An ai decision pedestrian simulation method for station-city integration area

Abstract

The application discloses an AI decision pedestrian simulation method for a station-city integration area. The method is based on a single-agent independent iteration mechanism, constructs a perception-decision-execution standardized closed-loop cycle simulation architecture, and is suitable for a complex dynamic passenger flow scene of the station-city integration area. Specifically, the method is characterized in that: the state of a single agent and the visual field environment data are collected in real time, and a standardized perception description is generated in a text form; a large language model is driven by combining a personalized role prompt word to complete personification situation reasoning and behavior decision; through semantic analysis and a special mapping rule library, a natural language decision intention is accurately converted into an executable atomic action instruction sequence of a simulation engine; the agent is driven to complete an action and the individual state and the global environment are updated in real time, continuous iteration and circulation are realized, and multi-agent parallel high-fidelity simulation is realized. The application specifically solves the problems of behavior homogenization, decision solidification, no independent cognitive cycle and poor dynamic adaptability in traditional simulation, accurately restores real personalized judgment, environment response and social interaction behavior of pedestrians, greatly improves the passenger flow simulation precision and engineering practical value, and can be widely applied to hub planning and design, operation deduction, large passenger flow early warning and emergency evacuation scenes.

Description

Technical Field

[0001] This invention belongs to the field of computer simulation technology and artificial intelligence large model fusion technology, specifically involving a highly realistic pedestrian simulation method for station-city integrated areas based on single agent closed-loop cycle and human-like decision-making of large language model. Background Technology

[0002] The station-city integration area is a large-scale comprehensive transportation hub that highly integrates rail transit, surface public transportation, commercial facilities, and office and accommodation facilities. It is characterized by high passenger density, complex personnel composition, strong randomness of pedestrian behavior, frequent individual interactions, and dynamic and ever-changing scene environment. It is a core, key, and challenging area for urban transportation operation and safety management. In the early planning and layout of the hub, daily passenger flow operation scheduling, and emergency evacuation plan preparation, pedestrian simulation technology is an indispensable digital analysis tool. By simulating the movement, waiting, ticket checking, shopping, avoidance, and wayfinding behaviors of massive numbers of pedestrians, it is possible to quantitatively assess the rationality of facility layout, passage capacity, congestion risk points, and emergency evacuation efficiency, providing accurate data support for hub design optimization and operation management.

[0003] Currently, mainstream pedestrian simulation technologies in the industry are based on social force models, cellular automata, and finite state machine rule models. These technologies adopt a group-wide rule-driven mode, where all pedestrian agents share fixed mathematical formulas and preset behavioral logic, resulting in inherent technical defects. First, agents lack independent perception and decision-making capabilities, and the behavioral logic of agents with similar roles is completely homogenized, failing to reproduce the differentiated behaviors brought about by individual personality, emotions, travel experience, and urgency in catching a bus in real-world scenarios, leading to monotonous and rigid simulation results. Second, the decision-making loop is fixed and static, lacking a closed-loop iterative mechanism for real-time perception and dynamic reasoning. Agents cannot respond to dynamic scenario variables such as broadcast notifications, sudden congestion, and changes in facility status, and their behavioral patterns remain fixed throughout. Third, simulation of individual social interaction is lacking, failing to simulate the micro-behaviors of individual pedestrians autonomously avoiding, following, conforming, and cooperating. It can only achieve macro-level group movement simulation, resulting in highly idealized simulation results that deviate significantly from the micro-operational state of passenger flow in real transportation hubs, greatly reducing the reliability and engineering reference value of simulation data.

[0004] Large language models possess powerful natural language understanding, complex scene reasoning, and anthropomorphic autonomous decision-making capabilities, providing a new technological direction for building pedestrian intelligent agents with independent cognitive abilities. However, existing technologies cannot achieve deep adaptation between the cognitive capabilities of large models and high-precision pedestrian simulation systems, resulting in several technical bottlenecks: large model decision outputs are highly random and lack scene rule constraints, easily leading to invalid decisions that violate physical rules and station operation rules; there is a lack of standardized single-agent perception-decision-execution closed-loop processes, resulting in a disconnect between semantic decision-making and simulated physical actions, making continuous iterative simulation impossible; and the large model reasoning mechanism is mismatched with the simulation step-size iteration mechanism, making it difficult to achieve real-time dynamic behavior updates for a single intelligent agent. Currently, the industry has not yet formed a mature, implementable, and reusable single-agent AI closed-loop decision-making station-city integrated pedestrian simulation solution.

