Fire-fighting evacuation path optimization method for crowded place

Abstract

The invention relates to a fire-fighting evacuation path optimization method for a crowded place. The method comprises a sensor module; an image acquisition module; a personnel counting module; a voice broadcast module; and a data processing module. The method comprises the following steps: performing real-time analysis on a monitoring picture acquired by an image acquisition module; for each detected personnel target, extracting a visual feature vector, and performing modeling and prediction on the movement track of the personnel in combination with spatio-temporal context information; a global optimal evacuation strategy is realized; and broadcasting the real-time evacuation strategy obtained by the collaborative evacuation model by voice until all people are evacuated safely. In the evacuation process, the states of people in all areas are obtained in real time, evacuation decisions are dynamically updated, and voice broadcast guidance is carried out. According to the invention, through multi-modal information fusion and a deep learning algorithm, intelligent and personalized evacuation guidance for large-scale crowds is realized, and the evacuation efficiency and safety are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for optimizing fire evacuation paths in crowded places. Background Art

[0002] How to achieve accurate, efficient and safe intelligent guidance in the process of emergency evacuation in crowded places is a complex systemic problem. Traditional evacuation methods are usually static, unable to dynamically adjust and monitor according to the on-site situation, lack efficient organization, lack effective transmission of on-site fire information, and are difficult to cope with the dynamically changing environment and personnel distribution, which can easily cause panic, congestion and chaos. To solve this pain point, the key challenge is how to perceive and predict the distribution and dynamics of the crowd in real time and accurately, and formulate the optimal evacuation strategy based on this. This involves multiple technical difficulties: first, how to quickly detect and track a large number of moving targets in complex scenes; second, how to integrate multi-source heterogeneous data to accurately grasp the distribution and flow of the crowd; third, how to dynamically adjust evacuation decisions according to real-time conditions to balance efficiency and safety; and finally, how to accurately convey evacuation instructions to individuals to achieve personalized guidance. These problems are interrelated and mutually constrained, constituting a unified technical problem. Solving this problem requires not only breaking through the bottleneck of single-point technology, but also systematically coordinating various links to achieve full-process intelligence from perception, decision-making to execution. Summary of the invention

[0003] The present invention provides a method for optimizing fire evacuation paths in crowded places, comprising:

[0004] Sensor modules for real-time monitoring of environmental parameters within the building;

[0005] Image acquisition modules are distributed at key nodes of floor passages and use surveillance video streams to collect image information;

[0006] People counting module, installed at the entrance of a single floor or the entrance of the first floor of a building, verifies personnel information and counts the actual number of people entering the floor;

[0007] Voice broadcast module, installed on each floor, used to broadcast escape information; and

[0008] The data processing module is connected with the sensor module, the image acquisition module, the people counting module and the voice broadcast module, and analyzes information including environmental parameters collected by the sensor module, personnel distribution information collected by the image acquisition module, personnel quantity information collected by the people counting module and engineering design drawings of the building, calculates and generates fire escape routes and sends them to the voice broadcast module;

[0009] The method comprises:

[0010] Step S01, using an algorithm based on image segmentation and target detection, the monitoring screen collected by the image acquisition module is analyzed in real time, and the human targets in the screen are identified at a speed of 30 frames per second, and the human targets in the real-time screen are identified and the position of each person is located;

[0011] Step S02: for each detected person target, extract its 512-dimensional visual feature vector through the CSP-Darknet53 convolutional neural network; based on the feature vectors of the person target in 10 consecutive frames of video, combined with the spatiotemporal context information, use the long short-term memory (LSTM) recurrent neural network to model and predict the movement trajectory of the person;

[0012] Step S03, obtaining environmental parameters, personnel quantity information, personnel distribution information and scene map information, and constructing a collaborative evacuation model. The collaborative evacuation model implements the global optimal evacuation strategy based on factors such as personnel density, flow direction and evacuation route in different areas collected in real time by the monitoring video stream;

[0013] Step S04, the real-time evacuation strategy obtained by the collaborative evacuation model is broadcast to people in different areas by voice until all people are safely evacuated.

