Pedestrian behavior simulation scene construction method and system, terminal and storage medium

Abstract

The invention discloses a pedestrian behavior simulation scene construction method and system, a terminal and a storage medium, and relates to the technical field of crowd simulation, and the method comprises the steps: obtaining a static three-dimensional virtual scene, and determining simulation environment data according to the static three-dimensional virtual scene, the simulation environment data comprising a two-dimensional structured navigation map and at least one global reference trajectory; obtaining state data corresponding to the virtual pedestrian, and obtaining a predicted motion trajectory corresponding to the virtual pedestrian through a trained trajectory diffusion generation model according to the state data and the simulation environment data; and controlling the virtual pedestrian to move in the static three-dimensional virtual scene according to the predicted motion track. Therefore, the pedestrian motion trail closer to the real scene can be generated based on the pre-trained trail diffusion generation model, so that the motion trail of the virtual pedestrian is closer to the pedestrian motion trail in the real scene, the authenticity of pedestrian behaviors can be improved, and the authenticity of the constructed pedestrian behavior simulation scene can be improved.

Description

Technical Field

[0001] This application relates to the field of crowd simulation technology, and in particular to a method, system, terminal and storage medium for constructing pedestrian behavior simulation scenarios. Background Technology

[0002] With the development of science and technology, the application of human-machine collaboration is becoming increasingly widespread. In human-machine collaborative environments, autonomous robot navigation needs to balance safety and natural human-machine interaction, and simulation of complex and diverse navigation scenarios is key to training reliable navigation strategies and evaluating algorithm performance.

[0003] In constructing pedestrian behavior simulation scenarios, existing technologies typically build a 3D virtual scene and then generate pedestrian behavior within that virtual scene according to predefined fixed rules (such as collision avoidance and following crowds). The problem with existing technologies is that pedestrian behavior generated based on predefined fixed rules is detrimental to improving the realism of pedestrian behavior, and consequently, to improving the realism of the pedestrian behavior simulation scenario.

[0004] Therefore, the relevant technologies still need to be improved and developed. Summary of the Invention

[0005] The main purpose of this application is to provide a method, system, terminal and storage medium for constructing pedestrian behavior simulation scenarios, aiming to solve the technical problem in related technologies where pedestrian behavior is generated based on predefined fixed rules, which is not conducive to improving the realism of pedestrian behavior, and thus not conducive to improving the realism of pedestrian behavior simulation scenarios.

[0006] To achieve the above objectives, the first aspect of this application provides a method for constructing a pedestrian behavior simulation scene, wherein the method includes: Obtain a static 3D virtual scene, and determine simulation environment data based on the static 3D virtual scene. The simulation environment data includes a 2D structured navigation map and at least one global reference trajectory. Obtain the state data corresponding to the virtual pedestrian. Based on the state data and the simulation environment data, obtain the predicted motion trajectory corresponding to the virtual pedestrian through the trained trajectory diffusion generation model. Based on the predicted motion trajectory, the virtual pedestrian is controlled to move within the static 3D virtual scene to obtain a pedestrian behavior simulation scene.

[0007] Optionally, the above-mentioned acquisition of a static 3D virtual scene and determination of simulation environment data based on the static 3D virtual scene include: Import a generic scene description format file containing geometric and semantic information to construct the aforementioned static 3D virtual scene; Spatial analysis is performed on the above static three-dimensional virtual scene to generate a two-dimensional occupied grid map representing the walkable area as the above two-dimensional structured navigation map. Based on the aforementioned two-dimensional occupied grid map, the global reference trajectory corresponding to the virtual pedestrian is determined through a path planning algorithm.

[0008] Optionally, the above-mentioned acquisition of the state data corresponding to the virtual pedestrian, and the acquisition of the predicted motion trajectory corresponding to the virtual pedestrian based on the state data and the simulation environment data through a trained trajectory diffusion generation model, including: The aforementioned state data and simulation environment data are input into the trained data-driven trajectory diffusion generation model. Obtain the predicted motion trajectory output by the data-driven trajectory diffusion generation model, wherein the predicted motion trajectory is the future trajectory of the virtual pedestrian predicted by the trajectory diffusion generation model.

[0009] Optionally, the status data corresponding to the aforementioned virtual pedestrian includes: The current status of the aforementioned virtual pedestrian, the historical trajectory of the aforementioned virtual pedestrian, and the trajectory of neighboring pedestrians; The predicted trajectory includes the coordinates of the two-dimensional trajectory points of the virtual pedestrian at multiple future moments, as well as the heading angle and velocity corresponding to each of the two-dimensional trajectory points.

[0010] Optionally, controlling the movement of the virtual pedestrian in the static three-dimensional virtual scene based on the predicted motion trajectory includes: Based on the predicted motion trajectory and the state data corresponding to the virtual pedestrian, a physical control signal that drives all joints of the virtual pedestrian is generated through a reinforcement learning controller trained on adversarial motion priors. The virtual pedestrian is controlled to move within the static three-dimensional virtual scene based on the aforementioned physical control signals.

[0011] Optionally, the aforementioned virtual pedestrian is a pedestrian agent configured based on a skinned multi-person linear model; The aforementioned pedestrian intelligent agent has a hinge body capable of driving joints; The above-mentioned control of the virtual pedestrian's movement in the above-mentioned static three-dimensional virtual scene based on the above-mentioned physical control signals includes: Based on the aforementioned physical control signals, the articulation of the aforementioned pedestrian intelligent agent is controlled so that the aforementioned virtual pedestrian moves along the aforementioned predicted motion trajectory in the aforementioned static three-dimensional virtual scene.

[0012] Optionally, the above method further includes: By setting up a robot in the above-mentioned static three-dimensional virtual scene, scene data corresponding to the above-mentioned static three-dimensional virtual scene and action change sequence data corresponding to the above-mentioned virtual pedestrian in the above-mentioned static three-dimensional virtual scene are collected. Based on the above scenario data and the above action change sequence data, navigation scenario data is constructed.

