Traffic participant behavior information prediction method and related device

Abstract

The embodiment of the invention provides a traffic participant behavior information prediction method and a related device. The method comprises the following steps: acquiring first behavior information corresponding to at least two types of traffic participants in a preset range; for any type of traffic participants in the at least two types of traffic participants, feature extraction is carried out on the first behavior information corresponding to any type of traffic participants to obtain first behavior features corresponding to any type of traffic participants, and behavior features of the first behavior features corresponding to different types of traffic participants are different; for any type of traffic participants, using a prediction model corresponding to any type of traffic participants to perform prediction based on the first behavior characteristics corresponding to any type of traffic participants to obtain second behavior information corresponding to any type of traffic participants, the second behavior information comprising behavior information after the first behavior information, therefore, the efficiency of behavior information prediction can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, and more particularly, to a behavior information prediction method of a traffic participant and related devices. BACKGROUND

[0002] With the rapid development of automatic driving, the related technologies of automatic driving also develop rapidly, for example, the simulation test of automatic driving also develops rapidly. In the related technologies, the input behavior information can be used for prediction to obtain predicted behavior information, and then the predicted behavior information is used for simulation test.

[0003] However, the inventors have found through long-term research that the predicted behavior information obtained by using the input behavior information for prediction is inaccurate. SUMMARY

[0004] The embodiments of the present application propose a behavior information prediction method of a traffic participant and related devices, which can improve the prediction accuracy of behavior information.

[0005] In a first aspect, the embodiments of the present application propose a behavior information prediction method of a traffic participant, comprising: obtaining first behavior information corresponding to at least two types of traffic participants in a preset range respectively; for any one type of traffic participant in the at least two types of traffic participants, performing feature extraction on the first behavior information corresponding to the any one type of traffic participant to obtain first behavior features corresponding to the any one type of traffic participant, wherein the behavior characteristics of the first behavior features corresponding to different types of traffic participants are different; for the any one type of traffic participant, using a prediction model corresponding to the any one type of traffic participant to perform prediction based on the first behavior features corresponding to the any one type of traffic participant to obtain second behavior information corresponding to the any one type of traffic participant, the second behavior information including behavior information after the first behavior information.

[0006] In the embodiment, the first behavior information corresponding to each of the at least two types of traffic participants in a preset range is acquired; for any one type of traffic participant in the at least two types of traffic participants, the first behavior information corresponding to the any one type of traffic participant is subjected to feature extraction to obtain the first behavior feature corresponding to the any one type of traffic participant, wherein the behavior characteristics of the first behavior features corresponding to different types of traffic participants are different; for the any one type of traffic participant, a prediction model corresponding to the any one type of traffic participant is used to perform prediction based on the first behavior feature corresponding to the any one type of traffic participant to obtain the second behavior information corresponding to the any one type of traffic participant, the second behavior information including behavior information after the first behavior information. Since the behavior characteristics of the first behavior features corresponding to different types of traffic participants are different, that is, the first behavior feature extraction is performed on the first behavior features of different types of traffic participants, the extraction accuracy of the first behavior features can be improved, and the accuracy of the behavior information prediction can be improved. Moreover, the prediction model corresponding to the any one type of traffic participant is used to perform prediction based on the first behavior feature corresponding to the any one type of traffic participant to obtain the second behavior information corresponding to the any one type of traffic participant, so that the feature extraction of different types of traffic participants can be performed synchronously, and the efficiency of the behavior information prediction can be improved.

[0007] In a possible implementation, the prediction based on the first behavior feature corresponding to the any one type of traffic participant by using the prediction model corresponding to the any one type of traffic participant to obtain the second behavior information corresponding to the any one type of traffic participant includes: the prediction based on the first behavior feature corresponding to the any one type of traffic participant by using the prediction model corresponding to the any one type of traffic participant to obtain the second behavior feature corresponding to the any one type of traffic participant, wherein the behavior characteristics of the second behavior features corresponding to different types of traffic participants are different; and the second behavior information corresponding to the any one type of traffic participant is determined based on the second behavior feature corresponding to the any one type of traffic participant.

[0008] In a possible implementation, the first behavior feature and the second behavior feature corresponding to the any one type of traffic participant include a motion feature, an interaction feature, and a decision feature, the motion feature is used to represent a motion condition of the any one type of traffic participant, the interaction feature is used to represent an interaction condition between the any one type of traffic participant and other traffic participants, the other traffic participants including participants in the preset range except the any one type of traffic participant, and the decision feature is used to represent an intention of the any one type of traffic participant.

[0009] In a possible implementation, the at least two types of traffic participants include at least two of motor vehicles, pedestrians, and non-motor vehicles; the motion feature corresponding to the motor vehicles includes at least one of instantaneous speed, instantaneous acceleration, jerk, and a travel direction angle, the interaction feature corresponding to the motor vehicles includes at least one of a distance from a preceding vehicle and a collision time, the collision time being used to represent an interval time required for the motor vehicle to collide with the preceding vehicle, the preceding vehicle including a vehicle located in a travel direction of the motor vehicle; the motion feature corresponding to the pedestrians includes at least one of walking speed, step length, and step frequency, the interaction feature corresponding to the pedestrians includes at least one of a distance from the motor vehicle and an avoidance probability, the avoidance probability being used to represent a possibility of avoidance between the pedestrians and the motor vehicle; the motion feature corresponding to the non-motor vehicles includes at least one of travel speed, acceleration, and an inclination angle, the interaction feature corresponding to the non-motor vehicles includes at least one of a distance from the motor vehicle and a distance from the pedestrians; the decision feature corresponding to the motor vehicles includes a lane-changing intention probability of the motor vehicles, the decision feature corresponding to the pedestrians includes at least one of a crossing selection at an intersection and a pause time, and the decision feature corresponding to the non-motor vehicles includes at least one of a degree of lane deviation of the non-motor vehicles and a violation probability.

[0010] In a possible implementation, the method further includes: for any type of traffic participant, obtaining a plurality of training samples corresponding to the any type of traffic participant, any training sample including third behavior information and fourth behavior information corresponding to the any type of traffic participant, the fourth behavior information including behavior information after the third behavior information; for any type of traffic participant, calling a to-be-trained model corresponding to the any type of traffic participant; and for any type of traffic participant, training the to-be-trained model corresponding to the any type of traffic participant by using the plurality of training samples corresponding to the any type of traffic participant, to obtain a prediction model corresponding to the any type of traffic participant.

[0011] In one possible implementation, a training model for any type of traffic participant is trained using multiple training samples corresponding to any type of traffic participant to obtain a prediction model for that type of traffic participant. This includes: extracting features from the third behavior information of any type of traffic participant to obtain third behavior features, wherein the behavioral characteristics of the third behavior features differ for different types of traffic participants; using the third behavior features of any type of traffic participant as input to the training model for that type of traffic participant to obtain the predicted behavior information output by the training model for that type of traffic participant; determining the training loss of the training model for that type of traffic participant using the fourth behavior information and the corresponding predicted behavior information; weighting the training losses of the training models for at least two types of traffic participants to obtain a target training loss; if the target training loss meets the training termination condition, then the training model for that type of traffic participant is used as the prediction model for that type of traffic participant; if the target training loss does not meet the training termination condition, then the training model for that type of traffic participant continues to be trained using multiple training samples corresponding to any type of traffic participant until the target training loss meets the training termination condition.

[0012] In one possible implementation, the third behavioral feature corresponding to any type of traffic participant is used as the input to the training model corresponding to any type of traffic participant to obtain the predicted behavioral information of the output of the training model corresponding to any type of traffic participant. This includes: using the third behavioral feature corresponding to any type of traffic participant as the input to the training model corresponding to any type of traffic participant, so as to make a prediction based on the third behavioral feature corresponding to any type of traffic participant to obtain a fourth behavioral feature, wherein the behavioral characteristics of the fourth behavioral feature corresponding to different types of traffic participants are different; and determining the predicted behavioral information corresponding to any type of traffic participant based on the fourth behavioral feature corresponding to any type of traffic participant.

[0013] Secondly, embodiments of this application propose a traffic participant behavior information prediction device, comprising: an acquisition module, configured to acquire first behavior information corresponding to at least two types of traffic participants within a preset range; a feature extraction module, configured to extract features from the first behavior information corresponding to any one of the at least two types of traffic participants to obtain first behavior features corresponding to any one type of traffic participant, wherein the behavior characteristics of the first behavior features corresponding to different types of traffic participants are different; and a prediction module, configured to predict, based on the first behavior features corresponding to any one type of traffic participant using a prediction model corresponding to any one type of traffic participant, to obtain second behavior information corresponding to any one type of traffic participant, wherein the second behavior information includes behavior information following the first behavior information.

[0014] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory, wherein: the memory is used to store computer programs; and the processor is used to execute the programs stored in the memory to implement the above-described method.

[0015] Fourthly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating an application scenario of a method for predicting the behavior information of traffic participants, as shown in an embodiment of this application.

[0017] Figure 2 This is a flowchart illustrating a method for predicting the behavior information of traffic participants, as shown in one embodiment of this application.

[0018] Figure 3 This is a schematic diagram of a behavioral feature system provided in an embodiment of this application.

[0019] Figure 4 This is a diagram illustrating a machine learning modeling and scene transformation architecture in an embodiment of this application.

[0020] Figure 5 This is a flowchart illustrating a conversion quality assessment and optimization process according to an embodiment of this application.

[0021] Figure 6 This is a flowchart illustrating an output simulation scene according to an embodiment of this application.

[0022] Figure 7 This is a schematic diagram of the structure of a traffic participant behavior information prediction device provided in an embodiment of this application.