[0014] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0015] 1. During the evacuation process, the status of personnel in each area is obtained in real time, evacuation decisions are dynamically updated, and voice broadcast guidance is given. The present invention realizes intelligent and personalized evacuation guidance for large-scale crowds through multimodal information fusion and deep learning algorithms, effectively improving evacuation efficiency and safety.

[0016] 2. The present invention adopts the YOLOv5 algorithm to detect personnel targets in the monitoring screen in real time, extracts features and predicts motion trajectories through CSP-Darknet53 and LSTM networks, combines environmental parameters, personnel distribution and other information to build a collaborative evacuation model, and uses deep reinforcement learning algorithms to learn the optimal evacuation strategy to improve the accuracy of personnel positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The present invention is a flow chart of a method for optimizing fire evacuation routes in crowded places.

[0018] Figure 2 It is a schematic diagram of a method for optimizing fire evacuation routes in crowded places according to the present invention. DETAILED DESCRIPTION

[0019] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and embodiments. The present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the relevant inventions, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the invention are shown in the accompanying drawings.

[0020] like Figure 1-2 , a method for optimizing fire evacuation paths in crowded places, comprising:

[0021] Sensor modules for real-time monitoring of environmental parameters within the building;

[0022] Image acquisition modules are distributed at key nodes of floor passages and use surveillance video streams to collect image information;

[0023] The personnel counting module is installed at the entrance of a single floor or at the entrance at the bottom of the building to verify personnel information and count the actual number of people entering the floor; it should be noted that the personnel counting module can be a door access control system or a high-definition camera that is separately set up to verify personnel information and count the number of people;

[0024] Voice broadcast module, installed on each floor, used to broadcast escape information; and

[0025] The data processing module is connected with the sensor module, the image acquisition module, the people counting module and the voice broadcast module, and analyzes information including environmental parameters collected by the sensor module, personnel distribution information collected by the image acquisition module, personnel quantity information collected by the people counting module and engineering design drawings of the building, calculates and generates fire escape routes and sends them to the voice broadcast module;

[0026] The method comprises:

[0027] Step S01, using an algorithm based on image segmentation and target detection, the monitoring screen collected by the image acquisition module is analyzed in real time, and the human targets in the screen are identified at a speed of 30 frames per second, and the human targets in the real-time screen are identified and the position of each person is located;

[0028] Step S02: for each detected person target, extract its 512-dimensional visual feature vector through the CSP-Darknet53 convolutional neural network; based on the feature vectors of the person target in 10 consecutive frames of video, combined with the spatiotemporal context information, use the long short-term memory (LSTM) recurrent neural network to model and predict the movement trajectory of the person;

[0029] Step S03, obtaining environmental parameters, personnel quantity information, personnel distribution information and scene map information, and constructing a collaborative evacuation model. The collaborative evacuation model implements the global optimal evacuation strategy based on factors such as personnel density, flow direction and evacuation route in different areas collected in real time by the monitoring video stream;

[0030] Step S04, the real-time evacuation strategy obtained by the collaborative evacuation model is broadcast to people in different areas by voice until all people are safely evacuated.

[0031] It should be noted that during the evacuation process, the status of personnel in each area is obtained in real time, the evacuation decision is dynamically updated, and voice broadcast guidance is given. The present invention realizes intelligent and personalized evacuation guidance for large-scale crowds through multimodal information fusion and deep learning algorithms, effectively improving evacuation efficiency and safety. The present invention uses the YOLOv5 algorithm to detect personnel targets in the monitoring screen in real time, extracts features and predicts motion trajectories through CSP-Darknet53 and LSTM networks, combines environmental parameters, personnel distribution and other information, builds a collaborative evacuation model, and uses deep reinforcement learning algorithms to learn the optimal evacuation strategy to improve personnel positioning accuracy.

[0032] Furthermore, the algorithm based on image segmentation and target detection used in step S01 is the YOLOv5 algorithm; the YOLOv5 algorithm can use the monitoring screen to identify and obtain the personnel information and location information in the area.