[0013] A second aspect of this application provides a pedestrian behavior simulation scene construction system, wherein the system includes: The data processing module is used to acquire a static three-dimensional virtual scene and determine simulation environment data based on the static three-dimensional virtual scene. The simulation environment data includes a two-dimensional structured navigation map and at least one global reference trajectory. The trajectory prediction module is used to obtain the state data corresponding to the virtual pedestrian. Based on the state data and the simulation environment data, the module obtains the predicted motion trajectory of the virtual pedestrian through the trained trajectory diffusion generation model. The motion control module is used to control the movement of the virtual pedestrian in the static three-dimensional virtual scene according to the predicted motion trajectory, so as to obtain a pedestrian behavior simulation scene.

[0014] A third aspect of this application provides a terminal, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements any of the steps of the pedestrian behavior simulation scene construction method described above.

[0015] A fourth aspect of this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the steps of the pedestrian behavior simulation scene construction method described above.

[0016] As can be seen from the above, in this application, a static three-dimensional virtual scene is obtained, and simulation environment data is determined based on the static three-dimensional virtual scene. The simulation environment data includes a two-dimensional structured navigation map and at least one global reference trajectory. The state data corresponding to the virtual pedestrian is obtained, and based on the state data and the simulation environment data, a predicted motion trajectory corresponding to the virtual pedestrian is obtained through a trained trajectory diffusion generation model. Based on the predicted motion trajectory, the virtual pedestrian is controlled to move in the static three-dimensional virtual scene.

[0017] Therefore, in the process of constructing a pedestrian behavior simulation scenario, pedestrian behavior is not simply generated according to predefined fixed rules. Instead, simulation environment data, including a two-dimensional structured navigation map and at least one global reference trajectory, is first determined based on the obtained static 3D virtual scene. Then, based on the simulation environment data and the virtual pedestrian's state data, a pre-trained trajectory diffusion generation model is used to predict the virtual pedestrian's predicted motion trajectory, thereby controlling the virtual pedestrian's movement based on the predicted trajectory. In this way, a pedestrian motion trajectory closer to the real scene can be generated based on the pre-trained trajectory diffusion generation model, making the virtual pedestrian's motion trajectory closer to the real-world pedestrian motion trajectory, which helps improve the realism of pedestrian behavior and thus improves the realism of the constructed pedestrian behavior simulation scenario. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a method for constructing a pedestrian behavior simulation scene according to an embodiment of this application; Figure 2 This is a schematic diagram of the technical route of a pedestrian behavior simulation scene construction method provided in an embodiment of this application; Figure 3 This is a schematic diagram of a whole-body pedestrian intelligent agent control algorithm framework based on adversarial motion prior provided in an embodiment of this application; Figure 4 This is a schematic diagram illustrating the visualization effect of a simulation platform provided in an embodiment of this application; Figure 5 This is a schematic diagram illustrating the training effect of an adversarial motion prior algorithm for full-body control of a virtual pedestrian provided in an embodiment of this application; Figure 6 This is a comparative diagram showing the training status of a robot navigation algorithm provided in an embodiment of this application; Figure 7 This is a schematic diagram of the constituent modules of a pedestrian behavior simulation scene construction system provided in an embodiment of this application; Figure 8 This is a block diagram illustrating the internal structure of a terminal provided in an embodiment of this application. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0021] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0022] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0023] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0024] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to classification." Similarly, the phrases "if determined" or "if classified to [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once classified to [the described condition or event]," or "in response to classification to [the described condition or event]."

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0026] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0027] In human-machine collaborative environments, the reliability and adaptability of robot autonomous navigation systems need to be verified through testing in a large number of complex scenarios, and the realism and efficiency of pedestrian behavior simulation scenarios determine the quality and efficiency of navigation algorithm training.

[0028] In some technologies, a 3D virtual scene is typically constructed, and then pedestrian behavior is generated within the 3D virtual scene according to predefined fixed rules (such as collision avoidance and following the crowd). The problem with existing technologies is that pedestrian behavior is generated based on predefined fixed rules, which is not conducive to improving the realism of pedestrian behavior, and thus not conducive to improving the realism of pedestrian behavior simulation scenes.

[0029] Current research on pedestrian behavior simulation and robot navigation in dense crowds faces several key bottlenecks. For example, the simulations suffer from insufficient realism and computational scalability: existing platforms (such as gym-avoid and SocNavBench) mostly approximate pedestrian movement using 2D geometric trajectories, lacking dynamic interaction simulations based on physics engines, and the rendering effects differ significantly from the visual feedback of the real environment. Furthermore, traditional simulation platforms (such as HuNavSim based on Gazebo and HabiCrowd based on Habitat) do not support parallel simulation using graphics processing units (GPUs), making it impossible to simultaneously simulate a large number of intelligent agents (pedestrians + robots), thus failing to meet the demand for efficient data generation in dense crowd scenarios. The fidelity of pedestrian behavior is also low: most simulation platforms (such as SEAN 2.0 and HuNavSim) rely on manually designed rules (such as social force models) or pre-recorded animations to generate pedestrian behavior, failing to simulate multimodal interactions in dense crowds (such as gait changes, obstacle avoidance intentions, and group collaboration), and exhibiting discontinuities in motion transitions (such as sudden changes in gait speed).

[0030] In terms of scene construction, existing simulation platforms mostly use simple 2D geometric scenes or low-precision 3D models, lacking realistic scenes containing complete geometric structures and semantic information. This results in significant differences between the generated navigation maps and the actual environment, and navigation algorithms trained on such scenes have extremely poor generalization ability in real-world environments. At the same time, global path planning often relies on manual settings and cannot adaptively generate feasible paths according to the scene structure, making it difficult to cover diverse navigation needs.

[0031] Pedestrian group behavior simulation is crucial for the development and evaluation of simulation platforms. In some applications, the Social Force Model (SFM) is used to simulate pedestrian behavior based on physical forces, generating trajectories by calculating repulsive forces between pedestrians and attractive forces from targets. However, the SFM model has significant drawbacks, such as difficulty capturing complex pedestrian interactions and frequent abrupt changes and discontinuities in generated trajectories (e.g., sudden turns, speed changes). In other applications, rule-based models are used, employing predefined rules (e.g., collision avoidance, following the crowd) to generate pedestrian behavior. These methods are computationally efficient but lack diverse representations of real-world pedestrian behavior, resulting in overly simplistic scenarios that fail to reflect complex interactions between real pedestrians. In still other applications, data-driven generative models, such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), can be employed. While these methods can generate diverse trajectories, they typically lack a physical basis, leading to significant deficiencies in the continuity and physical plausibility of the generated pedestrian behavior.