[0023] Figure 8This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0024] To make the technical problems, technical solutions, and beneficial effects solved by this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0025] With the rapid development of autonomous driving, related technologies are also advancing rapidly, such as autonomous driving simulation testing. These technologies utilize input behavioral information for prediction, yielding predicted behavioral information, which is then used for simulation testing. In the field of autonomous driving simulation testing technology, as autonomous driving technology advances to higher levels, the requirements for the realism and complexity of simulation scenarios are becoming increasingly stringent.

[0026] However, the inventors' long-term research revealed that related technologies lack sufficient specificity and refinement when extracting primary behavioral features from different traffic participants (e.g., vehicles, pedestrians, and two-wheeled vehicles). Most methods employ general feature extraction rules, failing to fully consider the unique behavioral characteristics of various participants. Furthermore, there is a lack of comprehensive and effective quantitative evaluation indexes in the process of converting real-world scenarios into simulation scenarios. In addition, existing simulation scenario optimization processes largely rely on manual experience to adjust parameters, and have not yet formed an automated and intelligent optimization loop.

[0027] In view of this, embodiments of this application propose a method and related apparatus for predicting the behavior information of traffic participants, which can improve the accuracy of behavior information prediction.

[0028] Please see Figure 1 , Figure 1 This is a schematic diagram illustrating an application scenario of a method for predicting the behavior information of traffic participants, as shown in an embodiment of this application. Figure 1The application scenario shown includes a vehicle 110 and a server 120. Specifically, the vehicle 110 sends first behavioral information corresponding to at least two types of traffic participants within a preset range to the server 120. The server can then obtain the first behavioral information corresponding to each of the at least two types of traffic participants. Next, for any one of the at least two types of traffic participants, features are extracted from the first behavioral information corresponding to that type of traffic participant to obtain first behavioral features for that type of traffic participant. The behavioral characteristics of the first behavioral features differ between different types of traffic participants. Then, for any one type of traffic participant, a prediction model is used based on the first behavioral features to predict the second behavioral information corresponding to that type of traffic participant. The second behavioral information includes behavioral information following the first behavioral information.

[0029] In this embodiment, server 120 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. Server 120 can also be referred to as the cloud.

[0030] It should be understood that vehicle 110 may also utilize the first behavioral information corresponding to at least two types of traffic participants within a preset range; then, for any one of the at least two types of traffic participants, feature extraction is performed on the first behavioral information corresponding to that type of traffic participant to obtain the first behavioral features corresponding to that type of traffic participant, wherein the behavioral characteristics of the first behavioral features corresponding to different types of traffic participants are different; then, for any one type of traffic participant, a prediction model corresponding to that type of traffic participant is used to predict based on the first behavioral features corresponding to that type of traffic participant to obtain the second behavioral information corresponding to that type of traffic participant, the second behavioral information including the behavioral information following the first behavioral information, thus eliminating the need for the cooperation of server 120.

[0031] It should be noted that this embodiment can be applied to simulation testing, autonomous driving decision-making during driving, and other situations requiring behavioral information prediction, without limitation. Taking simulation testing as an example, first behavioral information can be input, and then second behavioral information can be obtained through the scheme of this embodiment, which can then be used as the simulation result for autonomous driving. Taking autonomous driving decision-making as another example, first behavioral information can be input, and then second behavioral information can be obtained through the scheme of this embodiment, which can then be used for driving control.

[0032] Please see Figure 2 , Figure 2 This is a flowchart illustrating a method for predicting the behavior information of traffic participants, as shown in one embodiment of this application. Figure 2 The method shown can be performed by an electronic device, such as a vehicle or a server, as... Figure 2 The methods shown may include: S210. Obtain the first behavior information corresponding to at least two types of traffic participants within a preset range.

[0033] The first behavioral information can be manually input information or information collected from traffic participants within a preset range. In this embodiment, the first behavioral information can represent the behavior of traffic participants, such as their behavior within a certain time period, which can be a period preceding the current time. "At least two types of traffic participants" can refer to different types of traffic participants, such as categorization by motor vehicles and non-motor vehicles, or by categorization by vehicles (including motor vehicles and non-motor vehicles) and non-vehicles (e.g., pedestrians), etc., without limitation. For example, assuming the preset range includes motor vehicles, non-motor vehicles, and pedestrians, it can be the first behavioral information of motor vehicles, non-motor vehicles, and pedestrians detected by a motor vehicle detection device. For example, it could be the first behavioral information of the motor vehicle itself detected by the speed sensor and pose sensor in the motor vehicle detection device, and the first behavioral information of non-motor vehicles and pedestrians detected by the lidar in the motor vehicle detection device, without limitation.

[0034] S220. For any one of at least two types of traffic participants, feature extraction is performed on the first behavior information corresponding to any one type of traffic participant to obtain the first behavior feature corresponding to any one type of traffic participant, wherein the behavior characteristics of the first behavior feature corresponding to different types of traffic participants are different.

[0035] In this embodiment, feature extraction can be performed on the first behavioral information corresponding to any type of traffic participant to obtain the first behavioral features corresponding to any type of traffic participant. However, the behavioral characteristics of the first behavioral features corresponding to different types of traffic participants are different. In this embodiment, the behavioral characteristics are related to the type of traffic participant. For example, the behavioral characteristics of motor vehicles may include forward movement, backward movement, acceleration, deceleration, lane changing, and turning. The behavioral characteristics of non-motorized vehicles may include forward movement, backward movement, acceleration, deceleration, lane changing, turning, and tilting. The behavioral characteristics of pedestrians may include walking speed, cadence, and stride length, etc., without limitation.

[0036] S230. For any type of traffic participant, use the prediction model corresponding to any type of traffic participant to predict based on the first behavioral feature corresponding to any type of traffic participant, and obtain the second behavioral information corresponding to any type of traffic participant. The second behavioral information includes the behavioral information after the first behavioral information.

[0037] In this embodiment, a prediction model corresponding to any type of traffic participant can be used to predict the second behavior information corresponding to any type of traffic participant based on the first behavior feature of that type of traffic participant. That is, the prediction models for different types of traffic participants can be different. For example, the prediction models for motor vehicles, non-motor vehicles, and pedestrians can all be different. The second behavior information includes behavior information following the first behavior information; it can be understood as the future behavior information of the first behavior information. In this embodiment, the second behavior information can be used as the simulation result of a simulation test or as a reference for autonomous driving decisions.

[0038] In this embodiment, first behavioral information corresponding to at least two types of traffic participants within a preset range is acquired. For any one of the at least two types of traffic participants, feature extraction is performed on the first behavioral information corresponding to that type of traffic participant to obtain first behavioral features corresponding to that type of traffic participant. The behavioral characteristics of the first behavioral features corresponding to different types of traffic participants are different. For any one type of traffic participant, a prediction model corresponding to that type of traffic participant is used to predict based on the first behavioral features, to obtain second behavioral information corresponding to that type of traffic participant. The second behavioral information includes behavioral information following the first behavioral information. Since the behavioral characteristics of the first behavioral features corresponding to different types of traffic participants are different, i.e., first behavioral features are extracted for different types of traffic participants, the accuracy of first behavioral feature extraction can be improved, thereby improving the accuracy of behavioral information prediction. Furthermore, by using a prediction model corresponding to any one type of traffic participant to predict based on the first behavioral features, the second behavioral information corresponding to that type of traffic participant can be obtained. This allows feature extraction for different types of traffic participants to be performed simultaneously, thereby improving the efficiency of behavioral information prediction.

[0039] In one possible implementation, the first behavioral feature corresponding to any type of traffic participant includes motion features, interaction features, and decision features. Motion features are used to represent the motion of any type of traffic participant, interaction features are used to represent the interaction between any type of traffic participant and other traffic participants, which include participants other than any type of traffic participant within a preset range, and decision features are used to represent the intention of any type of traffic participant.

[0040] The first behavioral characteristic includes motion characteristics, interaction characteristics, and decision-making characteristics, which can be understood as different types of traffic participants having different motion characteristics, interaction characteristics, and decision-making characteristics. Other traffic participants include participants other than any type of traffic participant within the preset range. For example, assuming the preset range includes motor vehicle 1, motor vehicle 2, non-motorized vehicles, and pedestrians, then for motor vehicle 1, other traffic participants may include, but are not limited to, motor vehicle 2, non-motorized vehicles, and pedestrians; for motor vehicle 2, other traffic participants may include, but are not limited to, motor vehicle 1, non-motorized vehicles, and pedestrians; for non-motorized vehicles, motor vehicle 1, motor vehicle 2, and pedestrians are other traffic participants; and for pedestrians, non-motorized vehicles, motor vehicle 1, and motor vehicle 2 can all be considered other traffic participants, without restriction.

[0041] In this embodiment, the first and second behavioral features corresponding to any type of traffic participant include motion features, interaction features, and decision features. Motion features represent the motion of any type of traffic participant, interaction features represent the interaction between any type of traffic participant and other traffic participants, which include participants other than any type of traffic participant within a preset range. Decision features represent the intention of any type of traffic participant. Thus, the motion features, interaction features, and decision features of different types of traffic participants are different. Therefore, different types of traffic participants can be differentiated based on their behavioral characteristics from multiple aspects such as motion features, interaction features, and decision features, thereby improving the accuracy of predicting behavioral information.

[0042] In another possible implementation, the first behavioral feature corresponding to any type of traffic participant includes motion features, interaction features, or decision features, which can improve the efficiency of feature extraction and thus improve the prediction efficiency of behavioral information.

