A pedestrian behavior simulation method and system based on intention and attention mechanism

Abstract

The application discloses a kind of pedestrian behavior simulation method and system based on intention and attention mechanism, belong to unmanned driving simulation technical field.The application is to solve the technical problem that social psychological factor is missing in pedestrian behavior modeling and dynamic interaction logic is not strong, by global path planning based on pedestrian attribute and traffic environment information to obtain temporary target point;According to the attention weight distribution of pedestrian individual characteristics and traffic environment elements, obtain decision-related environmental information;Based on temporary target point and decision-related environmental information, the motion state of the pedestrian is determined, and the motion decision of the next moment is obtained.The application realizes the overall modeling of pedestrian behavior decision-making process, significantly improves the rationality and explainability of simulation, and can be widely applied to automatic driving, remote cooperative control and robot navigation scene.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving simulation technology, specifically relating to a pedestrian behavior simulation method and system based on intent and attention mechanisms. Background Technology

[0002] In the field of autonomous driving simulation, behavioral modeling of pedestrian agents has always been one of the key research focuses and challenges. Compared with other traffic participants such as vehicles, pedestrians have a series of significant and unique behavioral characteristics, making their behavior more complex and diverse.

[0003] On the one hand, pedestrians exhibit significantly greater speed variation and turning maneuverability than vehicles, with more flexible and complex trajectories, and are generally unconstrained by lane structures. This high degree of maneuverability allows pedestrians to move freely in narrow or high-density spaces, while vehicles must strictly adhere to road structures and traffic rules. On the other hand, pedestrians have a higher tolerance for minor physical contact; minor collisions in crowded environments are generally acceptable, whereas similar situations are often considered traffic accidents in vehicle interactions. This difference enables pedestrians to maintain small individual distances and continue moving in dense environments. Furthermore, pedestrians have a significantly shorter reaction time from a stationary state to full-speed movement than vehicles, allowing them to quickly fill gaps and efficiently utilize available space, especially in high-density urban environments.

[0004] Meanwhile, pedestrian behavior exhibits significant social characteristics, often manifesting as complex group behaviors such as walking in groups, queuing, and pushing and shoving. The formation, maintenance, and dispersal of pedestrian groups directly influence individual movement speed, path selection, and perception and response to the surrounding environment. This dynamic group behavior significantly increases the complexity of pedestrian behavior modeling and indicates that pedestrian modeling methods need to differ from traditional vehicle behavior modeling paradigms to characterize their unique individual and group behavioral features.

[0005] In traffic simulation research, traditional pedestrian behavior modeling methods often draw on dynamic concepts from physics, with the "Social Force Model" (SFM) being one of the most representative. This model, based on the concept of physical forces, explains and predicts pedestrian motion by introducing the interaction forces between pedestrians and between pedestrians and their environment. In SFM, each pedestrian is considered a force-acting particle, and their motion is determined by the combined action of multiple "social forces," primarily including: the attraction of the destination to the pedestrian, the repulsive force between pedestrians to avoid collisions, and the repulsive force between the pedestrian and environmental obstacles (such as walls and fences). The resultant force of these forces determines the pedestrian's acceleration and velocity changes at any given time.

[0006] While the social force model (SFM) has achieved some success, it still has significant limitations. First, it relies excessively on abstract representations of physical forces, failing to adequately characterize the social and psychological factors in pedestrian behavior, such as long-term behavioral intentions, individual differences, emotional states, and attention distribution. In reality, human drivers can distinguish between pedestrians attempting to hail a taxi, those distracted by mobile phones, and those trying to cross traffic, making accurate judgments accordingly. These high-level semantic intentions have limited impact on low-level autonomous driving systems but are crucial for high-level systems. Second, SFM performs poorly in complex interaction scenarios (such as multi-group convergence or multi-directional pedestrian flow). Furthermore, in actual simulation platforms, SFM often relies on manual settings or independent path planning modules to pre-determine routes for the agent, ignoring the dynamic influence of other traffic participants' behaviors on path selection and pedestrian decision-making, thus limiting the realism of the simulation results.

