Multi-traffic participant modeling method and system fusing intention reasoning and density gradient

Abstract

The invention relates to the technical field of traffic scene simulation, and discloses a multi-traffic participant modeling method and system fusing intention reasoning and density gradient, and the method comprises the steps: obtaining the dynamic data of multiple traffic participants in a complex traffic scene, and the multiple traffic participants comprise motor vehicles, non-motor vehicles and pedestrians; constructing a motor vehicle behavior model according to the dynamic data of the motor vehicle, wherein the motor vehicle behavior model predicts an overall behavior intention and a specific behavior intention of the motor vehicle; and according to the dynamic data of the non-motor vehicles and the dynamic data of the pedestrians, constructing a non-motor vehicle behavior model and a pedestrian behavior model based on the behaviors of the non-motor vehicles, the behaviors of the pedestrians and the traffic density. According to the method, the behaviors of the traffic participants can be modeled, and the simulation effect of pedestrians and vehicles is improved.

Description

Technical Field

[0001] This invention relates to the field of traffic scene simulation technology, and in particular to a multi-traffic participant modeling method and system that integrates intent reasoning and density gradient. Background Technology

[0002] Autonomous driving simulation testing is the foundation and key to autonomous vehicle testing and experimentation. In autonomous driving simulation testing, the establishment of traffic participant models is an essential part. The establishment of traffic participant models can provide scenarios that are more in line with actual roads and have a variety of complex traffic flows during testing.

[0003] However, existing methods for modeling traffic participants typically only consider the position and motion parameters of vehicles and pedestrians, without taking into account specific behaviors or real-time traffic density in different scenarios, which affects the simulation effect of the model. Summary of the Invention

[0004] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a multi-traffic participant modeling method and system that integrates intent reasoning and density gradient, which can take into account the specific behavior of multi-traffic participants and real-time traffic density in different scenarios, and improve the simulation effect of pedestrians and vehicles.

[0005] To address the aforementioned technical problems, this invention provides a multi-traffic participant modeling method that integrates intent reasoning and density gradient, comprising: To acquire dynamic data of multiple traffic participants in complex traffic scenarios, including motor vehicles, non-motor vehicles, and pedestrians; A motor vehicle behavior model is constructed based on the dynamic data of the motor vehicle, and the motor vehicle behavior model predicts the overall behavioral intention and specific behavioral intention of the motor vehicle. Based on the dynamic data of non-motorized vehicles and pedestrians, non-motorized vehicle behavior models and pedestrian behavior models are constructed based on the behavior of non-motorized vehicles, the behavior of pedestrians, and traffic density.

[0006] Furthermore, when predicting the overall behavioral intention of a motor vehicle, the posterior probability of each behavioral intention is calculated using Bayesian inference based on the current vehicle state and environmental information, and the behavioral intention with the highest posterior probability is taken as the prediction result of the overall behavioral intention. The specific behavioral intentions of motor vehicles include intersection passage decision intentions. When predicting the intersection passage decision intentions, machine learning is used to adjust and optimize the decisions in complex scenarios while adhering to traffic rules.

[0007] Furthermore, the use of machine learning to adjust and optimize decision-making in complex scenarios specifically involves: using a deep Q-network to adjust and optimize decision-making in complex scenarios. The defined state space includes the vehicle's own state, the states of surrounding vehicles, the states of pedestrians, and the state of traffic signals. The information in the state space is quantified into numerical values ​​and formed into state vectors, which are then input into the deep Q-network. The defined action space consists of the possible actions of the vehicle. Each possible action of the vehicle is quantified into a specific numerical value and formed into an action vector. Each possible action corresponds to a vector element in the action vector. The probability distribution or Q-value of the possible actions is output through the deep Q-network to obtain the intersection passage decision intention.

[0008] Furthermore, when using a deep Q-network to adjust and optimize decisions in complex scenarios, the constructed reward function includes a safety reward, an efficiency reward, and a rule compliance reward. The safety reward is as follows: if a motor vehicle does not collide while passing through an intersection, a positive reward is given; if a collision occurs, a negative reward is given and the current round is terminated. The efficiency reward is specifically as follows: a reward is given based on the time it takes for a vehicle to pass through the intersection. If the vehicle passes through the intersection quickly, a positive reward is given; if the waiting time at the intersection is too long, a negative reward is given. The reward for compliance with the rules is as follows: a positive reward is given for obeying traffic signals and traffic rules; a negative reward is given for violating the rules.

[0009] Furthermore, the step of constructing non-motorized vehicle behavior models and pedestrian behavior models based on the dynamic data of non-motorized vehicles and pedestrians, and on the basis of the behavior of non-motorized vehicles, the behavior of pedestrians, and traffic density, specifically involves: Traffic density regulation force is calculated based on traffic density, and the expected driving force, inter-individual repulsion force, and boundary or obstacle repulsion force on non-motorized vehicles and pedestrians are calculated by combining social force model and the behavior of non-motorized vehicles and pedestrians. The acceleration of non-motorized vehicles or pedestrians is calculated as follows: , In the formula, This represents the acceleration of the i-th non-motorized vehicle or pedestrian. Let represent the expected driving force experienced by the i-th non-motorized vehicle or pedestrian. This represents the inter-individual repulsive force experienced by the i-th non-motorized vehicle or pedestrian. This represents the repulsive force exerted by the boundary or obstacle on the i-th non-motorized vehicle or pedestrian. This represents the mass of the i-th non-motorized vehicle or pedestrian. It serves as a force for regulating traffic density.

