Automatic driving simulation agent method

Abstract

The invention provides an automatic driving simulation agent method, and the method comprises the steps: obtaining bag data; screening a to-be-controlled vehicle as a target agent based on the bag packet data; main vehicle track change features, target agent state features, interaction relation features of the main vehicle and the target agent and environment constraint features of the target agent are calculated respectively; performing feature fusion on the main vehicle track change feature, the state feature of the target agent, the interaction relationship feature of the main vehicle and the target agent and the environmental constraint feature of the agent; and making a driving decision based on a feature fusion result. By means of the method, the simulation scene better fits the actual interaction situation.

Description

An autonomous driving simulation agent method Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to an autonomous driving simulation agent method. Background Technology

[0002] Autonomous driving technology has been under development for decades. In the early days, it relied on pre-programmed rules to make cars drive simply (such as adjusting the direction when the lane lines were off). Now, it has been upgraded to rely on intelligent algorithms and various sensors (cameras, radar, lidar, etc.) to allow cars to make autonomous judgments. The core of the current research is to enable cars to "see the road, think, and drive" like humans: "seeing the road" relies on cameras and lidar, which can see the surrounding cars, pedestrians, and traffic lights in 360 degrees. Moreover, lidar is becoming cheaper and more and more cars are being equipped with it; "thinking" relies on large AI models, which are fed with massive amounts of data from real-world driving, enabling cars to cope with complex situations such as traffic jams, sudden lane changes, and unexpected road crossings.

[0003] Current autonomous driving technologies are primarily focused on Level 2 "partial automation." This allows for simultaneous control of steering and acceleration / deceleration, enabling features like automatic following and lane centering on highways, and basic automatic lane avoidance and traffic light recognition in urban areas. However, the core prerequisite is that the driver must constantly monitor the road, keeping their hands on the steering wheel for extended periods. In complex scenarios such as heavy rain, snow, or rural roads without lane markings, the driver must take over the vehicle promptly. Hardware-wise, most rely on a combination of cameras and millimeter-wave radar to perceive the environment, with a few models adding low-cost LiDAR to improve accuracy. Technically, it primarily uses multi-sensor fusion and rule-based algorithms, but it cannot yet handle all emergencies as flexibly as a seasoned human driver. For example, human intervention may still be necessary when faced with pedestrians suddenly crossing the road or temporary road closures due to construction. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide an autonomous driving simulation agent method that makes the simulation scenario more closely resemble the actual interaction situation.

[0005] In a first aspect, embodiments of this application provide an autonomous driving simulation agent method, comprising: acquiring bag data; filtering vehicles to be controlled as target intelligent agents based on the bag data; calculating the trajectory change features of the master vehicle, the state features of the target intelligent agent itself, the interaction relationship features between the master vehicle and the target intelligent agent, and the environmental constraint features of the target intelligent agent; performing feature fusion on the trajectory change features of the master vehicle, the state features of the target intelligent agent itself, the interaction relationship features between the master vehicle and the target intelligent agent, and the environmental constraint features of the intelligent agent; and making driving decisions based on the feature fusion results.

[0006] In conjunction with the first aspect, this application provides a first possible implementation of the first aspect, wherein the bag data includes: main vehicle data, agent data, and environmental data; the main vehicle data includes any one or more of the following: real-time vehicle trajectory, speed, acceleration, and turn signal status; the agent data includes any one or more of the following: perception frames of surrounding vehicles, historical trajectories, speeds, and accelerations; and the environmental data includes any one or more of the following: road topology and traffic rules.

[0007] In conjunction with the first aspect, this application provides a second possible implementation of the first aspect, wherein the step of filtering the controlled vehicle as the target intelligent agent based on bag data includes: calculating the spatiotemporal influence factor between the intelligent agent and the master vehicle according to the following formula. :

[0008] in, The relative distance between the main vehicle and the intelligent agent i; The relative speed between the main vehicle and the intelligent agent i; This is the lane relationship factor, with 1 for lanes in the same lane, 0.5 for adjacent lanes, and 0 for others; The weight is used to select the target intelligent agent from all intelligent agents according to a preset screening threshold; the screening threshold is dynamically adjusted according to the current scene.