[0007] In view of the above problems, in order to more accurately predict and characterize pedestrian behavior in the real world, there is an urgent need for a simulation scene modeling method that can comprehensively reflect the true behavioral intentions of pedestrians. Summary of the Invention

[0008] The purpose of this invention is to address the technical problems of traditional pedestrian simulation methods, such as over-reliance on physical force abstraction, difficulty in depicting social psychological factors, and lack of dynamic interactive feedback. This invention proposes a pedestrian behavior simulation method and system based on intention and attention mechanisms. By introducing a long-term and short-term behavioral intention and attention weight allocation mechanism under a unified modeling framework, the invention aims to improve the rationality of the simulation, the diversity of behaviors, and make the generated trajectories more consistent with the real human behavior patterns.

[0009] To achieve the above objectives, the present invention adopts the following technical solution.

[0010] A pedestrian behavior simulation method based on intent and attention mechanisms includes the following steps: Global path planning is performed based on pedestrian attributes and traffic environment information to obtain temporary target points for pedestrians; Attention weights are allocated based on individual pedestrian characteristics and traffic environment factors to obtain decision-related environmental information. Based on the pedestrian's temporary target point and the decision-related environmental information, the pedestrian's movement state is determined, and the pedestrian's movement decision for the next moment is obtained.

[0011] Furthermore, based on pedestrian attributes and traffic environment information, global path planning is performed to obtain temporary pedestrian target points, including: Construct a traffic environment graph that includes reachable location nodes and their connections; Starting from the pedestrian's current location, search for a path to the long-term behavioral goal in the traffic environment map; The node values ​​are adjusted in real time according to changes in the traffic environment, and the path is updated to obtain the temporary target point for pedestrians.

[0012] Furthermore, a traffic environment map is constructed, including: Obtain information on obstacle distribution and passable areas in the traffic environment; The passable area is divided into multiple nodes representing locations, and edges are established based on the connectivity between the nodes.

[0013] Furthermore, the route is updated based on changes in the traffic environment, including: The dynamic impact zone of pedestrians is determined based on their spatial needs and psychological state. When the influence disks of multiple pedestrians intersect, their positions are adjusted according to priority, and the path planning results are optimized by combining the adjusted environmental information.

[0014] Furthermore, attention weights are allocated based on individual pedestrian characteristics and traffic environment factors to obtain decision-related environmental information, including: Acquire a set of environmental stimuli related to pedestrians, and calculate the attention weight corresponding to each environmental stimulus based on the location and velocity information of the environmental stimuli; By prioritizing the set of environmental stimuli based on attention weights, decision-related environmental information can be obtained.

[0015] Furthermore, the attention weights corresponding to each environmental stimulus are calculated, including: Correlation indices are calculated based on the distance between environmental stimuli and the pedestrian's current behavioral goal; The urgency index is calculated based on the relative displacement and relative speed of environmental stimuli and pedestrians. Attention weights are obtained by fusing relevance and urgency indicators based on weighting coefficients.

[0016] Furthermore, determining the pedestrian's motion state and obtaining the pedestrian's motion decision at the next moment includes: The target driving force of pedestrians is determined based on individual pedestrian attributes, and the force components of pedestrians are calculated based on decision-related environmental information. The resultant force of the pedestrian is obtained by superimposing the target driving force and the force components, and the position and velocity of the pedestrian are updated based on the resultant force to obtain the pedestrian's motion decision at the next moment.

[0017] Furthermore, the target driving force and force components of the pedestrian are determined, including: Expected speed and relaxation time are calculated based on pedestrian patience index, and target driving force is determined accordingly; The interpersonal interaction force is calculated based on the relative displacement of the pedestrian to other pedestrians, and the force exerted by the obstacle is calculated based on the distance and direction between the pedestrian and the obstacle.

[0018] Furthermore, the pedestrian's movement decision at the next moment is obtained, including: The pedestrian's acceleration is calculated based on the resultant force, and the pedestrian's state parameters are updated in conjunction with the time step. The short-term benefit function is calculated based on the current region's congestion level and the updated state parameters. The decision on whether to trigger a re-invocation of global path planning is then made based on the short-term benefit function and the judgment result of preset conditions.