[0032] In the simulation application stage, existing platforms generally lack efficient data acquisition and utilization mechanisms, and cannot simultaneously acquire simulation scene data and pedestrian behavior data to build a comprehensive navigation scenario dataset. Moreover, most of them do not support GPU parallel computing, resulting in low efficiency in large-scale crowd simulations, which seriously restricts the development cycle of robot navigation algorithms.

[0033] To address at least one of the aforementioned technical problems, this application proposes a method for constructing a pedestrian behavior simulation scene. The method involves: acquiring a static three-dimensional virtual scene; determining simulation environment data based on the static three-dimensional virtual scene, wherein the simulation environment data includes a two-dimensional structured navigation map and at least one global reference trajectory; acquiring state data corresponding to a virtual pedestrian; obtaining a predicted motion trajectory for the virtual pedestrian based on the state data and the simulation environment data using a trained trajectory diffusion generation model; and controlling the virtual pedestrian's movement within the static three-dimensional virtual scene based on the predicted motion trajectory.

[0034] Therefore, in the process of constructing a pedestrian behavior simulation scenario, pedestrian behavior is not simply generated according to predefined fixed rules. Instead, simulation environment data, including a two-dimensional structured navigation map and at least one global reference trajectory, is first determined based on the obtained static 3D virtual scene. Then, based on the simulation environment data and the virtual pedestrian's state data, a pre-trained trajectory diffusion generation model is used to predict the virtual pedestrian's predicted motion trajectory, thereby controlling the virtual pedestrian's movement based on the predicted trajectory. In this way, a pedestrian motion trajectory closer to the real scene can be generated based on the pre-trained trajectory diffusion generation model, making the virtual pedestrian's motion trajectory closer to the real-world pedestrian motion trajectory, which helps improve the realism of pedestrian behavior and thus improves the realism of the constructed pedestrian behavior simulation scenario.

[0035] like Figure 1 As shown in the figure, this application provides a method for constructing a pedestrian behavior simulation scene. Specifically, the method includes the following steps: Step S100: Obtain a static three-dimensional virtual scene, and determine simulation environment data based on the static three-dimensional virtual scene. The simulation environment data includes a two-dimensional structured navigation map and at least one global reference trajectory. Step S200: Obtain the state data corresponding to the virtual pedestrian; based on the state data and the simulation environment data, obtain the predicted motion trajectory corresponding to the virtual pedestrian through the trained trajectory diffusion generation model. Step S300: Based on the predicted motion trajectory, control the virtual pedestrian to move in the static three-dimensional virtual scene to obtain a pedestrian behavior simulation scene.

[0036] It should be noted that the pedestrian behavior simulation scene construction method provided in this application embodiment can be implemented based on a simulation platform. This simulation platform can serve as an external data generation and management interface for a physics engine simulated on a GPU, aiming to support large-scale, high-fidelity simulation of human-computer interaction navigation scenarios. Relying on the core physical simulation functions and realistic rendering of the physics engine, this application embodiment provides a standardized framework that supports robot learning in a large-scale parallel environment to generate navigation scenarios with natural human activities.

[0037] Specifically, the simulation platform includes the following functions: Scene Construction: Constructing static 3D scenes and defining global navigation paths; Pedestrian Generation: Synthesizing natural human behavior using data-driven prior knowledge; Parallel Simulation: Synchronizing robots, pedestrians, and the environment through a physics engine to support high-throughput simulation generation. Specifically, it imports pre-made 3D static scene assets and obtains their 2D planar maps using ray scanning technology for navigation planning of virtual humans and robots; a global planner calculates a series of discrete path points. The virtual human agent is trained in these scenes to track the globally planned path points, mimicking the actions of real pedestrian datasets to control the virtual human to track assigned trajectories in various styles. Finally, the robot, equipped with sensors, is imported into the simulation engine's physical environment along with the virtual pedestrians. Parallel management of these agents is achieved through a vector environment, and the simulator's real-time feedback serves as input to the robot's navigation strategy and the pedestrian action generation network.

[0038] The aforementioned static 3D virtual scene can be pre-built through other platforms or built in real time; no specific limitations are made here.

[0039] In this embodiment of the application, the above-mentioned acquisition of a static three-dimensional virtual scene and determination of simulation environment data based on the static three-dimensional virtual scene include: Import a generic scene description format file containing geometric and semantic information to construct the aforementioned static 3D virtual scene; Spatial analysis is performed on the above static three-dimensional virtual scene to generate a two-dimensional occupied grid map representing the walkable area as the above two-dimensional structured navigation map. Based on the aforementioned two-dimensional occupied grid map, the global reference trajectory corresponding to the virtual pedestrian is determined through a path planning algorithm.

[0040] In some application scenarios, 3D scene assets in Universal Scene Description (USD) format are imported. These assets contain the geometry of the environment, semantic information (such as classification labels for each region), and visual textures (such as the textures of objects). These assets are then loaded through Isaac Lab's scene graph management module to construct a static 3D virtual scene.

[0041] A ray scanning technique is used to perform spatial analysis on a static 3D virtual scene, traversing the positional information of all physical entities in the scene and mapping the 3D space into a 2D grid. For example, a grid value of 1 indicates that the area is occupied by obstacles (such as shelves or walls), while a grid value of 0 indicates a walkable area (such as a passageway). A 2D occupancy grid map is generated as a 2D structured navigation map. The grid precision can be set and adjusted according to actual needs to ensure the accuracy of navigation path planning.

[0042] Based on a two-dimensional occupied grid map, the A* path planning algorithm is used to set the starting point and target point of the virtual pedestrian. The algorithm generates at least one global reference trajectory composed of discrete path points by calculating the minimum cost path from the starting point to the target point (the cost includes the path length and the distance to obstacles). Each path point contains (x, y) coordinate information.