[0009] In conjunction with the first aspect, this application provides a third possible implementation of the first aspect, wherein calculating the main vehicle trajectory change features includes: inputting the main vehicle trajectory deviation sequence into a temporal convolutional network to obtain a main vehicle trajectory influence feature matrix output by the temporal convolutional network that expresses the main vehicle trajectory change features; the main vehicle trajectory deviation sequence is:

[0010] in, The adjusted trajectory; The original trajectory of the bag; the trajectory change rate is:

[0011] The convolutional kernel of the temporal convolutional network is kernel=(5,1), and the stride is stride=(1,1). The fully connected layers in the temporal convolutional network are used to map the input features into high-dimensional features.

[0012] in, These are learnable parameters.

[0013] In conjunction with the third possible implementation of the first aspect, this application provides a fourth possible implementation of the first aspect, wherein calculating the target agent's own state characteristics includes: inputting the historical motion parameters of the target agent i into a GRU network to obtain an inertial feature matrix output by the GRU network that expresses the target agent's own state characteristics; the historical motion parameters include: rate of change of velocity. trajectory curvature .

[0014] In conjunction with the fourth possible implementation of the first aspect, this application provides a fifth possible implementation of the first aspect, wherein the calculation of the interaction relationship features between the master vehicle and the target intelligent agent includes: constructing an interaction graph based on the relative distance between the target master vehicle and the target intelligent agent i, the relative speed between the master vehicle and the target intelligent agent i, and the relative distance between the target intelligent agents; and inputting the feature information in the interaction graph into a graph convolutional network to generate an interaction influence feature matrix that expresses the interaction relationship features between the master vehicle and the target intelligent agent.

[0015] In conjunction with the fourth possible implementation of the first aspect, this application provides a sixth possible implementation of the first aspect, wherein the calculation of the environmental constraint features of the target agent includes: inputting the distance between the target agent i and the lane line, the current road segment speed limit and the traffic light status into an FC network with fully connected layer mapping features to obtain the environmental constraint feature matrix of the target agent output by the FC network.

[0016] In conjunction with the first aspect, this application provides a seventh possible implementation of the first aspect, wherein the feature fusion of the master vehicle trajectory change features, the target intelligent agent's own state features, the interaction relationship features between the master vehicle and the target intelligent agent, and the environmental constraint features of the intelligent agent includes: sequentially fusing the target intelligent agent's own state features, the interaction relationship features between the master vehicle and the target intelligent agent, and the environmental constraint features of the intelligent agent into the master vehicle trajectory change features to obtain fused features. .

[0017] In conjunction with the seventh possible implementation of the first aspect, this application provides an eighth possible implementation of the first aspect, wherein the driving decision based on the feature fusion result includes: fusing features... The input is fed into the decision network to obtain the behavior probability distribution of the target agent; the driving strategy of the target agent is determined based on the behavior probability distribution of the target agent.

[0018] In conjunction with the eighth possible implementation of the first aspect, this application provides a ninth possible implementation of the first aspect, which further includes: determining the position of the target agent at the next moment in the following manner:

[0019] in, For time intervals, Let the decision direction vector be... For based on The amount of correction;

[0020] For correction factor, To correct the weight matrix.

[0021] The autonomous driving simulation agent method provided in this application adopts a multi-feature branch modeling approach, specifically using four features: the trajectory change features of the master vehicle, the state features of the target intelligent agent, the interaction relationship features between the master vehicle and the target intelligent agent, and the environmental constraint features of the target intelligent agent. This allows for more precise and comprehensive control of feature fusion and feature decision-making, making the simulation scenario more closely resemble actual interaction situations.

[0022] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 shows a flowchart of an autonomous driving simulation agent method provided in an embodiment of this application; Figure 2 shows a flowchart of the complete logic of an autonomous driving simulation agent method provided in an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0026] In the research and development of autonomous driving technology, Logsim simulation (simulation testing based on bag data collected from real vehicles) is a core means of verifying the safety and robustness of algorithms. Bag data contains key data such as perception information of the main vehicle and surrounding traffic participants (such as radar point clouds and camera images), motion parameters (speed, position, acceleration), and environmental parameters (road structure, traffic signs), providing a real-world scenario foundation for simulation.