[0043] In one possible implementation, at least two types of traffic participants include at least two of motor vehicles, pedestrians, and non-motorized vehicles; The motion characteristics corresponding to motor vehicles include at least one of instantaneous speed, instantaneous acceleration, jerk, and driving direction angle; the interaction characteristics corresponding to motor vehicles include at least one of distance to the vehicle in front and collision time, where collision time represents the interval required for a collision between the motor vehicle and the vehicle in front, and the vehicle in front includes vehicles located in the direction of travel of the motor vehicle; the motion characteristics corresponding to pedestrians include at least one of walking speed, stride length, and stride frequency; the interaction characteristics corresponding to pedestrians include at least one of distance to motor vehicles and avoidance probability, where avoidance probability represents the possibility of avoidance between pedestrians and motor vehicles; the motion characteristics corresponding to non-motorized vehicles include at least one of driving speed, acceleration, and tilt angle; the interaction characteristics corresponding to non-motorized vehicles include at least one of distance to motor vehicles and distance to pedestrians; the decision characteristics corresponding to motor vehicles include the probability of lane-changing intent of motor vehicles; the decision characteristics corresponding to pedestrians include at least one of intersection crossing selection and stopping time; the decision characteristics corresponding to non-motorized vehicles include at least one of the degree of lane deviation of non-motorized vehicles and violation probability.

[0044] Instantaneous speed is defined as the speed of a motor vehicle along its current direction of travel at a specific instant (such as the instant it enters a pre-defined area of ​​an intersection or the instant the brakes are applied), typically measured in kilometers per hour (km / h) or meters per second (m / s). Instantaneous acceleration is defined as the rate of change of a motor vehicle's speed with time at a specific instant, typically measured in meters per second. 2 (m / s) 2 Positive values ​​indicate acceleration, and negative values ​​indicate deceleration (i.e., braking acceleration). Acceleration is defined as the rate of change of a vehicle's acceleration over time at a given instant, typically measured in meters per second. 3 (m / s) 3 Positive values ​​indicate increased acceleration (e.g., sudden acceleration), while negative values ​​indicate decreased acceleration (e.g., reduced braking force). The driving direction angle is defined as the angle between the tangent direction of the vehicle's current trajectory and a preset reference direction (e.g., due north, the center line of the road), usually expressed in degrees (°). The distance to the vehicle in front is defined as the straight-line distance between the front of the vehicle and the rear of the vehicle in front, usually expressed in meters (m), measured within a "preset safe range" (e.g., within 50 meters when following another vehicle, within 100 meters at an intersection). Under current driving conditions (within constant instantaneous and relative speeds of both vehicles), if neither takes any evasive action, the time interval required for a collision between the two vehicles is typically expressed in seconds (s). Example: The calculation formula is "distance to the vehicle in front". (Instantaneous speed of this vehicle - instantaneous speed of the vehicle in front) (when the speed of this vehicle is greater than that of the vehicle in front); In such cases, it is generally considered necessary to take immediate braking or evasive action.

[0045] The following definitions apply to walking speed: Walking speed is defined as the average speed of a pedestrian in the direction of travel at a given moment or over a short distance (such as the initial stage of crossing a road), typically measured in meters per second (m / s) or kilometers per hour (km / h). Stride length is defined as the straight-line distance between the heels of the back and front feet in two adjacent steps, typically measured in centimeters (cm) or meters (m). Step frequency is defined as the number of steps a pedestrian takes per unit of time, typically measured in steps per minute (step / min). Distance to motor vehicles is defined as the shortest straight-line distance between any part of a pedestrian's body (usually the torso or feet) and the vehicle's body (usually the front or side) within a pre-defined area (such as the area the pedestrian is crossing or a roadside waiting area), typically measured in meters (m). This distance is used to determine if a pedestrian is in a dangerous area; if the distance to a motor vehicle is less than 2 meters and the vehicle is not slowing down, the pedestrian must immediately stop walking. The probability of avoiding a collision is defined as the likelihood that a pedestrian will actively take evasive action such as stopping, backing up, or accelerating across the road, based on the current environment (e.g., distance from the vehicle, vehicle speed, and the pedestrian's own condition). The probability ranges from 0 (no avoidance) to 1 (definite avoidance). For example, the probability is determined by multiple factors. For instance, when a vehicle is 5 meters away at a speed of 30 km / h, the probability of avoiding a collision is approximately 0.8 (high probability of stopping and waiting); when a vehicle is 2 meters away at a speed of 50 km / h, the probability of avoiding a collision is approximately 0.3 (possibly due to insufficient reaction time, unable to avoid the collision).

[0046] The definition of travel speed is as follows: The speed of a non-motorized vehicle (such as a bicycle or electric bicycle) along its direction of travel at a given instant or over a short distance, usually measured in kilometers per hour (km / h) or meters per second (m / s). For example, bicycles typically travel at approximately 10-15 km / h, while compliant electric bicycles travel at approximately 20-25 km / h. The definition of acceleration is: The rate of change of speed of a non-motorized vehicle over time, usually measured in meters per second. 2 (m / s) 2 Positive values ​​indicate acceleration (e.g., twisting the throttle on an electric bicycle), while negative values ​​indicate deceleration (e.g., squeezing the brakes). For example, values ​​reflect the dynamic response of non-motorized vehicles; for instance, the acceleration of an electric bicycle is approximately 0.5-1.0 m / s². 2 When a bicycle (by human power) accelerates, it travels at approximately 0.2-0.5 m / s. 2* **Tilt Angle Definition:** The angle between the non-motorized vehicle body (e.g., the longitudinal centerline of the frame) and the horizontal plane, usually measured in degrees (°). The angle is negative / positive when tilting left / right, and 0° when vertical. Example: To determine the stability of a non-motorized vehicle, a tilt angle of approximately 5-15° (normal range) during a turn, and >20°, may indicate a risk of tipping over. Excessive fluctuations in tilt angle during straight-line travel may indicate unstable rider operation. * **Distance from Motor Vehicles Definition:** Within a preset range (e.g., non-motorized vehicle lanes, mixed traffic areas at intersections), the shortest straight-line distance between the non-motorized vehicle body (usually the front or wheels) and the motor vehicle body (usually the side or rear), usually measured in meters (m). Example: Reflects the safety redundancy of non-motorized vehicles. For example, when riding in a non-motorized vehicle lane, a distance of <1.5m from an adjacent motor vehicle increases the risk of collisions due to lane changes by the motor vehicle. Definition of distance from pedestrians: Within a preset area (such as a sidewalk or crosswalk at an intersection), the shortest straight-line distance between the non-motorized vehicle and the pedestrian, usually expressed in meters (m). For example, to determine whether a non-motorized vehicle affects pedestrian safety, if it is traveling on a sidewalk and the distance to a pedestrian is less than 1m, it must slow down or stop to yield.

[0047] Among them, the probability of lane change intention: This parameter quantifies the driver's decision-making tendency to "plan to change lanes" by analyzing vehicle driving behavior (such as steering operations and speed adjustments). The value ranges from 0 (no intention to change lanes at all) to 1 (certain to change lanes). It is a key indicator for predicting changes in vehicle paths and avoiding lateral conflicts. Intersection crossing choice definition: This refers to the decision options made by pedestrians at intersections (with / without crosswalks, with / without traffic lights) based on environmental judgments (such as vehicle flow and traffic light status), including "whether to cross the road," "from which location to cross," and "at what speed." It is a "goal-oriented attribute" of pedestrian crossing decisions. Dwell time definition: This refers to the duration of time a pedestrian remains stationary at an intersection before making a "crossing choice," observing the environment (such as traffic lights and oncoming vehicles). The unit is seconds (s). It is a key time dimension reflecting the pedestrian's "decision-making caution." Lane departure definition: refers to the lateral deviation of the non-motorized vehicle from the boundary / center line of the "legal non-motorized vehicle lane (or default driving area)" caused by decisions such as "actively avoiding obstacles," "needing to use the lane," or "distraction errors." It is a quantitative indicator for judging whether a cyclist "complies with lane rules." Violation probability definition: refers to the likelihood that, in specific scenarios (such as intersections or pedestrian crossings), a cyclist will violate the non-motorized vehicle traffic rules of the Road Traffic Safety Law due to decisions such as "rushing," "going against traffic," or "failing to yield." The value ranges from 0 (fully compliant) to 1 (inevitable violation), and it is a core indicator for assessing decision-making risk.

[0048] In one possible implementation, a prediction model corresponding to any type of traffic participant is used to predict based on the first behavioral features of that type of traffic participant, thereby obtaining the second behavioral information corresponding to that type of traffic participant, including: Using a prediction model corresponding to any type of traffic participant, a prediction is made based on the first behavioral feature corresponding to any type of traffic participant to obtain the second behavioral feature corresponding to any type of traffic participant. The behavioral characteristics of the second behavioral feature corresponding to different types of traffic participants are different. Based on the second behavioral feature corresponding to any type of traffic participant, the second behavioral information corresponding to any type of traffic participant is determined.

[0049] In this embodiment, the description of the second behavioral feature can be referred to the description of the first behavioral feature, and will not be repeated here.

[0050] In this embodiment, the second behavioral feature can represent the intention of any type of traffic participant following the first behavioral information. Then, according to the intention represented by the second behavioral feature, the second behavioral information corresponding to any type of traffic participant can be determined, and this second behavioral information matches the intention represented by the second behavioral feature. For example, if the second behavioral feature includes a lane-changing intention probability, then if the lane-changing intention probability is greater than an intention probability threshold, it is considered that a lane change will occur, and the second behavioral information is "lane change." If the lane-changing intention probability is not greater than the intention probability threshold, it is considered that a lane change will not occur. As another example, assuming the second behavioral feature includes intersection crossing selection and dwell time, then if intersection crossing selection indicates crossing the intersection, the second behavioral information is "crossing the intersection," and the dwell time is zero. Assuming the second behavioral feature includes not crossing the intersection, the second behavioral information is the dwell time of remaining in place. For example, suppose the second behavioral feature includes the degree to which the non-motorized vehicle deviates from the lane and the probability of violation. If the probability of violation is greater than the probability threshold, the second behavioral information can be that a violation has been determined and the non-motorized vehicle deviates from the lane according to the degree of lane deviation. If the probability of violation is not greater than the probability threshold, the second behavioral information can be considered as not deviating from the lane.