[0043] The diversity and realism of simulated scenarios are crucial for evaluating and improving the generalization ability of navigation strategies. In this embodiment, a visually and physically believable navigation space is constructed using a photorealistic 3D environment that includes geometric structure, semantic context, and visual textures. All scene elements are defined in USD format and managed by Isaac Lab's scene graph, where attributes such as mesh geometry and material textures are explicitly defined to achieve realistic rendering and accurate physical interaction. A scene consists of physically defined building structures (e.g., walls, floors), semantically defined functional areas (e.g., kitchen, living room, hallway), and household items that conform to the environment, approximating real-world deployment conditions. For task execution, NavIsaacLab generates an occupancy grid map based on the static scene layout to define the walkable space, thereby supporting effective sampling of the robot's initial state and navigation target. The global reference trajectory is calculated using the A* planner to ensure task solvability and support long-term and multi-agent navigation experiments requiring multiple feasible and collision-free paths.

[0044] It should be noted that other path planning algorithms can also be used to determine the global reference trajectory corresponding to the virtual pedestrian, and no specific limitations are made here.

[0045] Specifically, the process involves acquiring the state data corresponding to the virtual pedestrian, and based on this state data and the simulation environment data, using a trained trajectory diffusion generation model to obtain the predicted motion trajectory of the virtual pedestrian, including: The aforementioned state data and simulation environment data are input into the trained data-driven trajectory diffusion generation model. Obtain the predicted motion trajectory output by the data-driven trajectory diffusion generation model, wherein the predicted motion trajectory is the future trajectory of the virtual pedestrian predicted by the trajectory diffusion generation model.

[0046] The status data corresponding to the aforementioned virtual pedestrians includes: The current status of the aforementioned virtual pedestrian, the historical trajectory of the aforementioned virtual pedestrian, and the trajectory of neighboring pedestrians; The predicted trajectory includes the coordinates of the two-dimensional trajectory points of the virtual pedestrian at multiple future moments, as well as the heading angle and velocity corresponding to each of the two-dimensional trajectory points.

[0047] Furthermore, controlling the movement of the virtual pedestrian within the static three-dimensional virtual scene based on the predicted motion trajectory includes: Based on the predicted motion trajectory and the state data corresponding to the virtual pedestrian, a physical control signal that drives all joints of the virtual pedestrian is generated through a reinforcement learning controller trained on adversarial motion priors. The virtual pedestrian is controlled to move within the static three-dimensional virtual scene based on the aforementioned physical control signals.

[0048] Among them, the aforementioned virtual pedestrian is a pedestrian agent configured based on a skinned multi-person linear model; The aforementioned pedestrian intelligent agent has a hinge body capable of driving joints; The above-mentioned control of the virtual pedestrian's movement in the above-mentioned static three-dimensional virtual scene based on the above-mentioned physical control signals includes: Based on the aforementioned physical control signals, the articulation of the aforementioned pedestrian intelligent agent is controlled so that the aforementioned virtual pedestrian moves along the aforementioned predicted motion trajectory in the aforementioned static three-dimensional virtual scene.

[0049] In this embodiment, pedestrians are generated and their behavior is controlled within an acquired static 3D virtual scene. Specifically, a pedestrian generation module can be set up in the simulation platform, which uses a data-driven method to generate pedestrian behavior in dense crowds. Specifically, the pedestrian generation module may include a controllable pedestrian motion generation submodule and a trajectory diffusion generation submodule.

[0050] The virtual pedestrian has articulated joints that can reproduce natural whole-body kinematics. The aforementioned physical control signals can drive all joints of the virtual pedestrian to make the virtual pedestrian's movement conform to biomechanical constraints and track the corresponding predicted motion trajectory.

[0051] Each pedestrian agent is configured according to a Skinned Multi-Person Linear (SMPL) model, which represents the human body by combining learned body shape parameters with kinematic skeletons. To achieve full-body control of real-world pedestrian movement, in this embodiment, the skeleton of the SMPL model is mapped to 24 actuated joints rooted at the pelvis, covering major joints such as the hip, knee, and neck. The corresponding skinned mesh generated based on the SMPL shape parameters provides the surface geometry for photorealistic RGB-D rendering.

[0052] The controllable pedestrian motion generation submodule uses an Adversarial Motion Prior Controller (AMP) trained with adversarial motion priors. It trains a target-conditional policy through deep reinforcement learning (e.g., using proximal policy optimization algorithms), enabling the pedestrian to follow a 2D target trajectory and generate a realistic full-body posture. Its state space includes the pedestrian's current pose, obstacle features, and target trajectory; its motion space includes joint torques and angular velocities. Within the reinforcement learning framework, a motion discriminator is introduced to compare the actions generated by the current policy with real motion capture datasets (such as the AMASS dataset), calculating style rewards to ensure the generated motion possesses the realism of human movement while maintaining physical simulation. The total reward consists of AMP style rewards and task rewards. The task reward includes three parts: trajectory tracking error, calculated based on the distance difference between the agent's position and the 2D target trajectory; motion energy consumption, calculated based on joint torques and velocities, penalizing excessive energy consumption; and motion direction correction to prevent abnormal behaviors such as backward walking during training.

[0053] The trajectory diffusion generation submodule, based on the trajectory diffusion model, transforms pedestrian trajectory prediction into a conditional denoising process. Inputs include the pedestrian's historical trajectory (past 5-10 time steps), neighboring pedestrian trajectories, and a semantic map (e.g., obstacle locations, passage width). Output is a 2D trajectory for the next 10-20 time steps as the predicted motion trajectory. The predicted motion trajectory includes the position coordinates (x, y) and heading angle of each trajectory point. And velocity v. During the diffusion process, the forward process transforms the real trajectory into random noise by gradually adding Gaussian noise, while the reverse process learns to denoise through a 1D temporal convolutional network. At each step, the rationality of the trajectory is optimized by combining semantic map features (such as encoding obstacle information through 2D convolution).

[0054] In some application scenarios, virtual pedestrian state data is acquired first, and motion trajectory prediction is generated. Specifically, multiple (e.g., 10) virtual pedestrian agents are initialized. Each agent is configured based on the SMPL model and has 24 drivable joints (including lower limb joints such as hip, knee, and ankle joints, and upper limb joints such as neck and shoulder joints). The initial state data of each pedestrian is acquired through the simulation engine: current position, current heading angle (the angle with the positive x-axis), current speed, and historical trajectory of the past 5 time steps (containing 5 sets of state data). At the same time, the trajectory data of neighboring pedestrians within a 3m radius of each pedestrian is acquired through the sensor simulation module to form a set of neighboring pedestrian trajectories.