[0027] In the existing Logsim simulation, the traffic participant intelligent agent model has the following problems: (1) Behavior solidification: The behavior of the intelligent agent based on the playback of the original data of the bag completely replicates the real scene. When the trajectory of the main vehicle changes due to algorithm adjustment (such as lane change, deceleration, emergency avoidance), the behavior of the surrounding traffic participants (intelligent agents) cannot respond dynamically and still travel along the original trajectory, resulting in the simulation scene being disconnected from reality (for example, after the main vehicle actively avoids, the side vehicle still travels along the original trajectory, causing a virtual collision, which in turn affects the simulation results).

[0028] (2) Lack of interaction: The existing model does not model the dynamic interaction logic between traffic participants, ignores the "behavioral guidance" of the main vehicle's trajectory changes to surrounding vehicles (such as when the main vehicle turns on its turn signal, the surrounding vehicles usually slow down and give way), and only judges obstacle avoidance by preset distance thresholds, lacking the game and cooperation mechanism in real roads.

[0029] (3) Insufficient feature utilization: The multi-source heterogeneous data (perceptual features, motion features, environmental features) in the bag are not effectively integrated. The trajectory generation relies only on a single location information, resulting in poor rationality of the agent's response to complex scenarios (such as the main vehicle cutting into the lane or sudden deceleration).

[0030] (4) Adaptability limitations: The different scenarios of different bag data (such as highways, urban areas, and congested road conditions) will lead to inconsistent behavior of the agent. There is a lack of a unified dynamic adjustment mechanism, and parameters need to be tuned separately for each scenario, resulting in high adaptation costs.

[0031] (5) Control redundancy: The surrounding vehicles are not dynamically screened, and the same control strategy is used for all vehicles, which leads to a waste of computing resources and may affect the accuracy of the simulation due to unnecessary responses from unrelated vehicles.

[0032] To address the aforementioned problems in existing technologies, this invention provides an autonomous driving simulation agent method. This method achieves deep feature fusion and multi-objective optimization through multi-feature branch modeling and dynamic response fusion mechanisms, and combines this with a dynamic screening mechanism to determine the controlled vehicle, thereby improving simulation efficiency and accuracy.

[0033] As shown in Figure 1, the autonomous driving simulation agent method includes the following steps S101-S105: S101: Obtain bag data; S102: Filter the vehicles to be controlled as target intelligent agents based on bag data; S103: Calculate the trajectory change features of the master vehicle, the state features of the target intelligent agent, the interaction relationship features between the master vehicle and the target intelligent agent, and the environmental constraint features of the target intelligent agent; S104: Perform feature fusion on the trajectory change features of the master vehicle, the state features of the target intelligent agent, the interaction relationship features between the master vehicle and the target intelligent agent, and the environmental constraint features of the intelligent agent; S105: Make driving decisions based on the feature fusion results.

[0034] Figure 2 shows a schematic diagram of the complete logic of this scheme.

[0035] The solution requires acquiring bag data collected during autonomous driving at the beginning. The following three types of data can be parsed and extracted from the bag data: vehicle data, agent data, and environmental data.

[0036] Main vehicle data includes any one or more of the following: real-time vehicle trajectory (location) ), Main vehicle speed (i.e., speed) Main vehicle acceleration (i.e., acceleration) Main vehicle turn signal status (turn signal status) (0 = Off, 1 = Left turn, 2 = Right turn).

[0037] The agent data includes any one or more of the following: the perception boxes (location information) of surrounding vehicles (the agent). The historical trajectory of each intelligent agent (k=1,2,...,T, T=30 frames), the velocity of each agent's historical trajectory Acceleration of each agent's historical trajectory Environmental data includes any one or more of the following: road topology (Lane line location, intersection structure), traffic rules (Speed ​​limit, yield sign).