[0019] A pedestrian behavior simulation system based on intent and attention mechanisms includes: The long-term intent module is used to perform global path planning based on pedestrian attributes and traffic environment information to obtain temporary target points for pedestrians. The attention mechanism module is used to allocate attention weights based on individual pedestrian characteristics and traffic environment elements to obtain decision-related environmental information. The short-term intention module is used to determine the pedestrian's motion state based on the pedestrian's temporary target point and the decision-related environmental information, and to obtain the pedestrian's motion decision at the next moment.

[0020] The present invention has achieved the following beneficial effects.

[0021] 1. This invention introduces long-term intention, short-term intention, and attention mechanism within a unified framework, thereby achieving holistic modeling of pedestrian behavior decision-making with value function as the core. This makes up for the shortcomings of traditional models in characterizing socio-psychological factors and significantly improves the consistency and interpretability of pedestrian behavior simulation.

[0022] 2. This invention utilizes an attention mechanism to filter and assign weights to multi-source information in a scene. Combined with an improved social force model that considers changes in intent, it can accurately characterize the interactive behavior of pedestrians in complex dynamic environments, making the generated motion trajectory closer to the real pedestrian behavior characteristics.

[0023] 3. This invention introduces an influence disk mechanism and a bottleneck detection mechanism. By dynamically adjusting individual space requirements and congestion assessment, it effectively simulates the unique social characteristics and group behavior of pedestrians in high-density scenarios, greatly improving the rationality and diversity of pedestrian simulation behavior in complex traffic scenarios.

[0024] 4. The method described in this invention can respond to dynamic changes in the environment in real time and continuously optimize the path based on the short-term benefit function, providing simulation support with high semantic realism for the evaluation of autonomous driving systems and robot navigation planning. Attached Figure Description

[0025] Figure 1This is a flowchart of the pedestrian behavior simulation method based on intent and attention mechanisms in the embodiments; Figure 2 This is a framework diagram of the pedestrian behavior simulation method based on intent and attention mechanisms in the embodiment; Figure 3 This is a block diagram of the pedestrian behavior simulation system based on intent and attention mechanisms in the embodiment; Figure 4 This is a scatter plot comparing pedestrian flow-density from simulation experiments with real data. Figure 5 This is a data graph showing different pedestrian densities in the experiment. Detailed Implementation

[0026] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not limit it in any way.

[0027] This invention provides a pedestrian behavior simulation method based on intent and attention mechanisms, such as... Figure 1 and Figure 2 As shown, it includes the following steps: Step S1: Perform global path planning based on pedestrian attributes and traffic environment information to obtain temporary target points for pedestrians.

[0028] The system obtains the long-term behavioral goals of pedestrians from their attributes and generates at least one temporary target point based on the perception and understanding of the traffic environment using a global path planning method.

[0029] In an optional embodiment of the present invention, based on AnytimeD The algorithm performs path planning for pedestrians' destination; AnytimeD The algorithm is a graph search-based path planning algorithm that can continuously update and optimize the planned path from the starting position to the target position under dynamic environmental changes. Specifically, it constructs an environmental graph representing the traffic environment, where each node represents a reachable location in the environment, and the edges between nodes represent feasible paths for a pedestrian to move from one location to another. The algorithm uses the pedestrian's current position as the starting node to search and plan the path to the target position.

[0030] Step S11: Initialize the node values ​​in the environment graph.

[0031] During the initialization phase, the cost of each node in the environment graph is initialized to infinity, and the cost of the starting node is set to zero.

[0032] The starting node represents the pedestrian's current position.

[0033] Step S12: Select the node with the lowest cost to expand the path and obtain the path information.

[0034] During the path search process, a priority queue is used to select the node with the lowest current generation value from the unprocessed nodes for expansion, and the generation value of its adjacent nodes and the corresponding path information are updated accordingly.

[0035] Step S13: Assess the impact of environmental changes on node generation value and update path information.

[0036] When the environmental conditions change, the cost of the affected nodes is reassessed, and the existing paths are updated accordingly.