[0051] In this embodiment, a prediction model corresponding to any type of traffic participant is used to predict based on the first behavioral feature corresponding to any type of traffic participant to obtain the second behavioral feature corresponding to any type of traffic participant. The second behavioral feature is used to represent the intention of any type of traffic participant. The behavioral characteristics of the second behavioral features corresponding to different types of traffic participants are different. Based on the second behavioral features corresponding to any type of traffic participant, the second behavioral information corresponding to any type of traffic participant is determined. Since the behavioral characteristics of the second behavioral features corresponding to different types of traffic participants are different, the second behavioral information can also be determined using the second behavioral features with different behavioral characteristics, thereby improving the accuracy of behavioral information prediction.

[0052] In another possible implementation, the behavioral characteristics of the second behavioral feature of different types of traffic participants can be the same.

[0053] In summary, in this embodiment, the first behavioral feature and the second behavioral feature are differentiated features for different types of traffic participants. For easier understanding, accompanying drawings are provided below. Please refer to... Figure 3 , Figure 3 This is a schematic diagram of a behavioral feature system provided in an embodiment of this application. The behavioral features in this embodiment may be, for example, a first behavioral feature, a second behavioral feature, a third behavioral feature, and a fourth behavioral feature.

[0054] like Figure 3 The behavioral characteristic system shown can include vehicle (motor vehicle) characteristic branches, pedestrian characteristic branches, and two-wheeled vehicle (non-motor vehicle) characteristic branches.

[0055] Among them, the vehicle feature branch (motor vehicles): Motion characteristics: instantaneous speed The speed of a motor vehicle at a given moment; Instantaneous acceleration : Rate of change of speed over time (intensity of acceleration / deceleration); accelerometer : Rate of change of acceleration (reflects the smoothness of acceleration / deceleration operation); Driving direction angle The angle between the vehicle's trajectory and the reference direction (such as the center line of the road) is used to determine whether the vehicle has deviated from its lane.

[0056] Interactive features: Distance to the vehicle in front The real-time distance between this vehicle and the vehicle in front measures the safety margin for following other vehicles. Collision time If neither vehicle yields, the remaining time before a collision occurs between the vehicle in front and the vehicle in front is a key indicator for predicting the risk of a collision.

[0057] Decision-making characteristics: probability of lane change intention Based on the current scenario S (such as road conditions, traffic flow) and operation (For example, by using turn signals) to assess the likelihood of a vehicle changing lanes in the future.

[0058] In the characteristic branches of motor vehicles, 1) instantaneous speed: ,in , (1) Instantaneous velocity components of the vehicle in the x and y directions, respectively; (2) Instantaneous acceleration: (2) The first derivative of instantaneous velocity with respect to time; (3) Collision time: ,in To maintain distance from the vehicle in front, The speed of the vehicle in front. For the speed of this vehicle, (3) Minimum value (to avoid zero denominator); (4) Lane change intention probability: Where S represents the turn signal status (0 = off, 1 = on), and δ represents the vehicle heading angle. , σ is the weighting coefficient, b is the bias, and σ is the sigmoid function.

[0059] Among them, the pedestrian feature branch: Motion characteristics: walking speed The speed at which a pedestrian moves at a given moment; Step length The distance a pedestrian covers in one step reflects stride length; Step frequency The number of steps taken per unit of time reflects the walking rhythm.

[0060] Interactive features: Distance from vehicle The shortest distance between a pedestrian and a motor vehicle is used to determine whether the pedestrian is in a danger zone. Avoidance probability Based on pedestrian status P (such as age, attention) and vehicle status V (such as speed, distance), the probability of a pedestrian actively avoiding a motor vehicle.

[0061] Decision-making characteristics: Crossing the intersection Pedestrians' decision options at intersections: "whether to cross," "where to cross," and "at what speed." pause time The length of time pedestrians linger at intersections before making a decision reflects the level of caution in their decision-making.

[0062] In this embodiment, specific features are extracted based on pedestrian gait and interaction avoidance characteristics. The key feature calculation formula is as follows: (1) Step size: ,in , (2) Step frequency: (1) Pedestrian position coordinates at time tᵢ; ,in (3) Avoidance probability: ,in Let α be the shortest distance between pedestrians and vehicles, and α be the distance sensitivity coefficient.

[0063] Among them: Two-wheeled vehicle characteristic branch (non-motorized vehicle) Motion characteristics: driving speed Real-time speed of non-motorized vehicles; acceleration : Rate of change of velocity (intensity of acceleration / deceleration); tilt angle The angle between the non-motorized vehicle body and the horizontal plane is used to determine driving stability (such as the risk of turning or rolling over).

[0064] Interactive features: Distance from motor vehicles The shortest distance between non-motorized vehicles and motorized vehicles, used to measure the risk of lateral conflict; Distance from pedestrians The shortest distance between non-motorized vehicles and pedestrians is used to determine whether it affects pedestrian safety.

[0065] Decision-making characteristics: Lane Departure The lateral distance by which a non-motorized vehicle deviates from the legally designated lane reflects the degree of lane compliance. probability of violation Based on current behavior (such as going against traffic or running a red light), the likelihood of non-motorized vehicles violating traffic rules.

[0066] In this embodiment, core features are extracted based on the characteristics of two-wheeled vehicle mobility and stability. The key feature calculation formula is as follows: (1) Inclination angle: ,in , (1) The acceleration components of the two-wheeled vehicle in the x and y directions, where g is the acceleration due to gravity; (2) Other characteristics: travel speed acceleration Distance from motor vehicles Lane departure , etc., calculated using physical or geometric methods.

[0067] The following example, based on the above example, explains how to use the prediction model corresponding to any type of traffic participant to predict based on the first behavioral feature corresponding to any type of traffic participant, and obtain the second behavioral feature corresponding to any type of traffic participant.

[0068] Specifically, a vehicle's behavior can be predicted based on its first behavioral feature using a corresponding prediction model, thus obtaining the vehicle's second behavioral feature. This prediction model can also be called a vehicle behavior prediction model. The following section explains the vehicle behavior prediction model.

[0069] I. Adapted Scenarios To address the scenario transition requirements of long-term driving decisions (such as following other vehicles on highways, continuous lane changes, and crossing intersections), this paper solves the problem that traditional models have difficulty capturing long-term behavioral dependencies, ensuring that vehicle behavior in simulation scenarios is consistent with real-world driving logic.

[0070] II. Input / Output Design Input features: Vehicle history T in real-world scenarios V1 Multidimensional temporal features of time steps, including instantaneous velocity acceleration Driving direction angle Distance to the vehicle in front Collision time The total dimension of the features is ; Output: Future in the simulation scenario The time-step behavior prediction values ​​cover core parameters such as speed, acceleration, and lane change intention probability P(LC), with the output dimension matching the input feature dimension one by one.

[0071] III. Network Architecture Optimization 1. Location coding Sine and cosine position coding is used to supplement the position information of temporal features, as shown in the following formula: Where pos is the temporal position, i is the feature dimension index, and dmodel is the feature embedding dimension.

[0072] 2. Encoder Structure Stacked NEnc layers, each layer containing: a multi-head self-attention mechanism: H heads, each head with a feature dimension of... Capture long-term dependencies between different features; Feedforward network: The number of hidden layer nodes is dff, and the ReLU activation function is used to enhance the nonlinear representation of features.

[0073] 3. Decoder Structure Stack NDec layers, each layer containing: Masked multi-head self-attention: Masking future time series information to avoid leaking subsequent data during the prediction process; Encoder-decoder attention: Associates input temporal features with output prediction results to improve long-term prediction accuracy.

[0074] In this model, pedestrians can be predicted based on the first behavioral feature of pedestrians using the prediction model corresponding to motor vehicles, thus obtaining the second behavioral feature of pedestrians. The prediction model corresponding to pedestrians can also be called a pedestrian behavior prediction model. The following is an explanation of the pedestrian behavior prediction model.

[0075] I. Adapted Scenarios To address the scenario transformation needs of pedestrian group interaction behaviors (such as crossing intersections in groups, avoiding crowds, pedestrian-vehicle interactions, etc.), this paper solves the problem that traditional models cannot simultaneously capture individual temporal behaviors and group interaction relationships.

[0076] II. Input / Output Design Input features: Pedestrian history T in real-world scenarios P1 Individual characteristics of time step (walking speed v) P Step size s P , direction of travel Distance from vehicles The real-time location coordinates of the ) and K surrounding pedestrians, with a total feature dimension of . ; Output: Future T in the simulation scenario P2 Pedestrian position coordinates at time step (x P ,y P ) and intersection crossing decision C P (0 = no time travel, 1 = time travel).

[0077] III. Network Architecture Optimization 1. Long Short-Term Memory (LSTM) module (individual temporal feature extraction) We employ an NLSTM layer with bidirectional LSTM, and the hidden layer dimension is HLSTM: Forward LSTM: Input historical data in chronological order to capture past temporal features; Backward LSTM: Input historical data in reverse chronological order to capture future dependencies; Output: By concatenating the forward and backward outputs, we obtain the temporal representation vector of individual behavior.