[0055] A trajectory diffusion generation model is constructed. The input layer of this model includes a state data channel (current state, historical trajectory, and neighbor trajectories) and an environmental data channel (feature encoding of the 2D occupied grid map and path point information of the global reference trajectory). The intermediate layer uses a 1D temporal convolutional network to extract temporal and spatial features. The output layer is the predicted motion trajectory for the next 15 time steps, and each trajectory point includes position coordinates (x, y) and heading angle. And velocity v.

[0056] The acquired state data and the generated simulation environment data are input into the trained trajectory diffusion generation model. The model generates the predicted motion trajectory of each virtual pedestrian through a reverse denoising process (a denoising strategy trained based on the real pedestrian trajectory dataset AMASS).

[0057] Then, motion control of the virtual pedestrian is implemented based on the predicted trajectory. A reinforcement learning controller based on adversarial motion prior (AMP) is constructed, which consists of a policy network and a motion discriminator. The policy network takes as input pathpoint information of the predicted motion trajectory and the real-time state of the virtual pedestrian (joint angles, joint velocities), and outputs torque and angular velocity control signals for 24 joints. The motion discriminator outputs a style reward value by comparing the joint motion data generated by the policy network with the real human motion data in the AMASS dataset.

[0058] The predicted motion trajectory and the real-time status data of the virtual pedestrian are input into the AMP controller. The policy network generates an initial joint control signal based on the trajectory tracking error (the distance between the current position and the trajectory point). The motion discriminator calculates the similarity between the motion style corresponding to the control signal and the real data, and outputs a style reward. Combining the trajectory tracking reward (the smaller the distance, the higher the reward) and the energy consumption penalty (the points are deducted when the joint torque is too large), the policy network optimizes the control signal and finally generates a joint control signal that conforms to the physical constraints.

[0059] The joint control signals are input into the articulated body model of the virtual pedestrian, controlling each joint to move at a specified angle and speed, so that the virtual pedestrian walks along the predicted motion trajectory.

[0060] Furthermore, the above methods also include: By setting up a robot in the above-mentioned static three-dimensional virtual scene, scene data corresponding to the above-mentioned static three-dimensional virtual scene and action change sequence data corresponding to the above-mentioned virtual pedestrian in the above-mentioned static three-dimensional virtual scene are collected. Based on the above scenario data and the above action change sequence data, navigation scenario data is constructed.

[0061] Specifically, in this embodiment, parallel simulation of the robot and virtual pedestrian can also be performed. After the simulation starts, the generated positions of the robot and virtual pedestrian are first randomly sampled and allocated according to the offline sampled task paths. As the number of generated agents increases, various navigation conditions will be formed in diverse scenarios. At this time, the motion control of the virtual pedestrian is in the inference state, no longer relying on external real pedestrian motion data, and will calculate the optimal action based on the tracking trajectory planned by the trajectory diffusion generation module and the real-time feedback of the simulation. Multimodal sensors are integrated on the robot body, such as built-in RGBD cameras, LiDAR and other virtual sensors, which can output point cloud, RGB image, depth map and other data consistent with real devices. Therefore, the robot can capture the motion change sequence of pedestrians during navigation. As a synthetic data acquisition method, it can quickly collect a large amount of cross-scene multimodal navigation scenario data to support the training of robot visual navigation algorithms.

[0062] In some application scenarios, at least one mobile robot equipped with an RGBD camera and LiDAR is deployed in a static 3D virtual scene, and the robot's initial position and navigation target are set. During the virtual pedestrian's movement, the robot acquires RGB and depth images of the scene through the RGBD camera, and collects point cloud data of the surrounding environment through the LiDAR. Simultaneously, it records motion data such as joint angle change sequences and position change sequences of the virtual pedestrian. The environmental data (RGB images, depth maps, point clouds) and motion change sequence data are aligned by timestamps to construct a navigation scenario dataset containing multiple sets of samples, which is used for training the robot's navigation algorithm.

[0063] In this application, a static three-dimensional virtual scene is acquired, and simulation environment data is determined based on the static three-dimensional virtual scene. The simulation environment data includes a two-dimensional structured navigation map and at least one global reference trajectory. The state data corresponding to the virtual pedestrian is acquired, and based on the state data and the simulation environment data, a predicted motion trajectory corresponding to the virtual pedestrian is obtained through a trained trajectory diffusion generation model. Based on the predicted motion trajectory, the virtual pedestrian is controlled to move in the static three-dimensional virtual scene.

[0064] Therefore, in the process of constructing a pedestrian behavior simulation scenario, pedestrian behavior is not simply generated according to predefined fixed rules. Instead, simulation environment data, including a two-dimensional structured navigation map and at least one global reference trajectory, is first determined based on the obtained static 3D virtual scene. Then, based on the simulation environment data and the virtual pedestrian's state data, a pre-trained trajectory diffusion generation model is used to predict the virtual pedestrian's predicted motion trajectory, thereby controlling the virtual pedestrian's movement based on the predicted trajectory. In this way, a pedestrian motion trajectory closer to the real scene can be generated based on the pre-trained trajectory diffusion generation model, making the virtual pedestrian's motion trajectory closer to the real-world pedestrian motion trajectory, which helps improve the realism of pedestrian behavior and thus improves the realism of the constructed pedestrian behavior simulation scenario.

[0065] In this embodiment of the application, the above-mentioned method for constructing a pedestrian behavior simulation scenario is further described in detail based on a specific application scenario. Figure 2 This is a schematic diagram of the technical route of a pedestrian behavior simulation scene construction method provided in an embodiment of this application, such as... Figure 2 As shown, the system mainly includes modules for offline static scene construction, pedestrian full-body control network pre-training, and GPU parallel simulation. It may also include modules for robot body simulation and navigation algorithm settings to achieve complete fully automated simulation. Based on the constructed simulation scene, training can be performed on robot navigation algorithms, such as Socially Aware Reinforcement Learning (SARL), Optimal Reciprocal Collision Avoidance (ORCA), and Relational Graph Learning (RGL).