[0037] In an optional embodiment of the present invention, changes in the environmental state include, but are not limited to, changes in pedestrian distribution or the appearance of new obstacles.

[0038] Step S14: Optimize the path based on the latest environmental information to obtain a dynamically adaptive path planning result.

[0039] As pedestrians continue to move, the generated paths are continuously optimized based on the latest environmental information to ensure that the path planning results can adapt to dynamic changes in the environment.

[0040] In an optional embodiment of the present invention, each pedestrian corresponds to an influence disk, the size and shape of which are dynamically adjusted according to the pedestrian's personal space needs and current psychological state. When the influence disks of multiple pedestrians intersect, the pedestrians adjust their positions according to their respective priorities and goals, thereby managing queuing and waiting behavior.

[0041] In other embodiments of the present invention, global path planning may also be implemented using other path planning algorithms.

[0042] Step S2: Attention weights are assigned based on individual pedestrian characteristics and traffic environment factors to obtain decision-related environmental information.

[0043] In an optional embodiment of the present invention, a dynamic weight allocation mechanism is constructed to characterize pedestrians. The process of attention allocation for different stimuli in a traffic environment; the stimuli include, but are not limited to, other pedestrians, vehicles, obstacles and traffic signs; each stimulus is assigned a corresponding attention weight according to its relevance to the current behavioral goal and its urgency, and the attention weight is used to characterize the priority of each stimulus at the current moment.

[0044] Step S21: Obtain the set of environmental stimuli from the environmental perception unit and represent each environmental stimulus using a state vector.

[0045] Acquired from environmental sensing units and pedestrians The relevant set of environmental stimuli; for any environmental stimulus Using state vectors It is represented, among which It includes information on the location, velocity, and type of environmental stimuli.

[0046] In an optional embodiment of the present invention, the environmental stimuli include, but are not limited to, other pedestrians, vehicles, obstacles, and traffic signs.

[0047] Step S22: Calculate the correlation index and urgency index for each environmental stimulus.

[0048] For each environmental stimulus Calculate the relationship between them and the pedestrians respectively. Correlation indicators between current behavioral goals and urgency indicators : in, Indicates environmental stimuli Location; pedestrian Location; Indicates environmental stimuli The velocity vector; pedestrian The velocity vector; For scale parameters; It is a very small constant.

[0049] Step S23: Calculate attention weights based on relevance and urgency indicators.

[0050] Calculate the corresponding attention weight for each environmental stimulus. : in, and These are the weighting coefficients used to balance the effects of relevance and urgency; Indicates environmental stimuli The corresponding attention weights.

[0051] Step S24: Filter and prioritize environmental information based on attention weights.

[0052] For stimuli with high attention weight, retain their complete state information; for stimuli with attention weight below a preset threshold, reduce their processing priority.

[0053] Step S3: Determine the pedestrian's motion state based on the pedestrian's temporary target point and the decision-related environmental information, and obtain the pedestrian's motion decision at the next moment.

[0054] In an optional embodiment of the present invention, the collision avoidance strategy is optimized by comprehensively considering the pedestrian's emotional state and intention changes during the interaction process to improve the realism of the pedestrian's interaction behavior; a bottleneck detection mechanism is introduced to guide pedestrians to re-evaluate and adjust their decision path when local congestion is detected; and an influence disk mechanism is introduced to characterize the queuing and waiting behavior of pedestrians in high-density scenes.

[0055] Step S31, obtain pedestrians The patience index determines the desired speed and relaxation time.

[0056] At the start of each simulation moment, pedestrian data is acquired. Patience Index , The patience index is used to characterize the psychology of pedestrians in congested or waiting scenarios; the patience index is used to calculate pedestrian... Expected speed and relaxation time : in, and These represent the upper and lower limits of the pedestrian's expected speed, respectively; and These represent the maximum and minimum relaxation times, respectively.

[0057] Step S32: Based on the pedestrian's current position and the corresponding temporary target point, obtain the desired direction of movement and the target driving force.

[0058] Based on pedestrians Current location Its corresponding temporary target point Calculate the desired direction of motion and the target driving force: in, This represents the pedestrian's current velocity vector; the target driving force is used to guide the pedestrian toward the temporary target point.