[0078] 2. Graph Neural Network (GNN) module (group interaction feature extraction) Using NGCN layer graph convolutional network (GCN): Graph construction: Using individual actors as nodes, construct an adjacency matrix based on a distance decay function. (dij is the straight-line distance between pedestrians i and j, and dth is the distance threshold). Graph convolution: via the formula ( Given an adjacency matrix with self-loops, for The degree matrix (W is the weight matrix) is used to extract group interaction features.

[0079] 3. Attention Fusion Layer A multi-head attention mechanism is used to fuse individual and temporal features, as shown in the following formula: Where Q represents the individual temporal features output by LSTM, K and V represent the group interaction features output by GCN, and dk represents the feature dimension.

[0080] Among them, non-motorized vehicles can be predicted based on the first behavioral feature of the non-motorized vehicle using a prediction model, thus obtaining the second behavioral feature of the non-motorized vehicle. This prediction model for non-motorized vehicles can also be called a two-wheeled vehicle behavior prediction model. The following is an explanation of the two-wheeled vehicle behavior prediction model.

[0081] I. Adapted Scenarios To address the scenario transformation requirements of two-wheeled vehicles' local dynamic behaviors (such as tilting on curves, flexibly changing lanes in traffic, and avoiding pedestrians), this paper solves the problem that traditional models cannot take into account both local environmental characteristics and temporal behavior dependencies.

[0082] II. Input / Output Design Input features: Historical T-type two-wheeled vehicles in real-world scenarios M1 Individual characteristics of time step (driving speed) Inclination angle Lane departure Distance from motor vehicles The total dimensions of the survey include: road local structural features (lane curvature, intersection signage), and the overall road structure. ; Output: Future T in the simulation scenario M2 Two-wheel vehicle speed vM, tilt angle θlean, and lane departure L at time step M .

[0083] III. Network Architecture Optimization 1. Convolutional Neural Network (CNN) module (local feature extraction) Stacked NCNN layers of convolutional networks to adapt to mixed temporal and spatial features: Convolution kernel: The scale is k×1 (k is the time step window), and the number of convolution kernels is K1, K2, ..., ; Pooling layer: Max pooling (pooling kernel size 2×1) is used to preserve key local features (such as road curvature changes and close-range vehicle interactions).

[0084] 2. LSTM module (time-dependency capture) use LSTM with hidden layer dimension of 1. : Input: Receives local features from the CNN output, capturing long-term temporal dependencies in behavior (such as changes in the tilt angle of consecutive curves); Regularization: Adds a dropout layer (probabilistic) to the output. To suppress overfitting, the output layer maps the LSTM output to the predicted value through a linear mapping layer. The Sigmoid activation function is used to normalize the tilt angle (range -15° to 15°) and lane offset (range -0.5m to 0.5m) to ensure that the output conforms to real physical constraints.

[0085] Please see Figure 4 , Figure 4 This diagram illustrates a machine learning modeling and scene transformation architecture according to an embodiment of this application. This embodiment is explained in conjunction with simulations. Figure 4 The steps shown can be referred to the instructions above, and will not be repeated here.

[0086] In this embodiment, the obtained second behavioral feature can be written into the real-world behavioral feature library and used to generate simulated behaviors.

[0087] It should be noted that when constructing the standardized library (real-world behavioral feature library) and generating simulated behaviors, the first step is to cluster the behavioral characteristics of the three types of participants using a spatial-temporal density-based spatial clustering of applications with noise (ST-DBSCAN) algorithm with differentiated parameters (vehicles: , , ;pedestrian: Two-wheeled vehicle: This ensures that the clustering results accurately reflect the behavioral characteristics of different participants. Clustering effectiveness is evaluated using the silhouette coefficient, calculated using the following formula: Profile coefficient: ; in Let i be the average distance between sample i and samples in the same cluster. Let SC be the average distance between sample i and the nearest heterogeneous sample, when SC ≥ SC th The clustering is deemed valid at that time. The percentage of clustered cells removed is < Small clusters (considered as anomalous behavior) are identified through key features (such as the "high-speed following" pattern of vehicles, which must meet certain criteria). and The system automatically names behavioral patterns and ultimately builds a standardized library containing N_V class vehicle patterns, N_P class pedestrian patterns, and N_M class two-wheeled vehicle patterns. It can also convert real-world behavioral patterns into simulation patterns.

[0088] In this embodiment, to ensure the model continuously adapts to changes in real-world scenarios, a dynamic update mechanism is established—for every N_update new real-world scene behavior data accumulated, the bottom N_freeze layer feature extraction layers of the VAE and prediction model are frozen, and only the top-level γ_tune scaling parameter is fine-tuned to reduce training resource consumption; the performance of the new model is verified through A / B testing, and when the prediction error of the new model is reduced by more than Δ_err compared to the old model, the old model is automatically replaced; when N_new new behavior patterns are detected (meeting the requirements of the previous model), the new model is updated accordingly. When performing this operation, the behavior pattern library is expanded to ensure the timeliness and accuracy of real-world scenario transitions. Input: The vehicle's historical time step T_V1 in the real-world scenario. Temporal characteristics; Output: Predicted behavior at future T_V2 time steps in the simulation scenario.

[0089] The training of the model will be explained below. In one possible implementation, the method also includes: For any type of traffic participant, obtain multiple training samples corresponding to that type of traffic participant. Each training sample includes third and fourth behavior information for that type of traffic participant, and the fourth behavior information includes behavior information following the third behavior information. For any type of traffic participant, call the training model corresponding to that type of traffic participant. For any type of traffic participant, train the training model corresponding to that type of traffic participant using the multiple training samples, and obtain the prediction model corresponding to that type of traffic participant.

[0090] In this embodiment, the third row of information can be referred to in the description of the first row of information, and the fourth row of information can be referred to in the description of the second row of information, and will not be repeated here. The model to be trained can be referred to in the description of the prediction model, and will not be repeated here.

[0091] In this embodiment, for any type of traffic participant, multiple training samples corresponding to that type of traffic participant are obtained. Each training sample includes third and fourth behavioral information corresponding to that type of traffic participant, and the fourth behavioral information includes behavioral information following the third behavioral information. For any type of traffic participant, the model to be trained corresponding to that type of traffic participant is invoked. For any type of traffic participant, the model to be trained corresponding to that type of traffic participant is trained using the multiple training samples corresponding to that type of traffic participant to obtain a prediction model corresponding to that type of traffic participant. This allows for the processing of information on the behavioral characteristics of traffic participants of the corresponding category.

[0092] In one possible implementation, a training model for any type of traffic participant is trained using multiple training samples corresponding to any type of traffic participant to obtain a prediction model for any type of traffic participant, including: Feature extraction is performed on the third behavior information corresponding to any type of traffic participant to obtain third behavior features, where the behavioral characteristics of the third behavior features corresponding to different types of traffic participants are different. The third behavior features corresponding to any type of traffic participant are used as input to the training model corresponding to that type of traffic participant to obtain the predicted behavior information of the output of the training model corresponding to that type of traffic participant. The training loss of the training model corresponding to that type of traffic participant is determined using the fourth behavior information and the corresponding predicted behavior information. The training losses of the training models corresponding to at least two types of traffic participants are weighted and calculated to obtain the target training loss. If the target training loss meets the training termination condition, the training model corresponding to any type of traffic participant is used as the prediction model corresponding to that type of traffic participant. If the target training loss does not meet the training termination condition, the training model corresponding to that type of traffic participant continues to be trained using multiple training samples corresponding to that type of traffic participant until the target training loss meets the training termination condition.

[0093] In this embodiment, the fusion of training losses based on the training losses of the models to be trained for any type of traffic participant can be achieved by weighted averaging. For example, the training losses of the models to be trained for motor vehicles, non-motor vehicles, and pedestrians can be weighted and averaged to obtain the target training loss. Optionally, the fusion weights for fusing training losses can be set according to the importance of the traffic participant category; the higher the importance, the higher the fusion weight, and vice versa. The training termination condition can be, for example, that the loss value falls below a certain threshold and tends to stabilize.

[0094] In this embodiment, features are extracted from the third behavior information corresponding to any type of traffic participant to obtain third behavior features. The behavioral characteristics of the third behavior features differ for different types of traffic participants. The third behavior features corresponding to any type of traffic participant are used as input to the training model for that type of traffic participant to obtain the predicted behavior information output by the training model. The training loss of the training model for that type of traffic participant is determined using the fourth behavior information and the corresponding predicted behavior information. The training losses of the training models for at least two types of traffic participants are weighted and calculated to obtain the target training loss. If the target training loss meets the training termination condition, the training model for that type of traffic participant is used as the prediction model for that type of traffic participant. If the target training loss does not meet the training termination condition, the training model for that type of traffic participant continues to be trained using multiple training samples from that type of traffic participant until the target training loss meets the training termination condition. This approach takes into account the predictive capabilities of prediction models for different types of traffic participants, thereby improving the accuracy of behavior prediction. The phrase "continuing to train the model corresponding to any type of traffic participant using multiple training samples" can refer to simultaneously updating all models, or updating them step-by-step or selectively, without further limitation.

[0095] In this embodiment, the model training uses a hybrid loss function, and the calculation formula is as follows: Hybrid loss function: ; in, Reconstruction loss (L2 loss is used for vehicles / two-wheeled vehicles): (Pedestrians are subject to cross-entropy loss). KL divergence loss: β is the balance coefficient (β_V for vehicles, β_P for pedestrians, and β_M for two-wheeled vehicles). Training parameters: Adam optimizer (learning rate η_VAE), batch B_VAE, and iteration E_VAE, to ensure that the features retain the core information of the real scene after dimensionality reduction, while improving the conversion efficiency.