[0010] Furthermore, the aforementioned The calculation method is as follows: , In the formula, A is the repulsion strength parameter, B is the attenuation control parameter, and exp is the natural exponential function. Let i be the minimum safe distance that is expected to be maintained between the i-th non-motorized vehicle or pedestrian and the j-th non-motorized vehicle or pedestrian. Let i be the actual distance between the i-th non-motorized vehicle or pedestrian and the j-th non-motorized vehicle or pedestrian. This represents the speed difference adjustment coefficient. This represents the velocity vector of the i-th non-motorized vehicle or pedestrian in the direction that is not connected to the target point. This represents the velocity vector of the j-th non-motorized vehicle or pedestrian in the direction that is not connected to the target point. Let be the unit vector pointing from the j-th non-motorized vehicle or pedestrian to the i-th non-motorized vehicle or pedestrian.

[0011] Furthermore, the aforementioned The calculation method is as follows: , In the formula, Let be the repulsion strength parameter of the k-th obstacle. Let be the repulsion attenuation coefficient of the k-th obstacle, and exp be the natural exponential function. Let be the minimum safe distance that is expected to be maintained between the i-th non-motorized vehicle or pedestrian and the k-th obstacle. Let i be the actual distance between the i-th non-motorized vehicle or pedestrian and the k-th obstacle. This represents the deviation angle of the k-th obstacle relative to the direction of movement of the i-th non-motorized vehicle or pedestrian. Let be the visibility function, used to control whether the k-th obstacle is within the line of sight of the i-th non-motorized vehicle or pedestrian. Let be the unit vector pointing from the k-th obstacle to the i-th non-motorized vehicle or pedestrian.

[0012] Furthermore, the aforementioned The calculation method is as follows: ; In the formula, This is the threshold for the field of view angle.

[0013] Furthermore, the aforementioned The calculation method is as follows: , In the formula, This represents the local traffic density at the current location of non-motorized vehicles or pedestrians. This represents the gradient of local traffic density at the current location of non-motorized vehicles or pedestrians. This is the local traffic density adjustment coefficient.

[0014] This invention also provides a multi-traffic participant modeling system that integrates intent reasoning and density gradient, comprising: The data acquisition module is used to acquire dynamic data of multiple traffic participants in complex traffic scenarios, including motor vehicles, non-motor vehicles, and pedestrians; The motor vehicle behavior model construction module is used to construct a motor vehicle behavior model based on the dynamic data of the motor vehicle, and the motor vehicle behavior model predicts the overall behavioral intention and specific behavioral intention of the motor vehicle. The non-motorized vehicle behavior model and pedestrian behavior construction model are used to construct non-motorized vehicle behavior models and pedestrian behavior models based on the dynamic data of non-motorized vehicles and pedestrians, and on the behavior of non-motorized vehicles, the behavior of pedestrians and traffic density.

[0015] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention acquires dynamic data of multiple traffic participants in complex traffic scenarios, constructs a motor vehicle behavior model including overall behavioral intent and specific behavioral intent, and constructs non-motor vehicle behavior models and pedestrian behavior models based on the behavior of non-motor vehicles, pedestrian behavior, and traffic density. By considering the specific behaviors of multiple traffic participants in different scenarios and real-time traffic density, it can naturally simulate interactive behaviors such as congestion, exclusion, and following in complex scenarios, thereby improving the simulation effect of pedestrians and vehicles. Attached Figure Description

[0016] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart of a method in a preferred embodiment of the present invention. Detailed Implementation

[0017] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0018] The purpose of this invention is to simulate a typical urban intersection scenario, including three types of traffic participants: motor vehicles, non-motor vehicles, and pedestrians. This intersection is a four-way intersection equipped with traffic lights, and each direction has lanes for going straight, turning left, and turning right. The surrounding area includes commercial, office, and residential areas, experiencing high traffic volume and complex traffic composition during morning and evening rush hours. (Refer to...) Figure 1 As shown, this invention discloses a multi-traffic participant modeling method that integrates intent reasoning and density gradient, comprising the following steps:

[0019] S1: Obtain dynamic data of multiple traffic participants in a complex traffic scenario, including motor vehicles, non-motor vehicles, and pedestrians.

[0020] S1-1: Dynamic data such as trajectories, speeds, and accelerations of motor vehicles, non-motor vehicles, and pedestrians are collected through real-world road test logs and a fusion perception system (sensors such as cameras and LiDAR deployed at intersections). Simultaneously, traffic light timing schemes at intersections are collected, including the duration and countdown information for green, yellow, and red lights in each direction.

[0021] S1-2: Dynamic element data is synchronized, reconstructed, and denoised to form unified dynamic element data for motor vehicles, non-motorized vehicles, and pedestrians, ensuring the accuracy and consistency of the data.