[0059] Step S33: Traverse the set of pedestrians and calculate the relative displacement vector to obtain the interpersonal interaction force.

[0060] For pedestrians traversing other pedestrians Calculate its relative displacement vector: in, This represents the relative displacement vector.

[0061] Calculate the pedestrian interaction force between pedestrians based on the magnitude of the relative displacement vector: Where A and B are the interaction strength and range parameters.

[0062] pedestrian The total interpersonal interaction force received is: in, pedestrian The total interpersonal interaction force received.

[0063] Step S34: Estimate the local crowd density based on the distance information between pedestrians to obtain the congestion index.

[0064] While calculating interpersonal interaction forces, local crowd density is estimated based on distance information between pedestrians: in, Indicates local population density.

[0065] And calculate the congestion index for the current scenario: in, The congestion level represents the total number of pedestrians in the scene; the congestion level is used to characterize the current traffic status of the area.

[0066] Step S35: Based on the relative positional relationship between the pedestrian and the obstacle, obtain the force exerted by the obstacle.

[0067] Calculate the force exerted by the obstacle based on the relative positional relationship between the pedestrian and the obstacle in the environment: in, pedestrian and obstacles The shortest distance; Let be the unit normal vector pointing from the obstacle to the pedestrian.

[0068] Step S36: Superimpose the component forces to obtain the resultant force and acceleration, and update the pedestrian's motion state.

[0069] By superimposing the driving force of the goal, the force of interpersonal interaction, and the force of the obstacle, we obtain the force acting on the pedestrian. The resultant force and acceleration: in, This represents the quality parameters of pedestrians.

[0070] Based on time step Update pedestrian speed and location: in, This represents the speed parameter of pedestrians. This represents the location parameters of the pedestrian.

[0071] Step S37: Based on the congestion level and pedestrian movement status, obtain the short-term benefit function.

[0072] After completing the state update, a short-term benefit function is calculated based on the congestion level, the pedestrian's current position, and speed. This function is used to evaluate the rationality of the current movement decision and serves as the basis for triggering a replanning process. in, pedestrian Short-term benefits at the present moment.

[0073] This invention also provides a pedestrian behavior simulation system based on intent and attention mechanisms, such as... Figure 3 As shown, it includes: The long-term intent module is used to perform global path planning based on pedestrian attributes and traffic environment information to obtain temporary target points for pedestrians. The attention mechanism module is used to allocate attention weights based on individual pedestrian characteristics and traffic environment elements to obtain decision-related environmental information. The short-term intention module is used to determine the pedestrian's movement state based on the pedestrian's temporary target point and the decision-related environmental information, and to obtain the pedestrian's movement decision at the next moment.

[0074] Method performance testing: To verify the effectiveness of the pedestrian behavior modeling method described in this invention, a joint simulation platform was built based on CARLA and SUMO, and two real datasets widely used in the field of pedestrian behavior modeling, namely the EWAP dataset and the UCY dataset, were selected as reference benchmarks for model performance evaluation.

[0075] In terms of model configuration, the long-term intent module uses Anytime D. In the path planning algorithm experiment, the search parameters were set as follows: Set to 20, the heuristic function is defined as the Euclidean straight-line distance between the pedestrian's current position and the target position.

[0076] like Figure 4 As shown, by comparing the density-flow curve obtained from the simulation results with the corresponding curve obtained from the statistical analysis of the real dataset, it can be seen that the model described in this invention can accurately reproduce the actual observed pattern of pedestrian flow changing with density. The two are highly consistent in the overall trend and key intervals, thus verifying the effectiveness of the modeling method at the level of macroscopic statistical characteristics.

[0077] like Figure 5 As shown, the environmental scenario was constructed with reference to typical scenarios in the UCY dataset, creating a simulation environment for pedestrians crossing a two-way zebra crossing. Traffic flow results showing different pedestrian densities over time were generated using the simulation platform. The model's visual rationality and potential performance in real-world applications were qualitatively evaluated to further verify its applicability in complex interactive scenarios.