[0033] Furthermore, modeling and predicting the movement trajectory of the person in step S02 also includes the following steps:

[0034] Step S201, using a long short-term memory (LSTM) recurrent neural network to model and predict the movement trajectory of the personnel; based on the modeling and prediction of the movement trajectory of the personnel, the K-means clustering algorithm is used to dynamically adjust the division granularity of the monitoring area, and the division granularity is increased to 5 meters in areas where the population density exceeds 2 people per square meter, and the division granularity is reduced to 2 meters in areas where the population density is less than 5 people per square meter, and the area with the most people is divided into key areas;

[0035] Step S202: For each divided area, according to the position distribution of the people inside it, 1-3 anchor points are determined by the density peak clustering algorithm as the representative positions of the area; when tracking and predicting the trajectory of the people, a local coordinate system is established with the anchor point as the center, and the Kalman filter algorithm is used for motion estimation and compensation to control the positioning error within 3 meters; by dynamically adjusting the area division granularity and anchor point setting, the positioning accuracy of the people is guaranteed to reach more than 90%.

[0036] For each detected person target, its 512-dimensional visual feature vector is extracted through the CSP-Darknet53 convolutional neural network; CSP-Darknet53 is a deep convolutional neural network combined with the cross-stage partial connection mechanism. By reducing redundant gradient information and optimizing feature extraction efficiency, it effectively solves the gradient vanishing and network redundancy problems in deep neural network training, enabling the network to achieve higher efficiency and deeper layers.

[0037] In personnel detection, CSP-Darknet53 can be used to efficiently extract multi-level visual features of personnel targets. Specifically, the detected personnel target image is input into the pre-trained CSP-Darknet53 network, and the network automatically extracts the multi-level features of the image, and finally outputs a 512-dimensional feature vector in the fully connected layer, which is calculated as follows:

[0038] ,

[0039] in, represents the extracted 512-dimensional feature vector, Represents the input person target image, CSP-Darknet53 is a convolutional neural network; this feature vector contains visual information such as the appearance, texture, and contour of the person target, which can be used for subsequent person tracking and trajectory prediction;

[0040] According to the feature vector of the person target in 10 consecutive frames of video, combined with the spatiotemporal context information, the long short-term memory (LSTM) recurrent neural network is used to model and predict the movement trajectory of the person. The calculation formula is:

[0041] ,

[0042] in, represents the hidden state at time t, represents the feature vector input at time t, represents the hidden state at time t-1, represents the long short-term memory network;

[0043] It should be noted that in video surveillance scenarios, the movement trajectory of people is continuous and time-dependent. In order to accurately predict the movement trend of people, it is necessary to consider the position and status of people at the current moment, as well as the movement history in the past period of time. The LSTM network is a special recurrent neural network that can effectively capture and utilize long-term and short-term temporal information by introducing a gating mechanism. Specifically, the person feature vectors extracted from 10 consecutive frames of video are input into the LSTM network in chronological order. The network will automatically learn the spatiotemporal laws of people's movement and output the predicted position of the person at the next moment at each time step.

[0044] By recursively predicting the positions of multiple time steps in the future, the movement trajectory of the person can be obtained.

[0045] At the same time, the LSTM network can also integrate contextual information in the scene, such as scene layout, location of other people, etc., to further improve the accuracy of trajectory prediction.

[0046] Based on the trajectory prediction results of LSTM, the K-means clustering algorithm is used to dynamically adjust the granularity of the monitoring area. The K-means algorithm divides the data points into K clusters through iterative optimization. Each cluster has a cluster center. The calculation formula for dividing the data points into K clusters is:

[0047] ;

[0048] The trajectory prediction formula of LSTM is:

[0049]

[0050] Among them, S represents the set of divided regions, k represents the number of regions, and x represents the data point. represents the center point of the i-th region; represents the state vector at time t, F represents the state transfer matrix, represents the process noise at time t; when adjusting the area division, the k value and division granularity are dynamically set according to the personnel density. The personnel density calculation formula is:

[0051] ,

[0052] in, represents the local density of point i, represents the distance between points i and j, represents the cutoff distance, Indicates when A function that is 1 if < 0 and 0 otherwise.