[0022] S2: Construct a motor vehicle behavior model based on the dynamic data of the motor vehicle, and the motor vehicle behavior model predicts the overall behavioral intention and specific behavioral intention of the motor vehicle.

[0023] S2-1: Constructing the overall behavioral intent of motor vehicles. Based on collected feature data such as vehicle speed, distance to the vehicle in front, and lane occupancy, the model predicts the vehicle's intent to go straight, change lanes, or pull over. For example, for a vehicle traveling in a straight lane and whose distance to the vehicle in front is gradually decreasing, the model predicts that it has a high probability of continuing to go straight.

[0024] In this embodiment, when constructing the overall behavioral intent of a motor vehicle, a Bayesian network intent recognition model based on state sequences is used, based on collected feature data such as vehicle speed, distance to the vehicle in front, and lane occupancy. A probability table is constructed to describe the probability of each node taking different states given a parent node. Based on the current vehicle state and environmental information (i.e., evidence variables), Bayesian inference methods (such as variable elimination algorithms) are used to calculate the posterior probability of each behavioral intent. The behavioral intent with the highest posterior probability is taken as the predicted result of the overall behavioral intent.

[0025] In using a state sequence-based Bayesian network intent recognition model, nodes represent vehicle behavior intent (e.g., going straight, changing lanes, parking) and various influencing factors (e.g., vehicle speed, distance to the vehicle in front, traffic signals, etc.); edges represent conditional dependencies between variables. For example, the node "vehicle speed" might influence the node "lane-changing intent," while "distance to the vehicle in front" might influence multiple nodes such as "going straight" and "lane-changing intent." The parameters of the Bayesian network (i.e., the probability values ​​in the probability table) can be learned from historical data. Using algorithms such as Expectation-Maximization (EM) to estimate the parameters, with a large amount of labeled vehicle behavior data, the probabilistic relationships between various variables can be effectively learned.

[0026] The specific prediction process is as follows: S2-1-1: Data Preprocessing. The collected vehicle status and environmental data are cleaned and normalized to remove outliers and erroneous data, and the data is converted into a format suitable for model input. For example, continuous values ​​such as vehicle speed and acceleration are standardized to have zero mean and unit variance.

[0027] S2-1-2: Feature Extraction. Extract features related to the vehicle's behavioral intent from the raw data. For Bayesian networks, various influencing factors can be extracted as node features, such as current speed, distance to the vehicle in front, and lane occupancy.

[0028] S2-1-3: Model Training. For Bayesian networks, the conditional probability parameters between each node are learned.

[0029] S2-1-4: Overall Behavioral Intent Prediction. In practical applications, real-time collected vehicle status and environmental data are input into a trained model to obtain the probability distribution of the vehicle's future behavioral intent. Depending on the application scenario, the intent with the highest probability can be selected as the prediction result, or all possible intents and their corresponding probabilities can be output to provide a reference for subsequent decision-making.

[0030] S2-2: Constructing the specific behavioral intentions of motor vehicles. The specific behavioral intentions of motor vehicles include longitudinal following intentions, lateral lane changing intentions, and intersection passage decision intentions.

[0031] S2-2-1: The longitudinal following intention is constructed using the Intelligent Driver Model (IDM). Based on the vehicle's current speed, distance to the vehicle in front, and speed difference, the vehicle's acceleration is calculated to simulate the longitudinal following intention of the vehicle while waiting in line at a traffic light.

[0032] The acceleration formula is: , In the formula, It is the acceleration of the motor vehicle. It is the maximum acceleration of a motor vehicle. It is the current speed of the motor vehicle. It is the desired speed of the motor vehicle. These are model parameters used to adjust the effect of vehicle speed on acceleration. It is the distance between the vehicle and the vehicle in front. It is the speed difference between the motor vehicle and the vehicle in front. It is the expected safe distance. The calculation formula is: , In the formula, It is the minimum safe distance. It is the reaction time of the motor vehicle. It is the maximum deceleration of the vehicle in front.

[0033] Based on the vehicle's current speed, distance to the vehicle in front, and speed difference, the vehicle's acceleration is calculated to simulate the following behavior of a vehicle waiting in line at a traffic light.

[0034] S2-2-2: The intention of lateral lane changing is to simulate the determination of whether to change lanes to find a faster route by using the lateral lane changing model (Minimizing Overall Braking Induced by Lane changes, MOBIL) when traffic is congested.

[0035] The criteria for determining whether a motor vehicle is changing lanes on the left are: , In the formula, This refers to the expected speed change of a vehicle after it changes lanes to the left lane. It refers to the impact of changing lanes on the speed of vehicles behind. This is a courtesy factor used to measure the degree of consideration a vehicle gives to vehicles behind it when changing lanes. If the conditions are met, the vehicle changes lanes to the left lane. The conditions for determining whether a vehicle changes lanes to the right are similar to those for determining whether it changes lanes to the left. In this embodiment, the courtesy factor is set to 0.4, which makes lane-changing behavior more conservative and consistent with actual observations.

[0036] During traffic congestion, the simulated vehicle determines whether to change lanes to find a faster route based on the speed and spacing of vehicles in adjacent lanes.