[0096] In another possible implementation, the training loss of the training model corresponding to different types of traffic participants can be used to determine whether the training termination condition is met separately, without any restrictions.

[0097] The following section explains the calculation of the training loss for the training model corresponding to different categories of traffic participants.

[0098] The training loss for the model to be trained corresponding to motor vehicles is explained below: Loss function and training configuration 1. Loss Function A weighted mixed loss is used to balance the regression task (velocity, acceleration) and the classification task (lane change intention), as shown in the following formula: in, (λ1 and λ2 are regression loss weights, λ3 is classification loss weight), CE is the cross loss.

[0099] 2. Training parameters Optimizer: AdamW (learning rate ηTrans, weight decay λreg); Training configuration: Batch size BTrans, training epochs ETrans; Early stopping strategy: Stop training when the validation set loss does not decrease for consecutive Pstop rounds to avoid overfitting.

[0100] The training loss of the model to be trained for pedestrians is explained below: 1. Loss Function The formula for combining location prediction loss and decision classification loss is as follows: in, γ is the Euclidean distance (measuring prediction bias), γ is the decision loss weight, and CE is the cross-entropy loss.

[0101] 2. Training parameters Optimizer: RMSprop (learning rate ηLSTM-GNN); Training configuration: Batch size BLSTM-GNN, training epochs ELSTM-GNN; Regularization: Dropout rate pdrop, which suppresses model overfitting.

[0102] The training loss for the model to be trained, corresponding to non-motorized vehicles (e.g., two-wheeled vehicles), is explained below: 1. Loss Function Using a weighted L1 loss, focusing on the tilt angle (which affects the stability of the two-wheeled vehicle), the formula is as follows: Where β (tilt angle loss weight) > (Lane deviance loss weight) Prioritize ensuring the authenticity of dynamic behavior.

[0103] 2. Training parameters Optimizer: AdamW (learning rate ηCNN-LSTM, weight decay λreg-M); Training configuration: Batch size BCNN-LSTM, training epochs ECNN-LSTM; Learning rate strategy: cosine annealing (initial learning rate ηinit, final learning rate ηfinal), balancing convergence speed in the early stage of training with accuracy in the later stage.

[0104] In one possible implementation, the third behavioral feature corresponding to any type of traffic participant is used as the input to the training model corresponding to any type of traffic participant, to obtain the predicted behavioral information of the output of the training model corresponding to any type of traffic participant, including: The third behavioral feature corresponding to any type of traffic participant is used as the input to the training model corresponding to any type of traffic participant. The training model corresponding to any type of traffic participant makes predictions based on the third behavioral feature corresponding to any type of traffic participant to obtain the fourth behavioral feature. The behavioral characteristics of the fourth behavioral feature corresponding to different types of traffic participants are different. Based on the fourth behavioral feature corresponding to any type of traffic participant, the predicted behavioral information corresponding to any type of traffic participant is determined.

[0105] In this embodiment, the third behavioral feature can be referred to the description of the first behavioral feature, and the fourth behavioral feature can be referred to... In this embodiment, the third behavioral feature corresponding to any type of traffic participant is used as the input to the training model corresponding to any type of traffic participant. The training model corresponding to any type of traffic participant makes predictions based on the third behavioral feature corresponding to any type of traffic participant to obtain a fourth behavioral feature. The behavioral characteristics of the fourth behavioral feature corresponding to different types of traffic participants are different. Based on the fourth behavioral feature corresponding to any type of traffic participant, the predicted behavioral information corresponding to any type of traffic participant is determined. In this way, not only are the behavioral characteristics of the features input to the model related to the category of traffic participants, but the behavioral characteristics of the features output by the model are also related to the category of traffic participants, thereby improving the accuracy of behavior prediction.

[0106] It should be noted that if this embodiment is applied to simulation, the second information obtained can be the estimated real-world scene information, and then the real-world scene information can be converted into simulation scene information.

[0107] After completing the conversion from the real-world scenario to the simulation scenario, the conversion quality needs to be quantitatively evaluated using multi-dimensional indicators to determine whether the simulation scenario closely matches the real-world scenario, thus providing a basis for subsequent optimization. First, based on the behavioral characteristics of the transformed simulation scene and the real scene, the single-feature similarity is calculated. The core calculation formula is as follows: Trajectory Similarity (DTW): Where X is the trajectory sequence of the real scene, Y is the trajectory sequence of the simulated scene, and ω is the time warped path; Distribution similarity (KL divergence): ; Where P represents the feature distribution of the real scene, and Q represents the feature distribution of the simulated scene; Joint distribution similarity (Wasserstein distance): in, Let be the set of joint distributions of P and Q.

[0108] Based on this, the overall conversion quality score is calculated using a weighted formula: Overall rating: ; in, , , The scores are based on the overall similarity of behavior for vehicles, pedestrians, and two-wheeled vehicles. , , Weighting coefficients (satisfying) ); when If the conversion quality is deemed satisfactory, an optimization process is triggered.

[0109] like Figure 5 As shown, Figure 5 This is a flowchart illustrating a conversion quality assessment and optimization process according to an embodiment of this application. Figure 5 The process shown may include: S510, Current simulation parameters.

[0110] In this embodiment, the simulation scenario parameters to be evaluated (such as the movement, interaction, and decision-making characteristics of traffic participants) are input into the system.

[0111] S520 and PRO strategy updates.

[0112] S530, Reward function calculation R.

[0113] In this embodiment, the reward value R of the intelligent agent is calculated by quantifying simulation results (such as traffic flow and safety) to determine whether the parameters "meet the standards". If "not meeting the standards", the optimization process on the right is triggered; if "meeting the standards", the current round of evaluation is completed.

[0114] S540, reinforcement learning agent.

[0115] In this embodiment, the agent outputs the parameter adjustment amount based on the input parameters, combined with the "reward function R" and "PRO policy update" (policy optimization method), that is, "how to adjust the current parameters to get closer to the target".

[0116] S550, simulation scene data input.

[0117] This involves importing the original data of the simulation scenario to be optimized (such as the behavior records of virtual traffic participants).

[0118] S560, Reality / Simulation Feature Alignment.

[0119] In this embodiment, the behavioral characteristics in the simulation scenario (such as the intention of motor vehicles to change lanes and the choice of pedestrians to cross) are compared with the real characteristics in real traffic to find the differences between the two (such as the pedestrian avoidance probability in the simulation not matching reality).

[0120] S570, Similarity Calculation.

[0121] Among them, based on the results of "feature alignment", the similarity between simulation and real behavior is calculated (the value ranges from 0 to 1, and the closer it is to 1, the more consistent it is).

[0122] S580, weight allocation.

[0123] Among them, based on the importance of different behavioral characteristics (such as "collision time" being more critical to safety than "step length"), weights are assigned to each characteristic (e.g., safety-related characteristics have a weight of 0.6, and efficiency-related characteristics have a weight of 0.4).

[0124] S590, Overall Score.

[0125] The simulation scenario can be evaluated by combining "similarity" and "weight" to calculate a comprehensive score S, which measures the effectiveness of the simulation. If the score is low (i.e., "unsatisfactory"), it is fed back to the evaluation process on the left to drive the next round of parameter adjustments.

[0126] This embodiment involves simulation scene optimization based on reinforcement learning: For simulation scenarios with substandard conversion quality, reinforcement learning is used to automatically optimize behavioral parameters, making the simulation scenarios more closely resemble real-world scenarios. First, we construct a reinforcement learning framework: State space: The set of behavioral characteristic parameters of three types of participants in the simulation scenario. ( For vehicle parameters, For pedestrian parameters, (Parameters for two-wheeled vehicles) Motion space: parameter adjustment amount The adjustment range meets the requirements. (To avoid excessive parameter adjustments that could lead to scene anomalies); Reward function: Reward function: in, To improve the conversion quality, the reward coefficient is increased. Adjust the regularization coefficients for the parameters. To optimize the difference in overall scores before and after.

[0127] Secondly, the Proximal Policy Optimization (PPO) algorithm is used for optimization. The core optimization objective formula is as follows: PPO optimization goals: ; in, A represents the ratio of the current strategy to the old strategy. t For the dominant function, This refers to the pruning parameters (controlling the update magnitude of the strategy).

[0128] Finally, set the maximum number of iterations. After each iteration, the overall score S is recalculated. When S≥S threshol Or, when the maximum number of iterations is reached, the optimization stops and the optimized simulation scene parameters are output, completing the closed loop of the entire process of real-world scene conversion and optimization.

[0129] For ease of understanding, a complete embodiment is provided below.

[0130] Please see Figure 6 , Figure 6 This is a flowchart illustrating an output simulation scene according to an embodiment of this application.

[0131] S601: Input of real data from multiple sources.

[0132] Meaning: Input multi-dimensional behavioral data of vehicles, pedestrians, and two-wheeled vehicles (non-motorized vehicles) in real traffic scenarios, including motion parameters (such as speed and acceleration), interaction records (such as distance between vehicles and avoidance behavior), and decision-making behaviors (such as lane change intentions and intersection crossing choices).

[0133] In this step, to ensure the accuracy of subsequent behavioral feature extraction and scene transformation, the collected multi-source real-world scene data must first undergo standardization processing. First, multi-source real-world data, such as vehicle road test data and traffic monitoring data, must be collected. This data must comprehensively cover the behavioral information of three types of traffic participants: vehicles (V), pedestrians (P), and two-wheeled vehicles (M), including key data such as location, speed, and interaction actions. Second, spatiotemporal alignment processing is performed, using algorithms such as Kalman filtering to unify data from different sensors (cameras, LiDAR, millimeter-wave radar) to the world coordinate system and the same time axis. To eliminate the impact of spatiotemporal bias on subsequent transformations, the SMOTE algorithm is used for data augmentation. This augments rare behavioral samples such as lane changing and jaywalking, addressing the imbalance of behavioral samples and providing sufficient and balanced training data for subsequent machine learning modeling.

