, thereby affecting the accuracy of traffic flow strategies. The invention application "Object Behavior Data Generation Method, Related Methods and Devices" with publication number CN115032912A generates target object behavior data by defining factors related to object behavior and extracting these behavior features from the road test data of autonomous driving. In addition, the invention also includes the retrieval and adjustment of target behavior data to enhance the flexibility and efficiency of the test. However, the invention is difficult to implement and has poor generalization. Although the rule-based method can ensure a certain degree of controllability and interpretability, it is difficult to design and implement rules in the face of complex and changeable autonomous driving scenarios, and it is difficult to adapt to different traffic environments, resulting in weak generalization ability. At the same time, the invention is not accurate enough in generating behaviors at the interaction level. Although this method has made some progress in improving the authenticity of simulation tests, in practical applications, the generation of object behavior data may still not fully capture the interaction details between intelligent agents in complex scenarios, resulting in a gap between simulation results and the actual situation. Finally, the invention has high complexity and manual dependence on data adjustment. Although it allows the generated data to be adjusted, this often requires human intervention, which increases the complexity of the operation and may affect the efficiency and repeatability of simulation tests. Summary of the invention

[0004] In order to solve the problems of poor simulation authenticity, inaccurate prediction and low calculation efficiency in the prior art, the present invention provides the following solution: a method for micro-behavior interaction of autonomous driving simulation based on AI agent, the method comprising:

[0005] S1. Input and preprocessing step, preprocessing multi-source data to obtain trajectory data and map data;

[0006] S2. a fusion embedding step, in which the trajectory data and the map data are deeply fused and embedded into a unified feature space to obtain a multimodal feature vector;

[0007] S3.Sim Transformer encoding step, converting the multimodal feature vector into a high-dimensional feature representation of global interaction features;

[0008] S4. A dense prediction network step, generating prediction information according to the high-dimensional features of the global interaction features;

[0009] S5.Sim Transformer decoding step, used to output a specific behavior decision based on the prediction information combined with the Query;

[0010] S6. Agent behavior output step, adjusting the actual operation of the agent according to the specific behavior decision.

[0011] Furthermore, the multi-source data includes: open source data, simulation data and roadside data.

[0012] Furthermore, the format of the trajectory data is: [N,T h ,D t ], where N is the predicted number of vehicles, T h refers to the number of historical trajectory points, D t is the trajectory characteristic attribute;

[0013] The format of the map data is [N,L c ,D m ], where L c refers to the number of road lines, D m Refers to the characteristic attributes of the map.

[0014] Furthermore, the trajectory data and the map data are deeply fused and embedded into a unified feature space by:

[0015] X=φ k,s (g(F))

[0016] Implementation, where g represents the backbone network layer of the PointNet neural network model, φk,s Represents a maximum pooling layer.

[0017] Furthermore, the conversion of the multimodal feature vector into a high-dimensional feature representation of global interaction features is specifically achieved through a hierarchical sliding window method and a multi-head self-attention mechanism.

[0018] Further, the prediction information is obtained by:

[0019]

[0020] Get, where MLP represents the fully connected neural network layer, S i Represents the attention score of the input.

[0021] Furthermore, the Query includes a location query and an intent query.

[0022] An AI agent-based autonomous driving simulation micro-behavior interaction system, the system comprising: an input and preprocessing module, a fusion embedding module, a Sim Transformer encoder module, a dense prediction network module, a Sim Transformer decoder module and an agent behavior output module;

[0023] The input and preprocessing module is used to preprocess the multi-source data to obtain trajectory data and map data and send them to the fusion embedding module;

[0024] The fusion embedding module is used to deeply fuse the trajectory data and the map data and embed them into a unified feature space to obtain a multimodal feature vector and send it to the Sim Transformer encoder module;

[0025] The Sim Transformer encoder module is used to convert the multimodal feature vector into a high-dimensional feature representation of a global interaction feature, and send the high-dimensional feature representation of the global interaction feature to the dense prediction network module;

[0026] The dense prediction network module is used to generate prediction information according to the high-dimensional features of the global interaction features and send the prediction information to the Sim Transformer decoder module;

[0027] The Sim Transformer decoder module is used to output a specific behavior decision based on the prediction information combined with the Query, and send the specific behavior decision to the agent behavior output module;

[0028] The agent behavior output module is used to adjust the actual operation of the agent according to the specific behavior decision.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] (1) Based on the Sim Transformer architecture, this paper constructs a self-attention mechanism based on multi-agent trajectories to achieve behavior prediction and decision-making of multi-agents in complex traffic scenarios by comprehensively considering the spatiotemporal interaction relationship between multiple agents. The spatiotemporal interaction relationship between agents is comprehensively considered, which improves the accuracy and consistency of the behavior simulation of agents in the simulation.