[0037] S2-2-3: The intersection traffic decision-making intent uses a rule-based approach as the basic decision-making framework to ensure that the basic traffic rules for motor vehicles are followed. At the same time, machine learning methods are introduced as an optimization module. Under the premise of following traffic rules, machine learning is used to adjust and optimize the decision-making in complex scenarios. When the green light is on, motor vehicles are allowed to pass according to the rules. Meanwhile, the machine learning model fine-tunes the acceleration of motor vehicles based on the status of surrounding motor vehicles and pedestrians to ensure safe and efficient passage through the intersection.

[0038] Fusion strategy: When traffic signal rules permit passage, prioritize following traffic rules; when encountering special circumstances (such as interference from surrounding vehicles and pedestrians), activate machine learning models for decision optimization.

[0039] Priority Setting: Set the rule priority to be higher than the machine learning decision. When the decision output by the machine learning model conflicts with the rule, the rule decision prevails. For example, due to the limitations of the training data, the machine learning model may suggest that a vehicle run a red light in some rare scenarios. In this case, the rule module will immediately reject the decision and force the vehicle to stop and wait.

[0040] Information Interaction: Rule-based and machine learning methods share data such as vehicle status and environmental information. The rule module provides initial decision guidance for vehicles based on traffic regulations and common sense. The machine learning module further optimizes the decision results by combining historical data and real-time scenario information. For example, if the rule module determines that a vehicle should cross the intersection with 2 seconds remaining on the green light, the machine learning module will comprehensively consider the driving status of surrounding vehicles and the optimal decisions under similar historical scenarios, fine-tuning the vehicle's acceleration to ensure the vehicle crosses the intersection safely and efficiently.

[0041] In this embodiment, the rules formulated according to traffic regulations include traffic signal rules, intersection priority rules, vehicle spacing and speed rules, and traffic light control logic.

[0042] The specific traffic signal rules are as follows: Decisions are made based on the traffic light status. When the light is green, motor vehicles are allowed to proceed straight or turn in the permitted direction; when the light is yellow, vehicles already in the intersection may continue, while those not yet in the intersection should slow down and stop; when the light is red, motor vehicles must stop and wait. The countdown timer information is also considered: when the remaining time on the green light countdown is short (e.g., less than 3 seconds) and the vehicle is close to the stop line (e.g., less than 10 meters), the vehicle is allowed to proceed quickly; if the distance is greater, it is recommended to slow down and stop.

[0043] The specific rules for priority passage at intersections are as follows: vehicles on the main road have priority over vehicles on the side road; turning vehicles must yield to vehicles going straight; when turning from opposite directions, left-turning vehicles must yield to vehicles going straight and right-turning vehicles. At intersections without traffic signals, follow traffic rules such as yield signs and markings, for example, "yield" markings indicate that vehicles must yield to vehicles coming from other directions.

[0044] The specific rules for vehicle spacing and speed are as follows: When a vehicle approaches an intersection, if the distance between it and the vehicle in front is less than the safe distance (which can be dynamically calculated based on the vehicle speed; for example, when the vehicle speed is v, the safe distance is v×t, where t is the reaction time), then the vehicle should decelerate to maintain the safe distance; if the distance is sufficient and the traffic signal allows, the vehicle may accelerate appropriately to pass through.

[0045] The traffic light control logic is as follows: In a simulated urban intersection scenario, each intersection is equipped with a traffic signal controller responsible for managing the traffic light status and countdown time. Vehicles perceive the traffic light color and countdown information through sensors (such as cameras and lidar). When a vehicle is within a certain range (e.g., 30 meters) of the intersection stop line, it begins real-time monitoring of the traffic light status. The traffic signal controller switches the traffic light status according to a preset timing scheme. For example, east-west traffic has a 30-second green light, a 3-second yellow light, and a 37-second red light; north-south traffic has a 37-second green light, a 3-second yellow light, and a 30-second red light. Vehicles make passage decisions based on the received traffic light information, their own position, speed, and other status factors, combined with the above rules.

[0046] In this embodiment, machine learning is used to adjust and optimize the decision-making in complex scenarios. Specifically, a Deep Q-Network (DQN) is used to adjust and optimize the decision-making in complex scenarios. DQN combines Q-learning and deep neural networks, and can handle high-dimensional state space and action space, making it suitable for complex intersection traffic decision-making scenarios.

[0047] The defined state space includes the vehicle's own state, the states of surrounding vehicles, the pedestrian state, and the traffic signal state. The vehicle's own state includes its position (accurate to meters) from the intersection stop line, current speed (meters per second), acceleration (meters per second²), and direction of travel (straight, left turn, right turn). The surrounding vehicle state obtains the distance, speed, and direction of travel of surrounding vehicles (such as vehicles in adjacent lanes). For example, the position of the vehicle in front from the intersection stop line, its relative speed to the vehicle, and the approaching speed of vehicles behind. The pedestrian state detects the presence of pedestrians within the intersection area, their position (e.g., their position on the crosswalk), their walking direction, and their speed. The traffic signal state includes the traffic light color (red, yellow, green) and the countdown time (in seconds). Information in the state space is quantified into numerical values ​​and input into a deep Q-network to form a state vector. For example, the position of the distance from the parking line at the intersection can be a value in the range of 0-50 meters, the speed can be a value in the range of 0-30 m / s, and the acceleration can be a value in the range of -5-5 m / s². The traffic light colors are represented by one-hot encoding (red light: [1,0,0], yellow light: [0,1,0], green light: [0,0,1]), and the countdown time is normalized to the range of 0-1.