[0053] It should be noted that, for example, in densely populated areas, the k value can be set larger and the division granularity can be reduced to 2 meters to capture more detailed personnel distribution characteristics; for example, in the main channel of a shopping mall, due to the large number of customers, the K-means clustering algorithm can be used to quickly divide key areas and divide key areas into 2-meter grids. In sparsely populated areas, such as the parking area of ​​a shopping mall, the k value can be set smaller and the division granularity can be increased to 5 meters to reduce computing overhead (the algorithm can also be adjusted to make the division more powerful). By dynamically adjusting the area division density, the system can achieve more detailed monitoring in densely populated areas and effectively reduce the consumption of computing resources in sparse areas.

[0054] In step S202, for each divided area, according to the position distribution of the personnel inside it, 1-3 anchor points are determined as the representative positions of the area by the density peak clustering algorithm. The density peak clustering algorithm identifies the cluster center based on the local density of the data point and the distance to the point with higher density. It assumes that there is a higher local density around the cluster center and it is far away from other points with higher density. By calculating the local density and distance of each data point, the peak of the personnel density in the area can be automatically found and used as the anchor point. For example, in a 50 square meter area, two peaks of personnel density are identified by the density peak clustering algorithm, which are located in the northeast corner and southwest corner of the area respectively, and these two positions are set as anchor points. When tracking and predicting the trajectory of personnel, a local coordinate system is established with the anchor point as the center, and the Kalman filter algorithm is used for motion estimation and compensation. Kalman filtering is a recursive state estimation algorithm that estimates the system state through two steps of prediction and update. In personnel tracking, state variables usually include position, speed, etc. The prediction step predicts the state at the current moment based on the state estimation and motion model at the previous moment, and the update step uses the current observation value to correct the prediction result. Through Kalman filtering, historical trajectory information and current observation information can be effectively integrated to smooth and compensate for the movement state of personnel, thereby improving positioning accuracy;

[0055] For example, assuming that in the local coordinate system, the position of the person at the previous moment is estimated to be (3, 4) and the speed is estimated to be (1, 1), then in the prediction step, assuming that the person moves at a constant speed, the current position is predicted to be (4, 5). Then, based on the observed position at the current moment (4.2, 5.1), the position estimate is updated to (4.1, 5.05) through the Kalman filter, which effectively reduces the positioning error. By dynamically adjusting the area division granularity and anchor point setting, the personnel positioning accuracy can be guaranteed to reach more than 90% while reducing the computational complexity and resource consumption. The dynamic adjustment of the area division granularity enables more detailed analysis in densely populated areas and reduces unnecessary calculations in sparsely populated areas; the setting of anchor points provides representative positions of personnel distribution in the area, which facilitates the establishment of a local coordinate system and motion estimation.

[0056] In short, the comprehensive use of algorithms such as LSTM, K-means clustering, density peak clustering and Kalman filtering can achieve accurate modeling and prediction of personnel movement trajectories, providing reliable data support for subsequent evacuation guidance.

[0057] Furthermore, step S03 also includes the following steps:

[0058] Step S301, combining the sensor module, the image acquisition module and the people counting module, obtaining the environmental parameters collected by the sensor module, the people distribution information collected in real time by the image acquisition module in combination with the target recognition algorithm, and the number of people information collected by the people counting module, and combining the scene map and historical data to build a collaborative evacuation model, and set the state space, action space and reward function of the collaborative evacuation model;

[0059] Step S302, using high-definition cameras and intelligent access control devices deployed in key areas such as entrances and exits and key passages, and using technologies such as face recognition and identity verification to obtain personnel identity and entry and exit status information in real time;

[0060] The DeepQ-Network algorithm is used to train the collaborative evacuation model to learn the optimal evacuation strategy;

[0061] Step S303, simulate the evacuation strategy of fire escape, model and predict the division of key areas based on the movement trajectory of personnel, and give priority to the dynamic evacuation guidance of key areas when simulating the evacuation strategy of fire escape; at the same time, dynamically adjust the access control strategy according to the density and flow of personnel in the area, control the flow of personnel in and out, and provide real-time evacuation guidance information through voice broadcast module and display screen equipment, optimize evacuation routes, and improve evacuation efficiency;

[0062] Step S304: During the training process, factors such as the density of people in each area, flow direction, and evacuation route are comprehensively considered and used as the observation state of the collaborative evacuation model; by setting a suitable reward function, the intelligent agent is guided to learn to avoid congestion, reduce evacuation time, and improve evacuation efficiency.