[0038] Before performing step S102, preprocessing is required. The preprocessing steps include: time synchronization: unifying multi-source data to the same timestamp (accuracy ±10ms); coordinate transformation: transforming all location information to the world coordinate system, i.e., unifying the coordinate system (based on the positioning information and IMU calibration data in the bag); outlier cleaning: removing noise and retaining effective features through the 3σ criterion.

[0039] Step S102 mainly involves the selection of intelligent agents, specifically: first, calculating the spatiotemporal influence factors between each intelligent agent and the main vehicle. :

[0040] in, The relative distance between the main vehicle and the intelligent agent i; The relative speed between the main vehicle and the intelligent agent i; This is the lane relationship factor, with 1 for lanes in the same lane, 0.5 for adjacent lanes, and 0 for others; For weights.

[0041] Then set the filter threshold. (This value should be dynamically adjusted depending on the scenario; for example, in urban scenarios...) In high-speed scenarios Finally, the selection was made. The intelligent agents, acting as the controlled vehicles (target intelligent agents), are denoted as the set. .

[0042] As shown in step S103, in order to make subsequent policy judgments, four types of data need to be determined: the trajectory change characteristics of the main vehicle, the state characteristics of the target agent, the interaction relationship characteristics between the main vehicle and the target agent, and the environmental constraint characteristics of the target agent; among them, the main branch: the trajectory change characteristics of the main vehicle: the features of the input model are: the trajectory deviation sequence of the main vehicle. ;in, The adjusted trajectory; The original trajectory of the bag; the rate of change of the trajectory. The model output is: the feature matrix of the main vehicle trajectory influence. (T is the time step, C is the feature dimension); The processing flow for the main vehicle trajectory change features is as follows: A Temporal Convolutional Network (TCN) is used to capture the trajectory change trend, with a convolutional kernel kernel=(5,1) and a stride of stride=(1,1); it is then mapped to high-dimensional features through a fully connected layer:

[0043] in, These are learnable parameters.

[0044] Sub-branch 1: The features of the target agent's own state input model are: the historical motion parameters of target agent i. (k=1,...,T); The model output is: its own behavioral inertia feature matrix. Processing flow: Calculate the kinematic characteristics of each target agent: rate of change of velocity. and trajectory curvature .

[0045] Using a GRU network to model temporal dependencies, the output features are:

[0046] Sub-branch 2: Interaction relationship features between the main vehicle and the target agent: The feature of the input model is the relative distance between the main vehicle and the target agent i. The relative velocity of the target intelligent agent Relative distance between target intelligent agents (j represents other agents besides the target agent being computed); the model output is: the interaction influence feature matrix. Processing flow: Construct an interaction graph based on the relative distance between the target vehicle and target agent i, the relative speed between the vehicle and target agent i, and the relative distance between the target agents. Node V is an agent, and the edge weights are... ( (The closer the distance, the greater the weight); aggregate interactive information through graph convolutional networks (GCNs):

[0047] That is, the feature information in the interaction graph is input into a graph convolutional network (GCN) to generate an interaction influence feature matrix that expresses the interaction relationship features between the host vehicle and the target intelligent agent.

[0048] Sub-branch 3: Environmental constraint features. The input model features are: the distance between the target agent i and the lane line. Current road section is restricted Traffic light status The model output is: environmental constraint feature matrix. Processing flow: Standardize environmental parameters to the [0,1] range: ( (For lane width); the fully connected layer mapping features of the model in this step:

[0049] In the above content, the feature fusion in the scheme uses element-wise multiplication (⊙) instead of addition. The characteristic of multiplication is that "the feature multiplied first will become the 'base' of the feature multiplied later" - the feature with high priority (such as the main vehicle intention) is integrated first and will become the "reference benchmark" for all subsequent features. Subsequent features can only be adjusted on this benchmark, rather than overturned.

[0050] A fixed order is a prerequisite for the "implicit threshold" to take effect, thus avoiding meaningless responses.

[0051] One of the core objectives of the phased integration approach is to construct an "implicit response threshold"—only when the master vehicle's intent is strong enough ( High) + Self-response ( High) + Interaction Allowed ( High) + Environmental Compliance high), Only when the value is high enough will a valid response be triggered; if a certain feature does not meet the requirements (e.g., the environment is a red light), Low), It will be lowered to suppress unreasonable responses.