[0048] The defined action space comprises the possible actions of a motor vehicle, including acceleration, deceleration, maintaining speed, left turn, straight ahead, and right turn. Each possible action is quantified into a specific numerical value; for example, acceleration represents an acceleration of 2 m / s², deceleration represents an acceleration of -3 m / s², maintaining speed represents an acceleration of 0 m / s², and left turn, straight ahead, and right turn represent the vehicle's direction of travel. Action vectors are formed, with each possible action corresponding to a vector element in the action vector. A deep Q-network is used to output the probability distribution or Q-value of the possible actions, thus obtaining the intersection traffic decision intention.

[0049] The constructed reward function includes safety rewards, efficiency rewards, and rule compliance rewards. Specifically, the safety reward is as follows: if a vehicle passes through an intersection without a collision, a positive reward (e.g., +1) is given; if a collision occurs, a negative reward (e.g., -100) is given, and the current round terminates. The efficiency reward is as follows: rewards are given based on the time it takes for a vehicle to pass through the intersection. A fast passage through the intersection results in a positive reward (e.g., -0.1 per second; fast passage reduces the total reward to a negative value, equivalent to a positive incentive); a long waiting time at the intersection results in a negative reward (e.g., -0.5 per second). The rule compliance reward is as follows: if traffic signals and rules are obeyed (e.g., not running red lights, not violating right-of-way rules), a positive reward (e.g., +0.5 per compliant action) is given; if rules are violated, a negative reward (e.g., -10 per violation) is given.

[0050] In this embodiment, a large amount of intersection traffic scenario data is collected, including vehicle states, surrounding environment information, and corresponding optimal decisions (which can be achieved through expert demonstration data or simulated data). This data is used to pre-train a deep Q-network, initializing its parameters. In a simulated environment, vehicles select actionable actions based on their current state using the deep Q-network, interacting with the environment to obtain reward signals and the next state. Experiences (state, action, reward, next state) are stored in an experience replay buffer. At regular time steps, a small batch of experiences is randomly selected from the experience replay buffer to train the deep Q-network. The network parameters are updated by optimizing the loss function (e.g., mean squared error loss function, calculating the difference between the predicted Q-value and the target Q-value), allowing the model to gradually learn the optimal traffic strategy.

[0051] During training, hyperparameters such as the exploration rate and learning rate are continuously adjusted to balance the exploration and utilization capabilities of the deep Q-network. Initially, the exploration rate is high, and the probability of the vehicle randomly selecting a move is relatively large. As training progresses, the exploration rate is gradually reduced, allowing the vehicle to make decisions more based on the optimal moves predicted by the model.

[0052] By integrating rule-based and machine learning methods, when the green light is on, motor vehicles are allowed to pass according to the rules. At the same time, the machine learning model fine-tunes the acceleration of motor vehicles based on the status of surrounding motor vehicles and pedestrians to ensure safe and efficient passage through the intersection.

[0053] S3: Based on the dynamic data of non-motorized vehicles and pedestrians, construct non-motorized vehicle behavior models and pedestrian behavior models based on the behavior of non-motorized vehicles, the behavior of pedestrians, and traffic density.

[0054] This invention simulates the interaction forces between different traffic participants using a Social Force Model (SFM). For example, pedestrians will actively avoid approaching vehicles due to the repulsive force generated; non-motorized vehicles in high traffic density areas are affected by density adjustment terms and choose to slow down or deviate from their paths. By combining the effects of various forces through combined behavioral drivers, the invention simulates the behaviors of non-motorized vehicles and pedestrians, such as congestion, repulsion, and following, thereby improving the realism and effectiveness of simulation tests.

[0055] The forces acting on non-motorized vehicles and pedestrians include expected driving force, inter-individual repulsive force, boundary or obstacle repulsive force, and traffic density regulating force. The acceleration of non-motorized vehicles or pedestrians is: , In the formula, This represents the acceleration of the i-th non-motorized vehicle or pedestrian. Let represent the expected driving force experienced by the i-th non-motorized vehicle or pedestrian. This represents the inter-individual repulsive force experienced by the i-th non-motorized vehicle or pedestrian. This represents the repulsive force exerted by the boundary or obstacle on the i-th non-motorized vehicle or pedestrian. This represents the mass of the i-th non-motorized vehicle or pedestrian. It serves as a force for regulating traffic density.

[0056] Expectation drives individuals toward their goals and enables them to reach the desired speed. The calculation method is as follows: , In the formula, For the desired speed, Let i be the current speed of the i-th non-motorized vehicle or pedestrian. To adjust the time.