[0066] As can be seen, to address the problems of low realism, poor behavioral fidelity, limited scene selection, and weak scalability in existing dense crowd simulation platforms, this application provides a dense crowd pedestrian behavior simulation platform based on physical simulation and data-driven approaches. This platform, in the form of a simulation data generation framework, is based on a GPU-accelerated physical simulation environment, with added frameworks that integrate rendering and physical dynamics simulation to provide real-time 3D visual feedback. It employs a trajectory diffusion model combined with an adversarial motion learning controller (AMP) data-driven scheme to generate controllable, high-fidelity full-body postures and movement trajectories of dense crowds. Furthermore, it can construct a cross-scale dynamic scene library and standardized benchmark test metrics (success rate, collision rate, path length, etc.). Ultimately, it can achieve real-time physical simulation supporting intelligent agents for crowd pedestrians, serving as a general simulation and testing tool in the field of human-machine collaborative navigation.

[0067] Based on the solution proposed in this application, the core pain points of existing technologies, such as low simulation realism, poor behavior fidelity, limited scene diversity, and weak computational scalability, can be addressed. By integrating photorealistic rendering, physically based dynamics simulation, data-driven pedestrian behavior generation, and GPU parallel computing architecture, it is possible to efficiently generate dense crowd navigation interaction data that closely approximates the real world. This provides a complete closed-loop solution for developing robot navigation strategies in complex human-robot collaborative environments, from data synthesis to policy training, significantly lowering the barrier to relying on expensive, time-consuming, and difficult-to-scale-obtain real-world data.

[0068] Figure 3 This is a schematic diagram of a full-body pedestrian intelligent agent control algorithm framework based on adversarial motion priors provided in this application embodiment. This algorithm is a motion prior modeling method that introduces adversarial learning into reinforcement learning. It performs adversarial training between the motion state sequence generated by the policy and real motion capture data (Mocap Data), thereby implicitly characterizing the high-dimensional motion distribution. Each Env represents a parallel vectorized intelligent agent environment on the GPU, which defines the state and actions of the humanoid intelligent agent in the form of tensors, and reads real-time feedback signals from the simulation data stream. Specifically, each time the humanoid intelligent agent interacts with the training environment, the data consists of tuples of Markov Decision Processes (MDPs). This describes the humanoid agent's state, actions, previous state, and reward at time t. This data is stored in a rollout buffer to support asynchronous policy optimization and is used as feature vectors during network training. Along with other observation data (trajectory, obstacles), this data serves as network input. Similar to traditional reinforcement learning (RL), the policy network computes a policy that maximizes task rewards through gradient optimization. Furthermore, by introducing a discriminator and optimizing the adversarial reward from it, and using a dataset of real motion as supervision, the actions generated by the policy network are statistically difficult to distinguish from expert motion. Figure 3 The illustrated full-body pedestrian agent control algorithm is used to compute joint-level realistic motion. Specifically, adversarial learning is used to guide the human agent to mimic motion in a real-world dataset while tracking a specific navigation target. This target-conditional reinforcement learning training leverages the capabilities of a parallel physics engine to efficiently perform simultaneous computations for multiple virtual pedestrians.

[0069] Figure 4This is a visualization diagram of a simulation platform provided in an embodiment of this application. The first row shows the robot's first-person perspective during movement, the second row shows depth information from the robot's perspective, and the third and fourth rows show external perspectives, displaying different environments such as supermarkets, homes, and lobbies. Figure 4 As shown, based on the solution presented in this application, the realism and scene diversity of the simulation can be improved. Specifically, Figure 4 This paper demonstrates the advantages of the simulation scenes generated by the proposed solution in terms of visual realism and environmental diversity. The first and second rows present the RGB images and corresponding depth information from the robot's first-person perspective, respectively. Their texture details, lighting effects, and geometric accuracy closely resemble real sensor data, providing high-quality synthetic training data for vision-based navigation algorithms. The third and fourth rows, showcasing external perspectives, demonstrate diverse cross-scale environments such as supermarkets, homes, and lobbies, validating the richness of the system's scene library. By importing 3D assets with detailed geometry and material textures and utilizing the ray tracing rendering capabilities of the physics engine, the system can output RGB-D image streams with realistic physical properties. Furthermore, the pre-built scene library, encompassing various spatial layouts (such as narrow corridors, open lobbies, and complex homes), ensures that simulation evaluations can cover a wide range of real-world navigation conditions, from simple to extremely complex, significantly improving the generalization ability of the trained navigation strategy.

[0070] Figure 5 This is a schematic diagram illustrating the training effect of an adversarial motion prior algorithm for full-body control of a virtual pedestrian provided in an embodiment of this application. Figure 5 In the diagram, the left side shows that as the number of training simulation steps increases, the virtual pedestrian interacts more frequently in one round, meaning its tracking trajectory becomes longer, until the maximum number of rounds (600) is reached, obtaining a reference trajectory with a planned duration of 3 seconds. The right side shows that as the number of training simulation steps increases (i.e., as training progresses), the final reward obtained by the algorithm gradually increases, corresponding to the improvement in the virtual pedestrian's performance (the comprehensive reward includes completing the tracking task, avoiding falls, and making movements similar to those of a real pedestrian, etc.), until the final reward converges to around 11, indicating that the policy network training is complete.

[0071] based on Figure 4 and Figure 5 As can be seen, the proposed solution can achieve high-fidelity and diverse pedestrian group behavior simulation. The generated pedestrians not only have a realistic appearance, but their movement patterns also possess biomechanical rationality and behavioral diversity. Pedestrians can exhibit natural walking, running, turning, and avoidance actions, and in multi-agent interactions, they can exhibit group behavior patterns such as cooperation, following, and diversion, avoiding problems such as trajectory abrupt changes and stiff movements caused by traditional rule-based models.

[0072] The AMP controller is responsible for executing this path at the physical level. Through adversarial learning, it mimics subtle joint-level movements from real human motion data (AMASS), ensuring that every step and every weight transfer conforms to human dynamics constraints. Specific training results can be found in [reference needed]. Figure 5 This is in the pre-training step of the pedestrian motion generation network, where a virtual pedestrian is set up to track a 3-second planned trajectory in order to train the policy network to calculate the control quantities of the virtual pedestrian's whole body joints. The planned trajectory is calculated in real time based on the pedestrian's real-time position, meaning that the reference trajectory changes over time. Figure 5 The training process is shown. As training progresses, the strategy approaches convergence and the task can be completed (the round ends if the deviation from the trajectory is greater than 0.8m).