[0052] This "threshold effect" relies entirely on a fixed sequence: the main vehicle's intent is the "start threshold" (first determine whether to respond), its own state is the "physical threshold" (then determine whether to respond), the interaction is the "safety threshold" (then determine whether to crash), and the environment is the "rule threshold" (finally determine whether it complies with regulations) - after layer-by-layer screening, only responses that meet all dimensions will be triggered.

[0053] If the order is disordered (e.g., merging with the interaction before merging with itself): This may result in the message "The interaction allows avoidance, but the vehicle itself cannot physically slow down (e.g., a heavily loaded vehicle)". It will still calculate a high value, triggering an unreasonable decision to "brake suddenly to avoid an obstacle," which violates the laws of physics.

[0054] And, a fixed order guarantees Its interpretability facilitates debugging and deployment.

[0055] Autonomous driving simulation algorithms must not only "make decisions," but also "explain why those decisions are made" (a core requirement for engineering implementation), with a fixed fusion sequence. The numerical changes can be traced back: if A low value indicates the agent failed to avoid the obstacle; this can be investigated by examining the driver's intent characteristics. Is the low lane change signal due to inherent characteristics (or is it due to the vehicle's own condition)? Low (vehicle cannot slow down) or environmental characteristics ( Low (red light constraint).

[0056] If the fusion order is disordered The value is the result of randomly multiplying all features, making it impossible to pinpoint "which constraint level caused the decision anomaly." For example, the same... =0.5, which may be due to weak intentions from the main vehicle or strong environmental constraints. Algorithm debugging will become "blind debugging" and completely lose its engineering value.

[0057] The above describes the calculation process of four data points: the trajectory change characteristics of the master vehicle, the state characteristics of the target agent, the interaction characteristics between the master vehicle and the target agent, and the environmental constraint characteristics of the target agent. The following explains the feature fusion and decision-making process: Feature Fusion Mechanism: As shown in Figure 2, a progressive enhancement fusion strategy is adopted. The main branch features and sub-branch features are multiplied element-wise, and the main branch features are dynamically updated: Initial Features = First fusion: fusing with its own state. =F0⊙( +1) (⊙ indicates element-wise multiplication, +1 avoids feature annihilation); Second fusion: Fusion interaction relationship: = ⊙( +1); Third fusion: Fusion environment constraints: = ⊙( +1).

[0058] This fusion mechanism constructs an implicit response threshold through a multiplication operation: when the trajectory of the master vehicle changes significantly ( (High value) and the intelligent agent is close to the main vehicle ( When the value is high, the fusion features Trigger a strong response; if environmental constraints are strict (such as when the light is red). (A low value) will inhibit irrational behavior.

[0059] Driving Decision and Trajectory Correction: Decision Network: Input The feature vectors after global average pooling are then passed through a fully connected layer to output the probability distribution of the target agent's behavior. (Such as acceleration, deceleration, steering, and holding):

[0060] Trajectory correction formula: Update the agent's position at the next moment based on the decision result:

[0061] in, =0.1s is the time interval. Let the decision direction vector be... For based on Correction amount:

[0062] =0.05 is the correction factor. To correct the weight matrix.

[0063] The fused temporal feature matrix T×C undergoes global pooling or dimensionality reduction to obtain a 1×C comprehensive feature matrix. This comprehensive feature matrix generates a behavior probability distribution through a fully connected layer, supporting the agent's trajectory correction and behavior decision-making. Through progressive multiplication and fusion, the feature matrix gradually incorporates the agent's own state, interaction relationships, and environmental constraints, evolving from a single primary influence feature matrix to one that incorporates its own state, interaction relationships, and environmental constraints, achieving precise control of dynamic responses.