[0005] In summary, existing traditional pedestrian simulation methods suffer from low individual realism, lack of independent cognitive loops, insufficient intelligence, and poor dynamic adaptability. Pure large-scale model decision-making simulations lack rule constraints and closed-loop implementation capabilities. The industry urgently needs a new pedestrian simulation method based on a single intelligent agent's independent closed-loop loop, which takes into account anthropomorphic intelligent decision-making, scene rule constraints, and high-precision simulation execution. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies, such as the lack of independent cognitive loops, homogenized behavior, rigid decision-making, insufficient realism, and uncontrollable AI decision-making in pedestrian intelligent agents. It provides an AI-based pedestrian simulation method for urban-station integrated areas. This invention innovatively constructs a complete closed-loop system of perception-decision-execution for a single intelligent agent. Through a large language model, it endows each pedestrian intelligent agent with independent high-order anthropomorphic cognition and dynamic decision-making capabilities. Combined with a standardized semantic mapping mechanism, it balances the intelligence, realism, and standardization of single-agent simulation, accurately reproducing pedestrian micro-level personalized behavior and social interaction. This solves the technical problem that traditional simulation cannot simulate individual dynamic response and differentiated decision-making, significantly improving the accuracy and engineering application value of pedestrian simulation in urban-station integrated areas.

[0007] To achieve the above objectives, this invention provides an AI-based pedestrian simulation method for station-city integrated areas. It employs a mode of independent closed-loop simulation by a single agent and parallel simulation by multiple agents, comprising the following iteratively executed steps: S1. Environmental perception and textual description: At each simulation step, the self-state information of a single target pedestrian intelligent agent and the surrounding environment information within the preset fan-shaped field of view in the three-dimensional simulation environment of the station-city fusion area are independently collected. According to the predefined natural language description template, standardized natural language perception description text adapted to the input of the large model is generated. S2. Intelligent Decision-Making Based on a Large Language Model: The natural language perception description text corresponding to a single agent is concatenated and fused with the preset agent role attribute system prompt words to form a complete input prompt, which is then imported into the large language model. The large language model performs independent scene understanding, contextual reasoning, and behavioral decision-making for the current single agent scenario, and outputs natural language decision text representing the agent's next behavioral intention. The knowledge base of scene behavior rules in the station-city fusion area is retrieved to perform multi-dimensional verification on the original decision text of the single agent output by the large language model, correcting, replacing, or rejecting illegal decisions, and outputting compliant standardized decision text. S3. Decision text parsing and action instruction mapping: Perform multimodal semantic parsing on the compliant natural language decision text of a single agent, extract the core action intent and key behavioral parameters, and accurately map the semantic behavioral intent into atomic action instructions and parameter sequences that can be recognized and executed by the simulation engine according to the preset mapping rule library. S4. Simulation Action Execution and Environment Update: The simulation engine loads and executes the atomic action instruction sequence of a single agent, driving the pedestrian agent to complete the corresponding behavior in the simulation scene, and synchronously updates the individual state of the agent, the state of surrounding facilities and the global passenger flow environment, completing a single simulation cycle for a single agent; steps S1 to S4 are executed in a loop to achieve continuous closed-loop simulation of a single agent, and multiple agents are used in parallel iteration to complete the dynamic simulation of pedestrians in the whole domain.

[0008] Furthermore, in step S1, the self-state information includes the agent's unique identity identifier, real-time location coordinates, physiological state, psychological and emotional state, travel goal, planned train number, and remaining departure time; the surrounding environment information is scene data within the agent's preset fan-shaped field of view, including the location, movement speed, and walking posture of other agents, the location, occupancy status, and vacancy status of various station public facilities, as well as on-site directional signs and real-time broadcast operation information; the station public facilities include turnstiles, ticket vending machines, restrooms, commercial shops, elevators, and staircases.