[0057] Inter-individual repulsive forces simulate avoidance behavior between individuals, preventing collisions or excessive closeness. The calculation method is as follows: , In the formula, A is the repulsion strength parameter, B is the attenuation control parameter, and exp is the natural exponential function. Let i be the minimum safe distance that is expected to be maintained between the i-th non-motorized vehicle or pedestrian and the j-th non-motorized vehicle or pedestrian. Let i be the actual distance between the i-th non-motorized vehicle or pedestrian and the j-th non-motorized vehicle or pedestrian. This represents the speed difference adjustment coefficient. This represents the velocity vector of the i-th non-motorized vehicle or pedestrian in the direction that is not connected to the target point. This represents the velocity vector of the j-th non-motorized vehicle or pedestrian in the direction that is not connected to the target point. Let be the unit vector pointing from the j-th non-motorized vehicle or pedestrian to the i-th non-motorized vehicle or pedestrian.

[0058] Speed ​​difference adjustment coefficient Used to quantify the contribution weight of relative speed to the psychological stress of avoidance, low When the value is high, it simulates a "sluggish" or "calm" behavior, meaning that even if a vehicle approaches rapidly from behind, the individual in front will not exhibit a violent evasive reaction. Value: Simulates a "vigilant" or "sensitive" behavior, generating greater psychological pressure towards rapidly approaching targets, thus producing a stronger repulsion and avoidance. Empirical value range. The value is typically set between 0.1 and 0.8. In typical urban intersection scenarios, a value of 0.5 is preferred, as it effectively simulates the pre-deceleration behavior of non-motorized vehicles when facing a rapidly approaching lateral target. Scene Adaptive Settings The adjustment is dynamically based on traffic density ρ. In high-density scenarios, the value is appropriately reduced. A value (e.g., 0.2) is used to simulate an individual's "forced tolerance" to speed differences in a crowded environment; in low-density scenarios, the value is increased. A value (e.g., 0.7) can be used to increase the simulated safety warning threshold.

[0059] Boundary or obstacle repulsion simulates the avoidance behavior of traffic participants towards fixed facilities such as road boundaries and medians, preventing traffic participants from hitting walls or running off the road. The calculation method is as follows: , In the formula, Let be the repulsion strength parameter of the k-th obstacle. Let be the repulsion attenuation coefficient of the k-th obstacle. Let be the minimum safe distance that is expected to be maintained between the i-th non-motorized vehicle or pedestrian and the k-th obstacle. Let i be the actual distance between the i-th non-motorized vehicle or pedestrian and the k-th obstacle. This represents the deviation angle of the k-th obstacle relative to the direction of movement of the i-th non-motorized vehicle or pedestrian. Let be the visibility function, used to control whether the k-th obstacle is within the line of sight of the i-th non-motorized vehicle or pedestrian. Let be the unit vector pointing from the k-th obstacle to the i-th non-motorized vehicle or pedestrian.

[0060] The calculation method is as follows: ; In the formula, This is the threshold for the field of view angle.

[0061] When the local traffic density exceeds a certain threshold This will prompt non-motorized vehicles to actively choose to slow down or deviate from their path to avoid congested areas. The calculation method is as follows: , In the formula, This represents the local traffic density at the current location of non-motorized vehicles or pedestrians. This represents the gradient of local traffic density at the current location of non-motorized vehicles or pedestrians. This is the local traffic density adjustment coefficient.

[0062] The local traffic density adjustment coefficient determines the sensitivity of traffic participants to congestion perception. High Value: Simulates a congestion-averse personality. When local density increases slightly, participants will exhibit a strong intention to slow down or change direction to avoid congested areas. Low Value: Simulates highly tolerant or forced-following behavior. Even with high density ahead, participants maintain their original paths, commonly seen in non-motorized vehicle lanes during peak hours when congestion is severe.

[0063] According to experimental calibration, the adjustment coefficient The value range is usually between 0.5 and 10.0.

[0064] Pedestrian model: Pedestrians are typically set to a higher value (e.g., 5.0–8.0) because they are more flexible in their movement and have a stronger desire to avoid congestion.

[0065] Non-motorized vehicle model: Typically set to a smaller value (e.g., 1.0–3.0), its physical space for avoiding congestion is limited by lane width and driving inertia.

[0066] This invention also discloses a multi-traffic participant modeling system that integrates intent reasoning and density gradient, comprising: The data acquisition module is used to acquire dynamic data of multiple traffic participants in complex traffic scenarios, including motor vehicles, non-motor vehicles, and pedestrians; The motor vehicle behavior model construction module is used to construct a motor vehicle behavior model based on the dynamic data of the motor vehicle, and the motor vehicle behavior model predicts the overall behavioral intention and specific behavioral intention of the motor vehicle. The non-motorized vehicle behavior model and pedestrian behavior construction model are used to construct non-motorized vehicle behavior models and pedestrian behavior models based on the dynamic data of non-motorized vehicles and pedestrians, and on the behavior of non-motorized vehicles, the behavior of pedestrians and traffic density.

[0067] The present invention also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements a multi-traffic participant modeling method that integrates intent reasoning and density gradient.

[0068] The present invention also discloses an apparatus including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a multi-traffic participant modeling method that integrates intent reasoning and density gradient.