[0064] The following is a specific example to illustrate the solution of this application: The bag data includes: an urban road scene, the main vehicle's original trajectory traveling along a straight lane, and agent A (the adjacent vehicle) located in the right lane of the main vehicle, with an initial distance... =5m, speed =30km / h; Main vehicle trajectory change: At t=1s, the main vehicle activates its right turn signal; at t=2s, it begins to change lanes to the right, resulting in trajectory deviation. Characteristic response: main branch The characteristic value of "turn signal + trajectory change" reaches 0.8 (triggering a strong response); sub-branch 2 medium relative distance =3m, eigenvalue reaches 0.75; after fusion Triggering the "decelerate and avoid" decision, probability P=0.92; Behavioral output: Agent A decelerates from t=2.3s, acceleration... At t=3s, the speed drops to 20km / h, yielding to the main vehicle. The trajectory correction amount... (Slightly adjust to the right to increase the lateral distance).

[0065] The core content of the solution provided in this application includes the following three points: 1. Multi-branch feature extraction architecture: the main branch extracts the trajectory change features of the main vehicle, and the sub-branches extract the features of the agent's own state, interaction relationship, and environmental constraints, and perform deep fusion; 2. Gradual enhancement fusion mechanism: dynamic feature update is achieved by multiplying the features of the main branch and sub-branch element by element, and implicit response threshold is constructed; 3. Behavioral decision and trajectory correction method based on fused features: behavioral decisions are output by combining probability distribution, and the agent's trajectory is adjusted in real time by correction formula.

[0066] The solution provided in this application has the following beneficial effects: (1) Improved dynamic response accuracy: When the trajectory of the main vehicle changes, the matching degree between the intelligent agent's behavior response and human driving behavior is significantly improved, the virtual collision rate in the simulation system is reduced, and the simulation scene is more in line with the actual interaction situation; (2) Enhanced interaction realism: Through multi-feature fusion modeling, the intelligent agent can simulate complex interactive behaviors such as "the main vehicle's turn signal induces the adjacent vehicle to decelerate" and "the main vehicle cuts in", which improves the simulation scene restoration and more realistically reproduces the game and collaboration mechanism in the road; (3) Scene adaptability expansion: The unified model can adapt to various typical scene bags such as highways, urban areas, and congestion, without the need to adjust parameters separately for each scene, which significantly improves the adaptation efficiency and reduces the scene adaptation cost; (4) Optimized control efficiency: Through three-layer optimization of dynamic filtering to reduce objects, sharing features to reduce redundancy, and lightweight fusion to support parallelism, while realizing independent modeling of multiple intelligent agents, resource consumption is effectively controlled, and the impact of unnecessary responses of irrelevant vehicles on the simulation accuracy is reduced.

[0067] Compared to patent application number CN202510011835.5, which is an AI agent interaction strategy for autonomous driving simulation based on Sim Transformer encoding, this solution requires a large amount of data for training and is based on a Transformer model, making it difficult to achieve real-time simulation. Compared to patent application number CN202510471420.6, which constructs a dynamic driving model based on image and measurement data to simulate the driving decision-making process of a background vehicle; and jointly trains the background vehicle and the test vehicle using a multi-agent reinforcement learning algorithm, this solution also relies on image data and requires a large amount of training data to train the dynamic driving model, making real-time simulation difficult.

[0068] Compared to the patent with application number CN201911172410.3, this solution analyzes the car-following characteristics of autonomous vehicles, constructs a car-following model, analyzes the lane-changing characteristics of autonomous vehicles, and compares lane-changing simulation with the lane-changing effects of manually driven vehicles. It incorporates factors influencing lane changing, lane-changing characteristics, and the characteristics of the autonomous driving simulation platform, introducing speed tolerance and space allowance to constrain lane-changing behavior and construct a matching model. It also confirms the lane-changing space and constructs a lane-changing preparatory model. After a lane-changing action occurs, based on the actual size of the current lane-changing space and considering the actual information processing latency, it constructs a lane-changing speed control model based on a two-lane forward-driving vehicle and a trajectory optimization model based on lane-changing space. This method relies on traditional rules to control the agent, which, while possessing a certain degree of controllability and interpretability, lacks human-like behavior and requires separate adaptation for different scenarios, making it difficult to adapt to various traffic scenarios and exhibiting poor generalization.