[0078] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A pedestrian behavior simulation method based on intent and attention mechanisms, characterized in that, Includes the following steps: Global path planning is performed based on pedestrian attributes and traffic environment information to obtain temporary target points for pedestrians; Attention weights are allocated based on individual pedestrian characteristics and traffic environment factors to obtain decision-related environmental information. Based on the pedestrian's temporary target point and the decision-related environmental information, the pedestrian's movement state is determined, and the pedestrian's movement decision for the next moment is obtained.

2. The method as described in claim 1, characterized in that, Global path planning is performed based on pedestrian attributes and traffic environment information to obtain temporary pedestrian target points, including: Construct a traffic environment graph that includes reachable location nodes and their connections; Starting from the pedestrian's current location, search for a path to the long-term behavioral goal in the traffic environment map; The node values ​​are adjusted in real time according to changes in the traffic environment, and the path is updated to obtain the temporary target point for pedestrians.

3. The method as described in claim 2, characterized in that, Constructing a traffic environment map includes: Obtain information on obstacle distribution and passable areas in the traffic environment; The passable area is divided into multiple nodes representing locations, and edges are established based on the connectivity between the nodes.

4. The method as described in claim 2, characterized in that, Update routes based on changes in traffic conditions, including: The dynamic impact zone of pedestrians is determined based on their spatial needs and psychological state. When the influence disks of multiple pedestrians intersect, their positions are adjusted according to priority, and the path planning results are optimized by combining the adjusted environmental information.

5. The method as described in claim 1, characterized in that, Attention weights are assigned based on individual pedestrian characteristics and traffic environment factors to obtain decision-related environmental information, including: Acquire a set of environmental stimuli related to pedestrians, and calculate the attention weight corresponding to each environmental stimulus based on the location and velocity information of the environmental stimuli; By prioritizing the set of environmental stimuli based on attention weights, decision-related environmental information can be obtained.

6. The method as described in claim 5, characterized in that, Calculate the attention weight corresponding to each environmental stimulus, including: Correlation indices are calculated based on the distance between environmental stimuli and the pedestrian's current behavioral goal; The urgency index is calculated based on the relative displacement and relative speed of environmental stimuli and pedestrians. Attention weights are obtained by fusing relevance and urgency indicators based on weighting coefficients.

7. The method as described in claim 1, characterized in that, Determine the pedestrian's motion state and obtain the pedestrian's motion decision for the next moment, including: The target driving force of pedestrians is determined based on individual pedestrian attributes, and the force components of pedestrians are calculated based on decision-related environmental information. The resultant force of the pedestrian is obtained by superimposing the target driving force and the force components, and the position and velocity of the pedestrian are updated based on the resultant force to obtain the pedestrian's motion decision at the next moment.

8. The method as described in claim 7, characterized in that, Determine the target driving force and force components of the pedestrian, including: Expected speed and relaxation time are calculated based on pedestrian patience index, and target driving force is determined accordingly; The interpersonal interaction force is calculated based on the relative displacement of the pedestrian to other pedestrians, and the force exerted by the obstacle is calculated based on the distance and direction between the pedestrian and the obstacle.

9. The method as described in claim 7, characterized in that, To obtain the pedestrian's movement decision in the next moment, including: The pedestrian's acceleration is calculated based on the resultant force, and the pedestrian's state parameters are updated in conjunction with the time step. The short-term benefit function is calculated based on the current region's congestion level and the updated state parameters. The decision on whether to trigger a re-invocation of global path planning is then made based on the short-term benefit function and the judgment result of preset conditions.

10. A pedestrian behavior simulation system based on intent and attention mechanisms, characterized in that, include: The long-term intent module is used to perform global path planning based on pedestrian attributes and traffic environment information to obtain temporary target points for pedestrians. The attention mechanism module is used to allocate attention weights based on individual pedestrian characteristics and traffic environment elements to obtain decision-related environmental information. The short-term intention module is used to determine the pedestrian's movement state based on the pedestrian's temporary target point and the decision-related environmental information, and to obtain the pedestrian's movement decision at the next moment.

Images