[0069] This invention acquires dynamic data of multiple traffic participants in complex traffic scenarios, constructs a motor vehicle behavior model including overall behavioral intent and specific behavioral intent, and constructs non-motor vehicle behavior models and pedestrian behavior models based on the behavior of non-motor vehicles, pedestrian behavior, and traffic density. By considering the specific behaviors of multiple traffic participants in different scenarios and real-time traffic density, it can naturally simulate interactive behaviors such as congestion, exclusion, and following in complex scenarios, thereby improving the simulation effect of pedestrians and vehicles.

[0070] To further demonstrate the beneficial effects of this invention, in this embodiment, a simulation environment consistent with the actual intersection scenario is set up on an autonomous driving simulation platform (such as SUMO / CARLA), including road layout and traffic light configuration. The traffic participant model obtained using the method of this invention is imported into the simulation platform to reproduce the traffic scenario of a typical intersection in a city during the evening rush hour (17:30-18:30). The simulation is compared with real-world road test data and traditional models that only consider vehicle dynamics and do not introduce behavior and traffic density. The performance in simulating traffic flow, traffic efficiency, and safety is analyzed. The simulation comparison data is shown in Table 1.

[0071] Table 1. Comparison of the present invention with real data and traditional models.

[0072] As shown in Table 1, the model of this invention closely matches real-world traffic flow simulations for both motorized and non-motorized vehicles, with a relative error controlled within 2.5%. This is because traditional models neglect the interference of non-motorized vehicles with motorized vehicles (such as non-motorized vehicles occupying lanes), resulting in an overestimation of the simulated motorized vehicle throughput (1550 vehicles), which is an overly idealized state. In contrast, this invention successfully reproduces the mutual constraints in mixed traffic flow by introducing repulsive forces from the social force model and various interference mechanisms from the non-motorized vehicle behavior model, making the traffic flow data closer to the real congestion environment.

[0073] Regarding the average delay metric, the average delay in real data is 35.4 seconds, while the simulation result of this invention is 34.2 seconds, accurately reflecting the delay caused by congestion; whereas the delay calculated by the traditional model is only 22.1 seconds, which is seriously distorted. In the right-turn lane scenario, this invention successfully simulates the deceleration and stopping behavior of motor vehicles due to yielding to pedestrians crossing the street (cluster movement in the pedestrian behavior model). In this invention, the average speed fluctuation variance (VelocityVariance) of vehicles passing through the intersection is 1.8 m / s², which is very close to the real data of 2.1 m / s², proving that the model captures the microscopic characteristics of frequent vehicle starts and stops.

[0074] When evaluating safety performance, Time To Collision (TTC) is used as a safety evaluation metric. In real-world scenarios, 42 minor traffic conflicts (TTC < 3 seconds) were observed, primarily concentrated in areas with mixed motorized and non-motorized traffic. This invention successfully reproduced 38 similar conflict scenarios, achieving a conflict reproduction rate of 90.4%. In contrast, traditional models, assuming strict adherence to rules and no interference between vehicles, only recorded 5 conflicts, making them unsuitable for evaluating the safety of autonomous driving in real-world road conditions. Furthermore, the simulation successfully modeled an electric bicycle cutting into a straight-ahead motorized vehicle lane while making a left turn, triggering emergency braking of the motorized vehicle (IDM model response). This behavior aligns with the "agile and aggressive" driving characteristics of non-motorized vehicles at real intersections. This demonstrates the beneficial effects of this invention.

[0075] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0076] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0077] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0078] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0079] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A multi-traffic participant modeling method integrating intent reasoning and density gradient, characterized in that, include: To acquire dynamic data of multiple traffic participants in complex traffic scenarios, including motor vehicles, non-motor vehicles, and pedestrians; A motor vehicle behavior model is constructed based on the dynamic data of the motor vehicle, and the motor vehicle behavior model predicts the overall behavioral intention and specific behavioral intention of the motor vehicle. Based on the dynamic data of non-motorized vehicles and pedestrians, non-motorized vehicle behavior models and pedestrian behavior models are constructed based on the behavior of non-motorized vehicles, the behavior of pedestrians, and traffic density.

2. The multi-traffic participant modeling method integrating intent reasoning and density gradient as described in claim 1, characterized in that: When predicting the overall behavioral intention of a motor vehicle, the posterior probability of each behavioral intention is calculated using Bayesian inference based on the current vehicle state and environmental information, and the behavioral intention with the highest posterior probability is taken as the prediction result of the overall behavioral intention. The specific behavioral intentions of motor vehicles include intersection passage decision intentions. When predicting the intersection passage decision intentions, machine learning is used to adjust and optimize the decisions in complex scenarios while adhering to traffic rules.

3. The multi-traffic participant modeling method integrating intent reasoning and density gradient according to claim 2, characterized in that: The method of using machine learning to adjust and optimize decision-making in complex scenarios specifically involves: using a deep Q-network to adjust and optimize decision-making in complex scenarios. The defined state space includes the state of the vehicle itself, the state of surrounding vehicles, the state of pedestrians, and the state of traffic signals. The information in the state space is quantified into numerical values ​​and formed into state vectors, which are then input into the deep Q-network. The defined action space is the possible actions of the vehicle. Each possible action of the vehicle is quantified into a specific numerical value and formed into an action vector. Each possible action corresponds to a vector element in the action vector. The probability distribution or Q-value of the possible actions is output through the deep Q-network to obtain the intersection passage decision intention.