[0069] The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and other division methods may be used in actual implementation. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0070] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0071] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0072] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0073] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

Claims

1. An autonomous driving simulation agent method, characterized in that, include: Retrieve data from the bag; Based on the data in the bag, the vehicles to be controlled are selected as target intelligent agents; The characteristics of the master vehicle's trajectory change, the target agent's own state, the interaction relationship between the master vehicle and the target agent, and the environmental constraints of the target agent are calculated separately; the characteristics of the master vehicle's trajectory change, the target agent's own state, the interaction relationship between the master vehicle and the target agent, and the environmental constraints of the agent are then fused. Driving decisions are made based on feature fusion results.

2. The method according to claim 1, characterized in that, The bag data includes: main vehicle data, agent data, and environmental data; the main vehicle data includes any one or more of the following: real-time vehicle trajectory, speed, acceleration, and turn signal status; the agent data includes any one or more of the following: perception boxes of surrounding vehicles, historical trajectories, speed, and acceleration; the environmental data includes any one or more of the following: road topology and traffic rules.

3. The method according to claim 1, characterized in that, The step of selecting the controlled vehicle as the target intelligent agent based on bag data includes: calculating the spatiotemporal influence factor between the intelligent agent and the master vehicle according to the following formula. ： in, The relative distance between the main vehicle and the intelligent agent i; The relative speed between the main vehicle and the intelligent agent i; This is the lane relationship factor, with 1 for lanes in the same lane, 0.5 for adjacent lanes, and 0 for others; The weight is used to select the target intelligent agent from all intelligent agents according to a preset screening threshold; the screening threshold is dynamically adjusted according to the current scene.

4. The method according to claim 1, characterized in that, The calculation of the main vehicle trajectory change features includes: inputting the main vehicle trajectory deviation sequence into a temporal convolutional network to obtain the main vehicle trajectory influence feature matrix output by the temporal convolutional network, which expresses the main vehicle trajectory change features; the main vehicle trajectory deviation sequence is: in, The adjusted trajectory; The original trajectory of the bag; the trajectory change rate is: The convolutional kernel of the temporal convolutional network is kernel=(5,1), and the stride is stride=(1,1). The fully connected layers in the temporal convolutional network are used to map the input features into high-dimensional features. in, These are learnable parameters.

5. The method according to claim 4, characterized in that, The calculation of the target agent's own state features includes: inputting the historical motion parameters of the target agent i into the GRU network to obtain the inertial feature matrix output by the GRU network that expresses the target agent's own state features; the historical motion parameters include: rate of change of velocity. trajectory curvature 。 6. The method according to claim 5, characterized in that, The calculation of the interaction relationship features between the master vehicle and the target intelligent agent includes: constructing an interaction graph based on the relative distance between the target master vehicle and the target intelligent agent i, the relative speed between the master vehicle and the target intelligent agent i, and the relative distance between the target intelligent agents; and inputting the feature information in the interaction graph into a graph convolutional network to generate an interaction influence feature matrix that expresses the interaction relationship features between the master vehicle and the target intelligent agent.

7. The method according to claim 5, characterized in that, The calculation of the environmental constraint features of the target agent includes: inputting the distance between the target agent i and the lane line, the current road segment speed limit and the traffic light status into an FC network with fully connected layer mapping features, so as to obtain the environmental constraint feature matrix of the target agent output by the FC network.

8. The method according to claim 1, characterized in that, The feature fusion process, which involves fusing the main vehicle trajectory change features, the target agent's own state features, the interaction relationship features between the main vehicle and the target agent, and the agent's environmental constraint features, includes: sequentially fusing the target agent's own state features, the interaction relationship features between the main vehicle and the target agent, and the agent's environmental constraint features into the main vehicle trajectory change features to obtain fused features. 。 9. The method according to claim 8, characterized in that, The driving decision based on feature fusion results includes: fusing features The input is fed into the decision network to obtain the behavior probability distribution of the target agent; the driving strategy of the target agent is determined based on the behavior probability distribution of the target agent.

10. The method according to claim 9, characterized in that, It also includes: determining the target agent's position at the next moment in the following manner: in, For time intervals, Let the decision direction vector be... For based on The amount of correction; For correction factor, To correct the weight matrix.

Images