[0009] Furthermore, in step S2, the system prompts are used to construct personalized decision-making logic exclusive to a single intelligent agent. The configuration content of the system prompts includes the agent's basic identity attributes, personality tendencies, decision-making preference styles, and output format constraint instructions. Among them, the basic identity attributes include business travelers, family passengers with children, pick-up personnel, ordinary waiting passengers, and station staff; the personality tendencies include impatient and sensitive, cautious and conservative, and calm and steady; and the decision-making styles include efficiency-first, comfort-first, and cautious conformity.

[0010] Furthermore, in step S3, the semantic parsing integrates three methods: keyword rule matching, dependency parsing, and lightweight semantic understanding model, to achieve accurate deconstruction and parameter extraction of single-agent natural language decision text; the atomic action commands include fixed-point path movement, directional attitude turning, stationary waiting, orderly queuing, facility interaction operation, dynamic adjustment of movement speed, and pedestrian posture and expression switching.

[0011] Furthermore, in step S2, the knowledge base of behavior rules for the station-city integration area is divided into three major constraint dimensions: physical constraint rules, which limit pedestrian passage boundaries, facility interaction logic and motion limit parameters; social behavior rules, which regulate crowd interaction behaviors such as queuing, yielding, following, cooperating and conforming; and station operation rules, which limit ticket checking time periods, passenger flow restrictions, facility opening and use standards and emergency scenario constraint requirements.

[0012] Furthermore, in step S1, the natural language perception description text supports adaptive switching between first-person and third-person dual templates, automatically matching the corresponding description logic according to the individual agent role type and simulation scenario conditions, adapting to the cognitive expression habits of different groups of people, and improving the decision fit of the single agent large model.

[0013] Furthermore, in step S4, a pedestrian feature adaptation model is preset. For special travel agents such as those carrying large luggage, elderly pedestrians, accompanying children, and people with disabilities, the simulation parameters such as moving speed, passing and avoidance distance, and start-stop buffer time are dynamically corrected during the execution of single agent actions to restore the differentiated movement characteristics of real people.

[0014] Furthermore, in step S3, the mapping rule base establishes a natural language semantic intent-multi-atomic action sequence association and adaptation mechanism. For a single semantic decision intent of a single intelligent agent, it can automatically match and combine multiple continuous atomic action instructions to adapt to the complex continuous travel behavior of pedestrians walking, finding their way, and interacting with facilities.

[0015] Furthermore, each intelligent agent operates independently in a complete closed-loop cycle, with no interference between the perception, decision-making, and execution processes of each agent. Only during the action execution phase is a socialized interactive linkage achieved through global environmental data to enable pedestrian avoidance, following, and collaboration.

[0016] Furthermore, during the iterative simulation cycle, the environmental perception data, historical decision text, action execution logs, and environmental interaction records of each agent are stored in real time to form a multi-agent simulation behavior dataset, which is used for simulation result review, customer behavior pattern mining, model parameter iterative optimization, and simulation accuracy verification.

[0017] The beneficial effects of this invention are: Compared with existing technologies, this invention, based on a single-agent independent closed-loop architecture, has the following outstanding advantages and beneficial effects: 1. Constructing a unique cognitive loop for each individual pedestrian agent to completely solve the problem of homogenized behavior. This invention builds an independent perception-decision-execution closed-loop iterative system for each pedestrian agent. Relying on the high-order personalized cognitive reasoning capabilities of a large language model, each pedestrian agent can generate unique decisions and differentiated behaviors based on its own role, personality, emotions, and the urgency of its journey. This completely breaks away from the drawbacks of traditional simulation groups with uniform rules and rigid individual behaviors, highly reproducing the micro-personalized behavioral characteristics of real pedestrians in urban-station integrated areas, and significantly improving the realism of the simulation.