[0031] (2) The present invention uses the Pointnet network to deeply fuse trajectory information and map information with feature alignment, embedding them into a unified feature space, and accurately modeling the spatiotemporal relationship between intelligent agents. This fusion embedding method can automatically adapt to the complexity of different traffic scenarios, ensure the invariance of rotation, translation and multi-scale of the input point cloud, and improve the adaptability of the model method in a changing traffic environment.

[0032] (3) The present invention uses a hierarchical sliding window mechanism. Local behavior response prediction (such as obstacle avoidance and lane change) is performed in a shorter time window, while global planning and path optimization are processed in a longer time window, ensuring the consistency of local behavior and global strategy. This can significantly improve computing efficiency and decision-making accuracy when processing large-scale multi-agent simulations.

[0033] (4) The present invention uses a dual query mechanism of location and intention. The location query dynamically adapts to changes in the current traffic environment based on the real-time location information of the intelligent agent to ensure that the operation of the intelligent agent is consistent with the surrounding traffic conditions; the intention query optimizes its decision-making process by analyzing the intelligent agent's future behavioral intentions (such as changing lanes, accelerating, etc.), so that the intelligent agent can predict and respond to potential risks in complex traffic scenarios. The dual query mechanism works together to improve the accuracy and consistency of the prediction of the intelligent agent's behavior in the simulation.

[0034] (5) The present invention provides a rich and diverse simulation environment by integrating open source data, simulation data and roadside data, significantly improving the authenticity and representativeness of the simulation. This data integration method not only expands the coverage of simulation scenarios, but also enhances the reliability of simulation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 is a flow chart of the method of the present invention;

[0036] Figure 2 This is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0037] Embodiment 1, combination Figure 1A method for micro-behavior interaction of autonomous driving simulation based on an AI agent, the method comprising:

[0038] S1. Input and preprocessing step, preprocessing multi-source data to obtain trajectory data and map data;

[0039] S2. a fusion embedding step for deeply fusing the trajectory data and the map data and embedding them into a unified feature space to obtain a multimodal feature vector;

[0040] S3.Sim Transformer encoding step, used to convert the multimodal feature vector into a high-dimensional feature representation of global interaction features;

[0041] S4. A dense prediction network step for generating prediction information based on the high-dimensional features of the global interaction features;

[0042] S5.Sim Transformer decoding step, used to output a specific behavior decision based on the prediction information combined with the Query;

[0043] S6. Agent behavior output step, used to adjust the actual operation of the agent according to the specific behavior decision.

[0044] Specifically, the present invention extracts and aligns features based on the fusion embedding of traffic scene context, and designs a dual query mechanism based on simulation scenes to accurately predict the interactive behavior of multiple agents. In terms of data training, the method uses massive multi-source data (including open source data, simulation data, and roadside data) for training and learning to improve the authenticity of autonomous driving simulation and the accuracy of intelligent behavior prediction.

[0045] The multi-source data includes: open source data, simulation data and roadside data.

[0046] The format of the trajectory data is: [N,T h ,D t ], where N is the predicted number of vehicles, T h refers to the number of historical trajectory points, D t is the trajectory characteristic attribute;

[0047] The format of the map data is [N,L c ,D m ], where L c refers to the number of road lines, D m Refers to the characteristic attributes of the map.

[0048] Specifically, the input and preprocessing steps are to standardize the above data and convert it into a unified vector representation to provide support for subsequent training and reasoning.

[0049]

[0050] Conversion, where represents the characteristics of the i-th trajectory, represents the i-th map feature, and merges the trajectory features and map features of all agents together to form a joint feature matrix F.