[0134] S602: Data preprocessing module.

[0135] Spatiotemporal alignment: unifying data from different sources and with different time / space references into the same spatiotemporal coordinate system to ensure the spatiotemporal consistency of the data (e.g., aligning vehicle trajectories captured by different cameras to the same road coordinate system).

[0136] Data cleaning: Filtering outlier data (such as urban road data where vehicle speed suddenly jumps to 100km / h), and supplementing missing values ​​to ensure data accuracy.

[0137] Sample augmentation: Increase the number and diversity of samples through data transformation (such as time interpolation, behavior pattern replication and expansion) to provide sufficient and rich input for subsequent model training.

[0138] S603: Feature extraction module.

[0139] Vehicle feature extraction: Extracting the motion features (instantaneous speed, acceleration, driving direction angle, etc.), interaction features (distance to the vehicle in front, collision time, etc.) and decision features (probability of lane change intention, etc.) of motor vehicles.

[0140] Pedestrian feature extraction: Extracting pedestrian motion features (walking speed, stride length, stride frequency, etc.), interaction features (distance from vehicles, probability of avoidance, etc.), and decision features (intersection crossing choices, pause time, etc.).

[0141] Two-wheeled vehicle feature extraction: Extract the motion features (speed, acceleration, tilt angle, etc.), interaction features (distance from motor vehicles / pedestrians, etc.) and decision features (lane deviation, probability of violation, etc.) of non-motorized vehicles.

[0142] S604: Model building module.

[0143] Behavioral clustering: Cluster analysis is performed on the extracted features to identify typical behavioral patterns of traffic participants (such as two clusters for vehicles: "smooth following" and "sudden braking and lane changing", and two clusters for pedestrians: "rapid crossing" and "hesitant stopping").

[0144] Feature dimensionality reduction: Reduce feature dimensionality using dimensionality reduction algorithms such as PCA, while retaining core information (e.g., reducing 10+ motion features of a vehicle to 3-5 key principal components), thereby improving subsequent computational efficiency.

[0145] Model training and prediction: Construct game models among traffic participants (such as "vehicle-pedestrian" avoidance game and "non-motorized vehicle-motorized vehicle" right-of-way game) to simulate the decision-making interaction process of multiple parties.

[0146] Scene parameter conversion: Convert real-world scene parameters into simulation parameters and verify the consistency between key parameters in the model (such as collision time threshold and avoidance probability threshold) and the real scene to ensure the authenticity of the model's basic logic.

[0147] S605: Realism assessment module.

[0148] Single-feature similarity calculation: Evaluate the behavioral indicators of each traffic participant individually (such as vehicle lane change success rate, pedestrian avoidance timeliness rate, and non-motorized vehicle violation frequency) to clarify individual performance.

[0149] Overall score: Combining the weights of each indicator (e.g., safety indicators have a higher weight than efficiency indicators), the overall similarity score between the simulation scene and the real scene is calculated to quantify the realism of the scene.

[0150] S606: Reinforcement Learning Optimization Module.

[0151] Parameter adjustment: Adjust simulation parameters (such as upper and lower limits of vehicle acceleration and pedestrian cadence range) based on the shortcomings of the scenario assessment.

[0152] PPO Policy Update: The Proximal Policy Optimization (PPO) algorithm is used to update the decision-making strategies of agents in the simulation (such as the strategy for choosing when to change lanes for vehicles and the decision-making strategy for pedestrians crossing intersections).

[0153] Quality verification: Re-evaluate the simulation scene after parameter adjustment to verify whether its feature similarity with the real scene has improved.

[0154] Loop judgment: Determine whether the comprehensive score S meets the threshold. If "No", return to the "Parameter Adjustment" step to continue optimization; if "Yes", proceed to the next step.

[0155] S607: Optimized simulation scene output.

[0156] The output is a simulation result that has been optimized through multiple rounds and is highly aligned with real traffic behavior. It can be used in real-world scenarios such as testing autonomous driving systems, optimizing traffic signal timing, and providing traffic risk warnings.

[0157] In summary, this embodiment addresses the problem of using a single rule to transform data into scenarios. It designs a hierarchical differentiated feature system, adding specific features such as acceleration and collision time, step length and stride frequency, tilt angle and lane departure for vehicles, pedestrians, and two-wheeled vehicles respectively. This classifies the behavioral patterns of the three types of participants into a standardized pattern library recognizable by the simulation scenario, solving the behavioral bias problem caused by a "one-size-fits-all" transformation and resulting in a higher behavioral pattern matching rate after scenario transformation. To address the problem of existing technologies where "a single model cannot adapt to the behavioral characteristics of different participants," a dedicated model combination architecture of "Transformer + LSTM-GNN + CNN-LSTM" is constructed. Vehicles use a Transformer model with an N1-layer encoder and an N2-layer decoder; pedestrians use an LSTM-GNN model that fuses an N3-layer bidirectional LSTM and an N4-layer GCN; and two-wheeled vehicles use a CNN-LSTM model with an N5-layer convolution and LSTM. Combined with a VAE enhancement model, high-dimensional feature dimensionality reduction (vehicle...) is achieved. ,pedestrian Two-wheeled vehicles This approach addresses the issue of high-dimensional feature redundancy, significantly improving conversion accuracy compared to a single model. To overcome the limitations of existing technologies in effectively evaluating data conversion scenario quality and their heavy reliance on manual verification, a closed-loop solution is constructed: "real-world feature extraction - scenario conversion - multi-dimensional evaluation - reinforcement learning optimization." This solution introduces DTW trajectory similarity, KL divergence distribution similarity, and Wasserstein joint distribution similarity to quantify behavioral bias; and uses a weighted formula... Achieve conversion quality quantification; design reward function based on PPO algorithm. To balance quality improvement and parameter stability, automated iterative optimization is achieved, significantly increasing the success rate of scene transformation. Addressing the problem of existing technologies where "fixed models cannot adapt to changes in real-world scenarios," a dynamic update scheme combining incremental training and model expansion is proposed. Every N new data points, the VAE and the underlying feature extraction layer of the prediction model are frozen, with only the top-level preset scaling parameters fine-tuned. This is achieved through the silhouette coefficient (SC≥SC). th It identifies new behavioral patterns and automatically expands the pattern library when more than N6 new patterns are detected, ensuring that the model continuously adapts to scene changes and greatly reducing the response time for new behavioral pattern recognition.

[0158] The method embodiments have been described above; the product embodiments are described below.

[0159] Please see Figure 7 , Figure 7This is a schematic diagram of the structure of a traffic participant behavior information prediction device provided in one embodiment of this application. Figure 7 The device shown can be applied to electronic devices, such as... Figure 7 The apparatus shown may include an acquisition module 710, a feature extraction module 720, and a prediction module 730, wherein: The acquisition module 710 is used to acquire first behavioral information corresponding to at least two types of traffic participants within a preset range; the feature extraction module 720 is used to extract features from the first behavioral information corresponding to any one of the at least two types of traffic participants to obtain first behavioral features corresponding to any one type of traffic participant, wherein the behavioral characteristics of the first behavioral features corresponding to different types of traffic participants are different; the prediction module 730 is used to predict based on the first behavioral features corresponding to any one type of traffic participant using the prediction model corresponding to any one type of traffic participant to obtain second behavioral information corresponding to any one type of traffic participant, wherein the second behavioral information includes behavioral information following the first behavioral information.

[0160] In one possible implementation, when the prediction module 730 uses the prediction model corresponding to any type of traffic participant to predict based on the first behavioral features corresponding to any type of traffic participant, and obtains the second behavioral information corresponding to any type of traffic participant, it is used for: Using a prediction model corresponding to any type of traffic participant, a prediction is made based on the first behavioral feature corresponding to any type of traffic participant to obtain the second behavioral feature corresponding to any type of traffic participant. The behavioral characteristics of the second behavioral feature corresponding to different types of traffic participants are different. Based on the second behavioral feature corresponding to any type of traffic participant, the second behavioral information corresponding to any type of traffic participant is determined.

[0161] In one possible implementation, the first and second behavioral features corresponding to any type of traffic participant include motion features, interaction features, and decision features. Motion features are used to represent the motion of any type of traffic participant, interaction features are used to represent the interaction between any type of traffic participant and other traffic participants, which include participants other than any type of traffic participant within a preset range, and decision features are used to represent the intention of any type of traffic participant.

[0162] In one possible implementation, at least two types of traffic participants include at least two of motor vehicles, pedestrians, and non-motorized vehicles; The motion characteristics corresponding to motor vehicles include at least one of instantaneous speed, instantaneous acceleration, jerk, and driving direction angle; the interaction characteristics corresponding to motor vehicles include at least one of distance to the vehicle in front and collision time, where collision time represents the interval required for a collision between the motor vehicle and the vehicle in front, and the vehicle in front includes vehicles located in the direction of travel of the motor vehicle; the motion characteristics corresponding to pedestrians include at least one of walking speed, stride length, and stride frequency; the interaction characteristics corresponding to pedestrians include at least one of distance to motor vehicles and avoidance probability, where avoidance probability represents the possibility of avoidance between pedestrians and motor vehicles; the motion characteristics corresponding to non-motorized vehicles include at least one of driving speed, acceleration, and tilt angle; the interaction characteristics corresponding to non-motorized vehicles include at least one of distance to motor vehicles and distance to pedestrians; the decision characteristics corresponding to motor vehicles include the probability of lane-changing intent of motor vehicles; the decision characteristics corresponding to pedestrians include at least one of intersection crossing selection and stopping time; the decision characteristics corresponding to non-motorized vehicles include at least one of the degree of lane deviation of non-motorized vehicles and violation probability.