[0018] 2. The closed-loop architecture has a clear hierarchy, balancing AI intelligence and simulation controllability. This invention uses natural language as a universal interaction medium, decoupling the high-level semantic cognitive decision-making of large models from the low-level high-precision physical control of the simulation engine. At the same time, it constrains the randomness of single-agent AI output through independent decision verification links, thus retaining the advantages of anthropomorphic intelligent decision-making while ensuring that the simulation process of single-agent and multi-agent systems is controllable, repeatable, and quantifiable, fully adapting to the accuracy requirements of engineering simulation.

[0019] 3. Supports dynamic adaptation and refined social interaction, with extremely high scenario fit. Based on a single-agent independent perception and decision-making mechanism, each pedestrian agent can capture dynamic scene changes within its field of vision in real time, autonomously adjust its travel decisions, and accurately respond to scenario variables such as broadcast notifications, passenger congestion, and facility status switching; at the same time, multiple agents can naturally achieve refined social interactions such as queuing, yielding, following, cooperating, and conforming through environmental data linkage, perfectly adapting to the dynamic and ever-changing real-world operational scenarios of station-city integration areas with complex individual interactions.

[0020] 4. The architecture boasts strong versatility and high engineering feasibility, adapting to various application scenarios. This invention's single-agent closed-loop architecture is adaptable to pedestrian agents with all identities, physical characteristics, and travel needs. It can be quickly migrated and applied to various engineering scenarios such as hub facility layout optimization, daily passenger flow simulation, large passenger flow risk early warning, and emergency evacuation simulation. The architecture is stable, reliable, and highly reusable, possessing extremely high engineering promotion value. (See attached figures for details.) Figure 1 A flowchart of the overall multi-agent closed-loop simulation of the AI ​​decision-making pedestrian simulation method for station-city integration areas provided in this embodiment of the invention; Figure 2 This is a flowchart illustrating the complete workflow of a single intelligent agent in a single simulation cycle, which is the core of this invention. Figure 3 This is an example diagram showing the single-agent environmental perception description text and the decision output results of the large language model corresponding to an embodiment of the present invention. Detailed Implementation

[0021] The following will refer to the appendices in the embodiments of the present invention. Figure 1 Appendix Figure 2 Appendix Figure 3 The specific technical solutions of the present invention will be clearly and completely described in detail. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0022] The overall simulation operation logic of this invention is as follows: Figure 1 As shown in the overall multi-agent closed-loop simulation flowchart, after the system completes initialization, it constructs a group of pedestrian intelligent agents covering multiple roles such as passengers rushing to catch trains, people picking up passengers, passengers waiting for trains, and station staff. All intelligent agents are deployed to a standardized three-dimensional station-city fusion area simulation environment. Each intelligent agent runs in parallel with an independent simulation step size to continuously complete the closed-loop iterative simulation.

[0023] The refined operational logic of the single agent in this invention strictly follows Figure 2 A detailed diagram of the closed-loop perception-decision-execution process for a single intelligent agent is presented. Each simulation cycle strictly follows the complete iterative chain of environmental perception and textual description, large-model intelligent decision-making, decision semantic parsing and action mapping, simulation action execution, and global environment update, achieving personalized, dynamic, and constrained human-like behavior simulation of a single intelligent agent. The following example, a typical business traveler carrying large luggage rushing to catch a train, illustrates the complete closed-loop behavioral decision-making and execution process within a single simulation step: Step S1: Environmental Perception and Textual Description correspond Figure 2 In the intelligent agent perception and acquisition module, at the beginning of the current simulation step, the agent's dedicated perception unit independently completes its own state acquisition and fan-shaped field-of-view environment acquisition.

[0024] S11. Status and Environmental Data Collection: Accurately collect the agent's own structured status data: The agent's role is a business traveler carrying large luggage, currently located in the center of the waiting hall, with the core travel goal being to reach the B12 ticket gate, 18 minutes remaining before train departure, and an anxious emotional state. Simultaneously, collect real-time environmental information within the agent's forward fan-shaped field of vision: 20 meters directly ahead is the entrance to the B area passage, with a loose queue of approximately 10 people at the entrance; a vertical elevator is located 15 meters to the left front, and the equipment is available for use; a pedestrian staircase is to the right front, with moderate pedestrian density in the area; and a live broadcast announces operational information urging passengers boarding train Gxxx to check in and board as soon as possible.