[0051] The trajectory data and the map data are deeply integrated and embedded into a unified feature space by:

[0052] X=φ k,s (g(F))

[0053] Implementation, where g represents the backbone network layer of the PointNet neural network model, φ k,s Represents a maximum pooling layer.

[0054] Specifically, traffic scenes are highly dynamic and have strong context dependence. The fusion embedding step processes the input trajectory data and map data as point cloud data, uses the PointNet neural network model backbone network to align the features of the two separately, and deeply fuses the multivariate heterogeneous data and embeds them into a unified feature space.

[0055] The conversion of the multimodal feature vector into a high-dimensional feature representation of global interaction features is specifically achieved through a hierarchical sliding window method and a multi-head self-attention mechanism.

[0056] Specifically, the Sim Transformer encoding step uses a hierarchical sliding window method to calculate the attention score, which can quickly process multi-dimensional input data, including spatiotemporal trajectory data and spatial position information, greatly improving efficiency. Then, through interactive calculations between windows, local and global information are integrated to capture the spatial interaction and time evolution characteristics between different agents, providing high-dimensional feature representation for subsequent behavior prediction.

[0057] First, given the input X, we can obtain Q, K, and V by linearly transforming them respectively. The calculation formulas are as follows:

[0058] Q=XW q ,K=XW k , V = XW v

[0059] Among them, Q, K, V are query vector, key vector and value vector respectively, and W q , W k and W v are the corresponding learned parameter matrices respectively.

[0060] Then, the attention score is calculated, which is obtained by:

[0061]

[0062] Get, where α i Represents the weight coefficient of the attention result of each level, Q i , K i and V i are the query, key, and value vectors of the i-th head, respectively.

[0063] Finally, multi-head attention calculation is performed, and the multi-head attention is calculated by:

[0064] MultiHeadAttn=Concat(head1,head2,...,head h )W o

[0065] Get, where W o Output linear transformation matrix.

[0066] The prediction information is obtained by:

[0067]

[0068] Get, where MLP represents the fully connected neural network layer, S i Represents the attention score of the input.

[0069] Specifically, the dense prediction network step utilizes a multi-layer fully connected neural network structure to predict the future position, speed, acceleration and other behaviors of multiple agents.

[0070] The Query includes a location query and an intent query.

[0071] Specifically, the query is obtained through model training and learning in multi-source massive data and is used to guide interactive behavior.

[0072] The location query is responsible for making behavioral decisions based on the real-time location information of the intelligent agent to ensure that the operation of the intelligent agent is consistent with the current traffic environment; the intention query analyzes the future behavioral intentions of the intelligent agent (such as changing lanes, accelerating, etc.). Based on the predicted information combined with the query output, the specific behavior decision is made through:

[0073]

[0074] To achieve, represents the interaction behavior predicted by the i-th agent, MLP represents the fully connected neural network layer, MutiHeadAttn represents multi-head attention calculation, Query, Key, and Value represent query, key, and value vectors respectively, and Q i p represents the position query vector of the ith agent, Q i i represents the intent query vector of the i-th agent, K i represents the key vector of the ith agent, X i represents the key vector of the ith agent, Represents the characteristics of a dense prediction network.

[0075] Embodiment 2, combined Figure 2 This embodiment is described. An AI agent-based autonomous driving simulation micro-behavior interaction system, the system comprising: an input and preprocessing module, a fusion embedding module, a Sim Transformer encoder module, a dense prediction network module, a Sim Transformer decoder module and an agent behavior output module;

[0076] The input and preprocessing module is used to preprocess the multi-source data to obtain trajectory data and map data and send them to the fusion embedding module;

[0077] The fusion embedding module is used to deeply fuse the trajectory data and the map data and embed them into a unified feature space to obtain a multimodal feature vector and send it to the Sim Transformer encoder module;

[0078] The Sim Transformer encoder module is used to convert the multimodal feature vector into a high-dimensional feature representation of a global interaction feature, and send the high-dimensional feature representation of the global interaction feature to the dense prediction network module;

[0079] The dense prediction network module is used to generate prediction information according to the high-dimensional features of the global interaction features and send the prediction information to the Sim Transformer decoder module;

[0080] The Sim Transformer decoder module is used to output a specific behavior decision based on the prediction information combined with the Query, and send the specific behavior decision to the agent behavior output module;

[0081] The agent behavior output module is used to adjust the actual operation of the agent according to the specific behavior decision.