[0073] Figure 6 This diagram illustrates a comparison of the training status of a robot navigation algorithm provided in an embodiment of this application. The solution presented in this application provides the ability to perform parallel training of robot navigation strategies in human-computer interaction navigation scenarios, thus quantitatively measuring the improvement in the efficiency and effectiveness of strategy training. Figure 6 In the diagram, the left side indicates that the algorithm training time decreases as the number of robots trained in parallel increases; the right side indicates that navigation performance improves as the number of robots trained in parallel increases.

[0074] The proposed solution enables efficient parallel computing and accelerates the algorithm development cycle. Figure 6 The efficiency advantages of the GPU-based parallel simulation architecture in this application were quantitatively verified, demonstrating that training mobile robots on the platform based on this application can quickly collect data and complete the training of robot navigation strategies more efficiently and with higher quality. This application's solution employs a SARL robot navigation algorithm (a reinforcement learning-based robot crowd navigation algorithm) to train its dense crowd navigation capabilities in a mixed pedestrian and robot environment. Figure 6 The left side shows that as the number of robot instances trained in parallel increases, the total training time required to complete a fixed number of policy updates decreases significantly, demonstrating a near-linear speedup. The right side shows that, with the same total number of simulation steps, using more parallel environments to collect data enables the navigation policy to obtain a higher final reward (i.e., better performance) because the algorithm can experience more diverse scenarios per unit of time.

[0075] Training time is significantly reduced: Based on 5000 policy updates, using the near-end policy optimization algorithm with parameters set to n_steps=32 and n_epochs=6, when the number of parallel robots increases from 1 (equivalent to traditional Gazebo single-scene simulation) to 32, the total training time is reduced from 235 seconds to 24 seconds, a reduction of 89%.

[0076] Batch size adaptation optimization: Parallel training maintains its efficiency advantage across different batch sizes (8, 16, 32). The batch size corresponds to the use of robot observation data of different dimensions for training. For example, when the batch size is 32 and the number of parallel robots is 16, the data processing volume for a single policy update is 16 times higher than that of a single robot scenario, but due to the parallel computing characteristics of GPUs, the update time per round only increases by 20%, achieving a balance between data volume and processing speed.

[0077] Excellent strategy training performance: Using 16,000-step simulation as the evaluation node, when the number of parallel robots increased from 1 to 32, the final reward of the robot navigation algorithm (defined as the average reward of the last 10 episodes at the end of training) increased from 5 to 20, a 3-fold increase in reward value; with a fixed training time of 15 minutes, the final reward of 32 parallel robot scenarios reached 18, a 125% improvement compared to a single robot scenario (reward value 8), proving that multi-scenario parallel training enables the robot navigation strategy to learn a more comprehensive obstacle avoidance strategy.

[0078] This application employs a data-driven approach to pedestrian behavior modeling: combining a trajectory diffusion model with an adversarial motion learning controller to achieve highly realistic generation of pedestrian trajectories and ensure physical consistency. The trajectory diffusion model provides diverse trajectory generation capabilities, while the AMP controller ensures that pedestrian motion conforms to biomechanical constraints. The combination of these two approaches addresses the problems of unnatural and insufficiently diverse pedestrian behavior in existing technologies.

[0079] Achieve the construction of a cross-scale scene library: Build a scene library containing 25 different sizes and complexities, covering a variety of environments from narrow indoor corridors to open outdoor plazas, providing a comprehensive evaluation benchmark for navigation algorithms and solving the problems of insufficient scene diversity in existing platforms.

[0080] It adopts a GPU parallel simulation architecture: based on the GPU parallel computing capabilities of Isaac Lab, it enables the efficient operation of large-scale simulation environments, supports the concurrent simulation of 100+ environments per second, and solves the problem of poor computing scalability of existing platforms.

[0081] Achieving multimodal perception signal simulation: It not only simulates the position and speed of pedestrians, but also their body posture, gait, gaze direction and other multimodal interactive behaviors, providing robots with richer environmental information and solving the problem that existing platforms only simulate planar trajectories.

[0082] It should be noted that in some application scenarios, the underlying simulation framework can be replaced from Isaac Lab to a Unity+PhysX combination. This requires adapting Unity's rendering interface and PhysX's physical parameters to maintain lighting, shadow rendering, and parallel computing capabilities. This makes it suitable for a wider range of development environments.

[0083] In other application scenarios, the trajectory generation and diffusion model can be replaced by a Long Short-Term Memory (LSTM) time-series prediction model. By adding an attention mechanism (paying attention to the positions of neighboring pedestrians), the rationality of the trajectory can be ensured, thereby reducing the model complexity and making it suitable for low-computing devices.

[0084] Furthermore, social comfort indicators (such as the minimum distance between pedestrians and robots, and the frequency of pedestrian gait adjustment) can be set to better meet the social needs of human-machine integration.

[0085] like Figure 7 As shown in the figure, corresponding to the above-mentioned pedestrian behavior simulation scene construction method, this application embodiment also provides a pedestrian behavior simulation scene construction system, which includes: The data processing module 710 is used to acquire a static three-dimensional virtual scene and determine simulation environment data based on the static three-dimensional virtual scene. The simulation environment data includes a two-dimensional structured navigation map and at least one global reference trajectory. The trajectory prediction module 720 is used to acquire the state data corresponding to the virtual pedestrian, and based on the state data and the simulation environment data, to acquire the predicted motion trajectory corresponding to the virtual pedestrian through the trained trajectory diffusion generation model. The motion control module 730 is used to control the movement of the virtual pedestrian in the static three-dimensional virtual scene according to the predicted motion trajectory, so as to obtain a pedestrian behavior simulation scene.

[0086] Therefore, in the process of constructing a pedestrian behavior simulation scenario, pedestrian behavior is not simply generated according to predefined fixed rules. Instead, simulation environment data, including a two-dimensional structured navigation map and at least one global reference trajectory, is first determined based on the obtained static 3D virtual scene. Then, based on the simulation environment data and the virtual pedestrian's state data, a pre-trained trajectory diffusion generation model is used to predict the virtual pedestrian's predicted motion trajectory, thereby controlling the virtual pedestrian's movement based on the predicted trajectory. In this way, a pedestrian motion trajectory closer to the real scene can be generated based on the pre-trained trajectory diffusion generation model, making the virtual pedestrian's motion trajectory closer to the real-world pedestrian motion trajectory, which helps improve the realism of pedestrian behavior and thus improves the realism of the constructed pedestrian behavior simulation scenario.