[0025] S12. Standardized Natural Language Text Generation: This function calls a preset first-person natural language description template, automatically matching the collected structured individual data with environmental data to generate standardized perceptual description text P that can be directly input into a large language model. The text content is completely consistent with the target language model. Figure 3 Example of a scenario description: I am a business traveler with large luggage, currently in the center of the waiting hall. I need to get to gate B12, and with 18 minutes until departure, I'm feeling a bit anxious. I see about 10 people queuing at the entrance to section B directly ahead; there's an empty elevator to my left; and a staircase to my right with moderate foot traffic. I hear an announcement that passengers taking train Gxxx should board as soon as possible. What should I do next? Step S2: Human-like Intelligent Decision Making Based on Large Language Model correspond Figure 2 The Zhongda Model Intelligent Decision Module completes situational reasoning and autonomous decision-making based on personalized role settings.

[0026] S21. Prompt Message Fusion Configuration: Retrieve the exclusive system prompt message S adapted to this agent's role. The prompt message is used to solidify the agent's personality traits and decision-making preferences. Specifically: You are a businessperson carrying a heavy suitcase, rushing to catch a train in the station-city fusion area. You value efficiency but have limited mobility and will prioritize the easiest and quickest methods. Based solely on what you see and hear, please state your most desired action in one sentence. Combine the system prompt message S with the perceptual description text P to form a complete, standardized input prompt.

[0027] S22, Large Model Inference Output: The combined input content is pushed to the large language model via a standard API interface. The large language model combines the agent's identity, real-time scene, emotional state, and travel needs to complete contextual inference and output accurate natural language action decision text D, corresponding to... Figure 3 The model's decision output is: proceed to the straight elevator to the left front.

[0028] Between steps S2 and S3, there is also step S2a: decision compliance verification. This step differs from traditional unconstrained AI decision-making simulations. It retrieves a pre-defined knowledge base of behavioral rules for station-city integrated areas, comprehensively validating the decision text output by the large model across three dimensions: physical access rules, social interaction norms, and station operation rules. The verification shows that the intention to go to the left-facing escalator complies with the hub's access boundaries, facility usage regulations, and the requirements of the catching-the-train scenario; there are no violations, logical conflicts, or issues, and it is deemed executable, thus a compliant decision and directly sent to the next semantic analysis stage. The decision text output by the large language model is then compared with a knowledge base of behavioral rules for station-city integrated areas. If the action intention described in the decision text violates pre-defined physical rules, social norms, or operational rules, it is corrected, replaced, or rejected according to the rule knowledge base, and the compliant decision text or corrected result is sent to subsequent analysis steps. This step ensures that AI decision-making, while demonstrating creativity, does not violate the basic logic of simulation and real-world constraints.

[0029] Step S3: Decision text parsing and action instruction mapping correspond Figure 2 The semantic parsing and instruction mapping module enables precise conversion from high-level semantic decisions to low-level simulation actions.

[0030] S31. Semantic parsing and parameter extraction: By combining keyword matching with dependency parsing, the compliance decision text D is deeply deconstructed to accurately identify the core action intent as using the elevator facility, and key behavioral parameters are extracted: the target facility is the empty elevator to the left front.

[0031] S32, Atomic Action Sequence Mapping: Calls the preset semantic-action mapping rule library, matches and generates a continuous and executable atomic action instruction sequence based on the semantic intent of using the facility. The specific chain is as follows: path planning, target point: simulation coordinates corresponding to the door of the elevator, posture turning, body facing the left elevator direction, facility interaction, facility type: elevator, execution action: call and enter, completing the standardized conversion from natural language decision to instructions that can be recognized by the simulation engine.