[0163] In one possible implementation, the device further includes: a training module, configured to acquire multiple training samples corresponding to any type of traffic participant, wherein each training sample includes third behavior information and fourth behavior information corresponding to any type of traffic participant, and the fourth behavior information includes behavior information following the third behavior information; to call the training model corresponding to any type of traffic participant for any type of traffic participant; and to train the training model corresponding to any type of traffic participant using the multiple training samples corresponding to any type of traffic participant for any type of traffic participant, thereby obtaining a prediction model corresponding to any type of traffic participant.

[0164] In one possible implementation, the training module trains the model to be trained for any type of traffic participant using multiple training samples corresponding to any type of traffic participant to obtain a prediction model for any type of traffic participant. When this yields the model, the module then: extracts features from the third behavior information of any type of traffic participant to obtain third behavior features, where the behavioral characteristics of the third behavior features differ for different types of traffic participants; uses the third behavior features of any type of traffic participant as input to the model to be trained for any type of traffic participant to obtain the predicted behavior information output by the model to be trained for any type of traffic participant; uses the fourth behavior information and the corresponding predicted behavior information of any type of traffic participant to determine the training loss of the model to be trained for any type of traffic participant; performs a weighted calculation of the training losses of the models to be trained for at least two types of traffic participants to obtain the target training loss; if the target training loss meets the training termination condition, the model to be trained for any type of traffic participant is used as the prediction model for that type of traffic participant; if the target training loss does not meet the training termination condition, the model to be trained for any type of traffic participant continues to be trained using multiple training samples corresponding to any type of traffic participant until the target training loss meets the training termination condition.

[0165] In one possible implementation, the training module, when using the third behavioral feature corresponding to any type of traffic participant as input to the model to be trained for any type of traffic participant to obtain the predicted behavioral information of the output of the model to be trained for any type of traffic participant, is used to: use the third behavioral feature corresponding to any type of traffic participant as input to the model to be trained for any type of traffic participant to predict based on the third behavioral feature corresponding to any type of traffic participant to obtain a fourth behavioral feature, wherein the behavioral characteristics of the fourth behavioral feature corresponding to different types of traffic participants are different; and determine the predicted behavioral information corresponding to any type of traffic participant based on the fourth behavioral feature corresponding to any type of traffic participant.

[0166] This application also provides an electronic device 80, please refer to... Figure 8 It includes a memory 810 and a processor 820, wherein the memory 810 is used to store computer programs; and the processor 820 is used to execute the programs stored in the memory 810 to implement the traffic participant behavior information prediction method described in any embodiment of this application.

[0167] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the traffic participant behavior information prediction method described in any embodiment of this application.

[0168] In this application, unless otherwise expressly defined, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0169] The terms “first,” “second,” “third,” “fourth,” etc., in this application (if any) are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. “Multiple” means no fewer than two.

[0170] In this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this application generally indicates that the preceding and following related objects have an "or" relationship. "Multiple" refers to no fewer than two.

[0171] Unless otherwise specified, all steps in this application may be performed sequentially or randomly. For example, if a method includes steps A and B, it means that the method may include steps A and B performed sequentially, or it may include steps B and A performed sequentially. For example, if a method may also include step C, it means that step C may be added to the method in any order. For example, the method may include steps A, B, and C, or it may include steps A, C, and B, or it may include steps C, A, and B, etc.

[0172] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for predicting the behavior information of traffic participants, characterized in that, include: Obtain the first behavioral information of at least two types of traffic participants within a preset range; For any one of at least two types of traffic participants, feature extraction is performed on the first behavior information corresponding to the first type of traffic participant to obtain the first behavior feature corresponding to the first type of traffic participant, wherein the behavior characteristics of the first behavior feature corresponding to different types of traffic participants are different. For any of the traffic participants, a prediction model corresponding to the traffic participant is used to predict based on the first behavioral feature corresponding to the traffic participant, thereby obtaining the second behavioral information corresponding to the traffic participant. The second behavioral information includes the behavioral information following the first behavioral information.

2. The method according to claim 1, characterized in that, The step of using the prediction model corresponding to any type of traffic participant to predict based on the first behavioral feature corresponding to any type of traffic participant to obtain the second behavioral information corresponding to any type of traffic participant includes: Using the prediction model corresponding to any type of traffic participant, a prediction is made based on the first behavioral feature corresponding to any type of traffic participant to obtain the second behavioral feature corresponding to any type of traffic participant, wherein the behavioral characteristics of the second behavioral feature corresponding to different types of traffic participants are different. Based on the second behavioral characteristics corresponding to any of the traffic participants, determine the second behavioral information corresponding to any of the traffic participants.

3. The method according to claim 2, characterized in that, The first and second behavioral features corresponding to any type of traffic participant include motion features, interaction features, and decision features. The motion features are used to represent the motion of any type of traffic participant. The interaction features are used to represent the interaction between any type of traffic participant and other traffic participants. The other traffic participants include participants other than any type of traffic participant within the preset range. The decision features are used to represent the intention of any type of traffic participant.

4. The method according to claim 3, characterized in that, The at least two types of traffic participants include at least two of motor vehicles, pedestrians and non-motor vehicles; The motion characteristics of the motor vehicle include at least one of instantaneous speed, instantaneous acceleration, jerk and driving direction angle. The interaction characteristics of the motor vehicle include at least one of distance to the vehicle in front and collision time. The collision time is used to represent the interval time required for the motor vehicle to collide with the vehicle in front. The vehicle in front includes vehicles located in the direction of travel of the motor vehicle. The pedestrian's motion characteristics include at least one of walking speed, stride length, and stride frequency; the pedestrian's interaction characteristics include at least one of distance to the motor vehicle and avoidance probability, wherein the avoidance probability is used to represent the probability of avoidance between the pedestrian and the motor vehicle. The motion characteristics of the non-motorized vehicle include at least one of driving speed, acceleration, and tilt angle; the interaction characteristics of the non-motorized vehicle include at least one of the distance between it and the motorized vehicle and the distance between it and the pedestrian. The decision features corresponding to the motor vehicle include the probability of the motor vehicle's lane-changing intention; the decision features corresponding to the pedestrian include at least one of intersection crossing selection and stopping time; and the decision features corresponding to the non-motor vehicle include at least one of the degree of the non-motor vehicle's deviation from the lane and the probability of violation.

5. The method according to any one of claims 1-4, characterized in that, The method further includes: For any type of traffic participant, obtain multiple training samples corresponding to the traffic participant. Each training sample includes third behavior information and fourth behavior information corresponding to the traffic participant. The fourth behavior information includes behavior information following the third behavior information. For any of the traffic participants, call the training model corresponding to any of the traffic participants; For any one type of traffic participant, the training model corresponding to any one type of traffic participant is trained using multiple training samples corresponding to any one type of traffic participant, thereby obtaining the prediction model corresponding to any one type of traffic participant.

6. The method according to claim 5, characterized in that, The step of training the model to be trained for any type of traffic participant using multiple training samples to obtain a prediction model for any type of traffic participant includes: The third behavior information corresponding to any type of traffic participant is used to extract features to obtain third behavior features, wherein the behavioral characteristics of the third behavior features corresponding to different types of traffic participants are different; Using the third behavioral feature corresponding to any type of traffic participant as the input to the training model corresponding to any type of traffic participant, the predicted behavioral information of the output of the training model corresponding to any type of traffic participant is obtained. Using the fourth behavior information and the corresponding predicted behavior information corresponding to any type of traffic participant, determine the training loss of the model to be trained corresponding to any type of traffic participant; The training loss of the training model corresponding to each of the at least two types of traffic participants is weighted and calculated to obtain the target training loss. If the target training loss satisfies the training termination condition, then the training model corresponding to any type of traffic participant is used as the prediction model corresponding to any type of traffic participant. If the target training loss does not meet the training termination condition, then the training model corresponding to any type of traffic participant will continue to be trained using multiple training samples corresponding to any type of traffic participant until the target training loss meets the training termination condition.

7. The method according to claim 6, characterized in that, The step of using the third behavioral feature corresponding to any type of traffic participant as input to the training model corresponding to any type of traffic participant to obtain the predicted behavioral information of the output of the training model corresponding to any type of traffic participant includes: The third behavioral feature corresponding to any type of traffic participant is used as the input to the training model corresponding to any type of traffic participant. The training model corresponding to any type of traffic participant makes a prediction based on the third behavioral feature corresponding to any type of traffic participant to obtain a fourth behavioral feature. The behavioral characteristics of the fourth behavioral feature corresponding to different types of traffic participants are different. Based on the fourth behavioral feature corresponding to any of the traffic participants, predictive behavioral information corresponding to any of the traffic participants is determined.

8. A device for predicting the behavior information of traffic participants, characterized in that, include: The acquisition module is used to acquire the first behavior information of at least two types of traffic participants within a preset range. The feature extraction module is used to extract features from the first behavior information corresponding to any one of at least two types of traffic participants, to obtain the first behavior feature corresponding to the first behavior feature of the first behavior feature corresponding to the first behavior feature of different types of traffic participants. The prediction module is used to predict, based on the first behavioral feature corresponding to the traffic participant, any traffic participant of any type using the prediction model corresponding to the traffic participant of any type, to obtain the second behavioral information corresponding to the traffic participant of any type, wherein the second behavioral information includes the behavioral information following the first behavioral information.

9. An electronic device, characterized in that, Includes processor and memory, of which: Memory, used to store computer programs; A processor for executing a program stored in memory to implement the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1-7.

Images