[0087] It should be noted that the specific structure and implementation of the pedestrian behavior simulation scene construction system and its various modules or units can be referred to the corresponding descriptions in the above method embodiments, and will not be repeated here.

[0088] It should be noted that the division of the various modules in the above pedestrian behavior simulation scenario construction system is not unique and is not intended as a specific limitation.

[0089] Based on the above embodiments, this application also provides a terminal, the principle block diagram of which can be as follows: Figure 8 As shown. The terminal includes a processor, memory, network interface, and display screen connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps of any of the pedestrian behavior simulation scene construction methods described above. The display screen can be a liquid crystal display (LCD) or an e-ink display.

[0090] Those skilled in the art will understand that Figure 8 The block diagram shown is only a partial structural diagram related to the solution of this application and does not constitute a limitation on the terminal on which the solution of this application is applied. The specific terminal may include more or fewer components than shown in the figure, or combine some components, or have different component arrangements.

[0091] In one embodiment, a terminal is provided, the terminal including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of any of the pedestrian behavior simulation scene construction methods provided in the embodiments of this application.

[0092] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any pedestrian behavior simulation scene construction method provided in this application.

[0093] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0094] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the above device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above device can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0095] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0096] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0097] In the embodiments provided in this application, it should be understood that the disclosed systems / terminal devices and methods can be implemented in other ways. For example, the system / terminal device embodiments described above are merely illustrative. For instance, the division of modules or units described above is merely a logical functional division, and in actual implementation, it can be divided in other ways. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0098] If the integrated modules / units described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, and software distribution media, etc. It should be noted that the content included in the computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0099] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions are not in essence a departure from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for constructing a pedestrian behavior simulation scene, characterized in that, The method includes: A static 3D virtual scene is acquired, and simulation environment data is determined based on the static 3D virtual scene, wherein the simulation environment data includes a 2D structured navigation map and at least one global reference trajectory; Obtain the state data corresponding to the virtual pedestrian; based on the state data and the simulation environment data, obtain the predicted motion trajectory corresponding to the virtual pedestrian through a trained trajectory diffusion generation model. Based on the predicted motion trajectory, the virtual pedestrian is controlled to move in the static three-dimensional virtual scene to obtain a pedestrian behavior simulation scene.

2. The method for constructing a pedestrian behavior simulation scene according to claim 1, characterized in that, The process of acquiring a static 3D virtual scene and determining simulation environment data based on the static 3D virtual scene includes: Import a generic scene description format file containing geometric and semantic information to construct the static 3D virtual scene; Spatial analysis is performed on the static three-dimensional virtual scene to generate a two-dimensional occupancy grid map representing the walkable area as the two-dimensional structured navigation map; Based on the two-dimensional occupied grid map, the global reference trajectory corresponding to the virtual pedestrian is determined by a path planning algorithm.

3. The method for constructing a pedestrian behavior simulation scene according to claim 1, characterized in that, The step of acquiring the state data corresponding to the virtual pedestrian, and based on the state data and the simulation environment data, obtaining the predicted motion trajectory corresponding to the virtual pedestrian through a trained trajectory diffusion generation model, includes: The state data and the simulation environment data are input into a trained data-driven trajectory diffusion generation model. Obtain the predicted motion trajectory output by the data-driven trajectory diffusion generation model, wherein the predicted motion trajectory is the future trajectory of the virtual pedestrian predicted by the trajectory diffusion generation model.

4. The method for constructing a pedestrian behavior simulation scene according to claim 3, characterized in that, The status data corresponding to the virtual pedestrian includes: The current state of the virtual pedestrian, the historical trajectory of the virtual pedestrian, and the trajectory of neighboring pedestrians; The predicted motion trajectory includes the coordinates of two-dimensional trajectory points of the virtual pedestrian at multiple future moments, as well as the heading angle and velocity corresponding to each two-dimensional trajectory point.

5. The method for constructing a pedestrian behavior simulation scene according to claim 1, characterized in that, The step of controlling the movement of the virtual pedestrian in the static three-dimensional virtual scene according to the predicted motion trajectory includes: Based on the predicted motion trajectory and the corresponding state data of the virtual pedestrian, a reinforcement learning controller based on adversarial motion prior training generates physical control signals to drive all joints of the virtual pedestrian. The virtual pedestrian is controlled to move within the static three-dimensional virtual scene according to the physical control signal.

6. The method for constructing a pedestrian behavior simulation scene according to claim 5, characterized in that, The virtual pedestrian is a pedestrian agent configured based on a skinned multi-person linear model; The pedestrian intelligent agent has a hinge body that can drive joints; The step of controlling the movement of the virtual pedestrian in the static three-dimensional virtual scene according to the physical control signal includes: According to the physical control signal, the hinge of the pedestrian intelligent agent is controlled so that the virtual pedestrian moves along the predicted motion trajectory in the static three-dimensional virtual scene.

7. The method for constructing a pedestrian behavior simulation scene according to any one of claims 1 to 6, characterized in that, The method further includes: By using a robot positioned within the static 3D virtual scene, scene data corresponding to the static 3D virtual scene and action change sequence data corresponding to the virtual pedestrian in the static 3D virtual scene are collected. Navigation scenario data is constructed based on the scenario data and the action change sequence data.

8. A pedestrian behavior simulation scene construction system, characterized in that, The system includes: The data processing module is used to acquire a static three-dimensional virtual scene and determine simulation environment data based on the static three-dimensional virtual scene, wherein the simulation environment data includes a two-dimensional structured navigation map and at least one global reference trajectory; The trajectory prediction module is used to acquire the state data corresponding to the virtual pedestrian, and based on the state data and the simulation environment data, to acquire the predicted motion trajectory corresponding to the virtual pedestrian through a trained trajectory diffusion generation model. The motion control module is used to control the movement of the virtual pedestrian in the static three-dimensional virtual scene according to the predicted motion trajectory, so as to obtain a pedestrian behavior simulation scene.

9. A terminal, characterized in that, The terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the pedestrian behavior simulation scene construction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the pedestrian behavior simulation scene construction method as described in any one of claims 1 to 7.

Images