[0032] Step S4: Simulation Action Execution and Environment Iterative Update correspond Figure 2 The middle action execution and environment update module completes single-cycle simulation execution and starts closed-loop iteration.

[0033] S41. Simulated Action Execution: The agent execution unit receives and parses the complete atomic action instruction sequence. First, it calls the global path planning algorithm to calculate the optimal obstacle avoidance path from the current position to the left front elevator door. Combining the preset luggage impact feature model, it adaptively reduces the moving speed and increases the passage buffer distance based on the agent's carrying large luggage, driving the agent to move smoothly along the planned path. During the movement, other pedestrian agents in the scene autonomously complete social interaction behaviors such as obstacle avoidance and fine-tuning of walking trajectories based on their own independent perception and decision-making logic, which conforms to the characteristics of real passenger flow interaction.

[0034] S42. Environmental Status Update and Closed-Loop Iteration: When the agent arrives at the target location of the elevator, it triggers the facility interaction command to complete the simulated behavior of waiting for and entering the elevator. The simulation engine updates the agent's spatial location, scene status, and behavioral trajectory information in real time, while simultaneously updating the occupancy status of the elevator equipment to complete the current full simulation step.

[0035] The simulation system automatically enters the next iteration cycle. The agent, based on its new state inside the elevator, restarts a new closed-loop process of environmental perception, intelligent decision-making, and action execution, achieving continuous dynamic simulation.

[0036] It should be noted that the above embodiments are only used to clearly illustrate the core technical solution of the present invention and are not intended to limit the scope of protection of the present invention. Although this embodiment provides a detailed closed-loop simulation description of an intelligent agent for a business traveler carrying large luggage, those skilled in the art can make conventional modifications or equivalent substitutions to the prompt word parameters, behavior mapping rules, and scene adaptation parameters of different intelligent agents based on the technical concept and core closed-loop architecture of the present invention. Any optimizations, improvements, and substitutions made based on the core technical logic of the present invention—single-agent independent perception and decision-making closed loop, large-model anthropomorphic reasoning combined with rule verification—do not depart from the core spirit and scope of protection of the present invention.

Claims

1. An AI-based pedestrian simulation method for station-city integrated areas, characterized in that, The simulation adopts a mode of independent closed-loop operation for a single agent and parallel simulation for multiple agents, including the following steps of continuous iterative execution: S1. Environmental perception and textual description: At each simulation step, the self-state information of a single target pedestrian intelligent agent and the surrounding environment information within the preset fan-shaped field of view in the three-dimensional simulation environment of the station-city fusion area are independently collected. According to the predefined natural language description template, standardized natural language perception description text adapted to the input of the large model is generated. S2. Intelligent Decision-Making Based on a Large Language Model: The natural language perception description text corresponding to a single agent is concatenated and fused with the preset agent role attribute system prompt words to form a complete input prompt, which is then imported into the large language model. The large language model performs independent scene understanding, contextual reasoning, and behavioral decision-making for the current single agent scenario, and outputs natural language decision text representing the agent's next behavioral intention. The knowledge base of scene behavior rules in the station-city fusion area is retrieved to perform multi-dimensional verification on the original decision text of the single agent output by the large language model, correcting, replacing, or rejecting illegal decisions, and outputting compliant standardized decision text. S3. Decision text parsing and action instruction mapping: Perform multimodal semantic parsing on the compliant natural language decision text of a single agent, extract the core action intent and key behavioral parameters, and accurately map the semantic behavioral intent into atomic action instructions and parameter sequences that can be recognized and executed by the simulation engine according to the preset mapping rule library. S4. Simulation Action Execution and Environment Update: The simulation engine loads and executes the atomic action instruction sequence of a single agent, driving the pedestrian agent to complete the corresponding behavior in the simulation scene, and synchronously updates the individual state of the agent, the state of surrounding facilities and the global passenger flow environment, completing a single simulation cycle for a single agent; steps S1 to S4 are executed in a loop to achieve continuous closed-loop simulation of a single agent, and multiple agents are used in parallel iteration to complete the dynamic simulation of pedestrians in the whole domain.