4. The multi-traffic participant modeling method integrating intent reasoning and density gradient as described in claim 3, characterized in that: When using deep Q-networks to adjust and optimize decisions in complex scenarios, the constructed reward function includes safety rewards, efficiency rewards, and rule compliance rewards. The safety reward is as follows: if a motor vehicle does not collide while passing through an intersection, a positive reward is given; if a collision occurs, a negative reward is given and the current round is terminated. The efficiency reward is specifically as follows: a reward is given based on the time it takes for a vehicle to pass through the intersection. If the vehicle passes through the intersection quickly, a positive reward is given; if the waiting time at the intersection is too long, a negative reward is given. The reward for compliance with the rules is as follows: a positive reward is given for obeying traffic signals and traffic rules; a negative reward is given for violating the rules.

5. The multi-traffic participant modeling method integrating intent reasoning and density gradient according to claim 1, characterized in that: The step of constructing non-motorized vehicle behavior models and pedestrian behavior models based on the dynamic data of non-motorized vehicles and pedestrians, and on the basis of the behavior of non-motorized vehicles, the behavior of pedestrians, and traffic density, specifically involves: Traffic density regulation force is calculated based on traffic density, and the expected driving force, inter-individual repulsion force, and boundary or obstacle repulsion force on non-motorized vehicles and pedestrians are calculated by combining social force model and the behavior of non-motorized vehicles and pedestrians. The acceleration of non-motorized vehicles or pedestrians is calculated as follows: ， In the formula, This represents the acceleration of the i-th non-motorized vehicle or pedestrian. Let represent the expected driving force experienced by the i-th non-motorized vehicle or pedestrian. This represents the inter-individual repulsive force experienced by the i-th non-motorized vehicle or pedestrian. This represents the repulsive force exerted by the boundary or obstacle on the i-th non-motorized vehicle or pedestrian. This represents the mass of the i-th non-motorized vehicle or pedestrian. It serves as a force for regulating traffic density.

6. The multi-traffic participant modeling method integrating intent reasoning and density gradient according to claim 5, characterized in that: The The calculation method is as follows: ， In the formula, A is the repulsion strength parameter, B is the attenuation control parameter, and exp is the natural exponential function. Let i be the minimum safe distance that is expected to be maintained between the i-th non-motorized vehicle or pedestrian and the j-th non-motorized vehicle or pedestrian. Let i be the actual distance between the i-th non-motorized vehicle or pedestrian and the j-th non-motorized vehicle or pedestrian. This represents the speed difference adjustment coefficient. This represents the velocity vector of the i-th non-motorized vehicle or pedestrian in the direction that is not connected to the target point. This represents the velocity vector of the j-th non-motorized vehicle or pedestrian in the direction that is not connected to the target point. Let be the unit vector pointing from the j-th non-motorized vehicle or pedestrian to the i-th non-motorized vehicle or pedestrian.

7. The multi-traffic participant modeling method integrating intent reasoning and density gradient according to claim 5, characterized in that: The The calculation method is as follows: ， In the formula, Let be the repulsion strength parameter of the k-th obstacle. Let be the repulsion attenuation coefficient of the k-th obstacle, and exp be the natural exponential function. Let be the minimum safe distance that is expected to be maintained between the i-th non-motorized vehicle or pedestrian and the k-th obstacle. Let i be the actual distance between the i-th non-motorized vehicle or pedestrian and the k-th obstacle. This represents the deviation angle of the k-th obstacle relative to the direction of movement of the i-th non-motorized vehicle or pedestrian. Let be the visibility function, used to control whether the k-th obstacle is within the line of sight of the i-th non-motorized vehicle or pedestrian. Let be the unit vector pointing from the k-th obstacle to the i-th non-motorized vehicle or pedestrian.

8. The multi-traffic participant modeling method integrating intent reasoning and density gradient according to claim 7, characterized in that: The The calculation method is as follows: ； In the formula, This is the threshold for the field of view angle.

9. The multi-traffic participant modeling method integrating intent reasoning and density gradient according to claim 5, characterized in that: The The calculation method is as follows: ， In the formula, This represents the local traffic density at the current location of non-motorized vehicles or pedestrians. This represents the gradient of local traffic density at the current location of non-motorized vehicles or pedestrians. This is the local traffic density adjustment coefficient.

10. A multi-traffic participant modeling system integrating intent reasoning and density gradient, characterized in that, include: The data acquisition module is used to acquire dynamic data of multiple traffic participants in complex traffic scenarios, including motor vehicles, non-motor vehicles, and pedestrians; The motor vehicle behavior model construction module is used to construct a motor vehicle behavior model based on the dynamic data of the motor vehicle, and the motor vehicle behavior model predicts the overall behavioral intention and specific behavioral intention of the motor vehicle. The non-motorized vehicle behavior model and pedestrian behavior construction model are used to construct non-motorized vehicle behavior models and pedestrian behavior models based on the dynamic data of non-motorized vehicles and pedestrians, and on the behavior of non-motorized vehicles, the behavior of pedestrians and traffic density.

Images