[0082] Although the present invention has been described according to a limited number of embodiments, it will be apparent to those skilled in the art, with the benefit of the above description, that other embodiments may be envisioned within the scope of the invention thus described. In addition, it should be noted that the language used in this specification is selected primarily for readability and teaching purposes, rather than for explaining or defining the subject matter of the present invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the present invention is illustrative, not restrictive, with respect to the scope of the present invention, which is defined by the appended claims.

Claims

1. A method for micro-behavior interaction of autonomous driving simulation based on AI agent, characterized in that: The method comprises: S1. Input and preprocessing step, preprocessing multi-source data to obtain trajectory data and map data; S2. a fusion embedding step, in which the trajectory data and the map data are deeply fused and embedded into a unified feature space to obtain a multimodal feature vector; S3.Sim Transformer encoding step, converting the multimodal feature vector into a high-dimensional feature representation of global interaction features; S4. A dense prediction network step, generating prediction information according to the high-dimensional features of the global interaction features; S5.Sim Transformer decoding step, used to output a specific behavior decision based on the prediction information combined with the Query; S6. Agent behavior output step, adjusting the actual operation of the agent according to the specific behavior decision.

2. According to claim 1, the method for interactive micro-behavior of autonomous driving simulation based on AI agent is characterized in that: The multi-source data includes: open source data, simulation data and roadside data.

3. The method for micro-behavior interaction of autonomous driving simulation based on AI agent according to claim 1, characterized in that: The format of the trajectory data is: [N,T h ,D t ], where N is the predicted number of vehicles, T h refers to the number of historical trajectory points, D t is the trajectory characteristic attribute; The format of the map data is [N,L c ,D m ], where L c refers to the number of road lines, D m Refers to the characteristic attributes of the map.

4. The method for interactive micro-behavior of autonomous driving simulation based on AI agent according to claim 1, characterized in that: The trajectory data and the map data are deeply integrated and embedded into a unified feature space by: X=φ k,s (g(F)) Implementation, where g represents the backbone network layer of the PointNet neural network model, φ k,s represents the maximum pooling layer, and F represents the joint feature matrix.

5. The method for interactive micro-behavior of autonomous driving simulation based on AI agent according to claim 1, characterized in that: The conversion of the multimodal feature vector into a high-dimensional feature representation of global interaction features is specifically achieved through a hierarchical sliding window method and a multi-head self-attention mechanism.

6. The method for interactive micro-behavior of autonomous driving simulation based on AI agent according to claim 1, characterized in that: The prediction information is obtained by: Get, where MLP represents the fully connected neural network layer, S i Represents the attention score of the input.

7. The method for interactive micro-behavior of autonomous driving simulation based on AI agent according to claim 1, characterized in that: The Query includes a location query and an intent query.

8. An AI-based autonomous driving simulation micro-behavior interaction system, characterized in that: The system includes: an input and preprocessing module, a fusion embedding module, a Sim Transformer encoder module, a dense prediction network module, a Sim Transformer decoder module and an agent behavior output module; The input and preprocessing module is used to preprocess the multi-source data to obtain trajectory data and map data and send them to the fusion embedding module; The fusion embedding module is used to deeply fuse the trajectory data and the map data and embed them into a unified feature space to obtain a multimodal feature vector and send it to the Sim Transformer encoder module; The Sim Transformer encoder module is used to convert the multimodal feature vector into a high-dimensional feature representation of a global interaction feature, and send the high-dimensional feature representation of the global interaction feature to the dense prediction network module; The dense prediction network module is used to generate prediction information according to the high-dimensional features of the global interaction features and send the prediction information to the Sim Transformer decoder module; The Sim Transformer decoder module is used to output a specific behavior decision based on the prediction information combined with the Query, and send the specific behavior decision to the agent behavior output module; The agent behavior output module is used to adjust the actual operation of the agent according to the specific behavior decision.

Images