2. The AI-based pedestrian simulation method for station-city integration areas according to claim 1, characterized in that, In step S1, the self-state information includes the agent's unique identity identifier, real-time location coordinates, physiological state, psychological and emotional state, travel goal, planned train number, and remaining departure time; the surrounding environment information is scene data within the agent's preset fan-shaped field of view, including the location, movement speed, and walking posture of other agents, the location, occupancy status, and vacancy status of various station public facilities, as well as on-site directional signs and real-time broadcast operation information; the station public facilities include turnstiles, ticket vending machines, restrooms, commercial shops, elevators, and staircases.

3. The AI-based pedestrian simulation method for station-city integration areas according to claim 1, characterized in that, In step S2, the system prompts are used to construct personalized decision-making logic for a single agent. The configuration of the system prompts includes the agent's basic identity attributes, personality traits, decision-making preference styles, and output format constraints. The basic identity attributes include business travelers, families with children, pick-up personnel, ordinary waiting passengers, and station staff. The personality traits include impatient and sensitive, cautious and conservative, and calm and steady. The decision-making styles include efficiency-first, comfort-first, and cautious conformity.

4. The AI-based pedestrian simulation method for station-city integration areas according to claim 1, characterized in that, In step S3, the semantic parsing integrates three methods: keyword rule matching, dependency parsing, and lightweight semantic understanding model, to achieve accurate deconstruction and parameter extraction of single-agent natural language decision text; the atomic action commands include fixed-point path movement, directional attitude turning, stationary waiting, orderly queuing, facility interaction operation, dynamic adjustment of movement speed, and pedestrian posture and expression switching.

5. The AI-based pedestrian simulation method for station-city integration areas according to claim 1, characterized in that, In step S2, the knowledge base of behavior rules for the station-city integration area is divided into three major constraint dimensions: physical constraint rules, which limit pedestrian passage boundaries, facility interaction logic and motion limit parameters; social behavior rules, which regulate crowd interaction behaviors such as queuing, yielding, following, cooperating and conforming; and station operation rules, which limit ticket checking time periods, passenger flow restrictions, facility opening and use standards and emergency scenario constraint requirements.

6. The AI-based pedestrian simulation method for station-city integration areas according to claim 1, characterized in that, In step S1, the natural language perception description text supports adaptive switching between first-person and third-person dual templates, automatically matching the corresponding description logic according to the individual agent role type and simulation scenario conditions, adapting to the cognitive expression habits of different groups of people, and improving the decision fit of the single agent large model.

7. The AI-based pedestrian simulation method for station-city integration areas according to claim 1, characterized in that, In step S4, a pedestrian feature adaptation model is preset. For special travel agents such as those carrying large luggage, elderly pedestrians, accompanying children, and people with disabilities, the simulation parameters such as moving speed, passing and avoidance distance, and start-stop buffer time are dynamically corrected during the execution of single agent actions to restore the differentiated movement characteristics of real people.

8. The AI-based pedestrian simulation method for station-city integration areas according to claim 1, characterized in that, In step S3, the mapping rule base establishes a natural language semantic intent-multi-atomic action sequence association and adaptation mechanism. For a single semantic decision intent of a single intelligent agent, it can automatically match and combine multiple continuous atomic action instructions to adapt to the complex continuous travel behavior of pedestrians walking, finding their way, and interacting with facilities.

9. The AI-based pedestrian simulation method for station-city integration areas according to claim 1, characterized in that, Each intelligent agent operates independently in a complete closed-loop process. The perception, decision-making, and execution processes of each intelligent agent do not interfere with each other. Only during the action execution phase, social interaction and linkage such as pedestrian avoidance, following, and cooperation are achieved through global environmental data.

10. The AI-based pedestrian simulation method for station-city integration areas according to claim 1, characterized in that, During the iterative simulation cycle, the environmental perception data, historical decision text, action execution logs, and environmental interaction records of each agent are stored in real time to form a multi-agent simulation behavior dataset, which is used for simulation result review, customer behavior pattern mining, model parameter iterative optimization, and simulation accuracy verification.

Images