[0063] It should be noted that in step S301, the sensor module can collect environmental parameters such as temperature, humidity, smoke concentration, etc. These parameters can reflect the severity and diffusion trend of the fire. The image acquisition module, combined with the target recognition algorithm, can monitor the distribution of people in the scene in real time, including the number of people, location, movement trajectory, etc. The personnel counting module can count the flow of people at the entrance and exit and grasp the total number of people in the scene. Combining these three types of data with scene maps and historical data, a comprehensive collaborative evacuation model can be constructed. The state space of the model includes information such as environmental parameters, personnel distribution and quantity in each area, and the action space includes evacuation guidance strategies, such as opening or closing certain access control, broadcasting guidance information, etc. The reward function is set according to indicators such as evacuation efficiency and personnel safety to guide the model to learn the optimal strategy; in step S302, high-definition cameras and intelligent access control devices are arranged in key areas such as entrances and exits and key channels. Through face recognition and identity authentication technology, the identity information of people entering and leaving can be accurately obtained, and the status of people can be tracked in real time. For example, the camera installed at the entrance of the office area can identify the identity of employees and record the time they enter the office area; the access control equipment installed in the fire passage can monitor the identity and flow of evacuated personnel. These identity and status information can be combined with the personnel distribution information in the collaborative evacuation model to more accurately predict the evacuation process and control the evacuation strategy. At the same time, deep reinforcement learning algorithms such as the DeepQ-Network algorithm can be used to train the collaborative evacuation model so that it can autonomously learn and optimize the evacuation strategy based on the real-time observed state, so as to adapt to complex and changeable evacuation scenarios; in step S303, by modeling and predicting the movement trajectory of personnel, key areas such as densely populated areas and congested areas can be identified. For example, in the office area, employees often stay in their own workstations, meeting rooms, tea rooms and other areas for a long time, forming a gathering of people; during the evacuation process, areas such as stairways and emergency exits are prone to congestion. The collaborative evacuation model can give priority to evacuation guidance to these key areas, and control the flow of people in the area by dynamically adjusting the access control strategy. At the same time, it is also possible to provide clear evacuation guidance information to the personnel in the area through voice broadcasts, display screens and other devices, such as "Please evacuate to Exit A in an orderly manner", etc., to guide the personnel to choose the optimal evacuation route. During the evacuation process, the model can also optimize the evacuation strategy in real time according to the changes in the density and flow of people in the area, such as opening alternative evacuation channels, adjusting guidance information, etc., to improve the overall evacuation efficiency; in step S304, by setting a reward function, such as rewarding +1 point for every 10% reduction in evacuation time, the intelligent agent is guided to learn to optimize the evacuation efficiency. In the actual evacuation process, the location, density and flow direction of personnel in each area are obtained in real time through cameras, sensors, etc., and input into the intelligent agent model. According to the current state, the intelligent agent combines the learned strategy, such as guiding personnel to change routes if the evacuation route is congested, controlling the flow speed to reduce by 20%, etc., to dynamically adjust the evacuation decision.Repeat the above process and continuously optimize the decision strategy until all personnel are safely evacuated. For example, in a certain scenario, there are 1,000 people who need to be evacuated, and the average evacuation speed of each exit is 50 people / min. The initial plan estimates that the evacuation time is 20 minutes. Through the dynamic decision optimization of the intelligent agent, people are guided to be diverted to different exits, the evacuation speed is increased to 60 people / min, and the final evacuation time is shortened to 17 minutes, improving the evacuation efficiency by 20%.

[0064] Furthermore, step S04 also includes the following steps:

[0065] Step S401, during the actual evacuation process, real-time acquisition of personnel status information in each area, including location, density, and flow direction, etc., as input to the collaborative evacuation model;

[0066] Step S402, the collaborative evacuation model updates the evacuation status, combines the learned evacuation strategy, dynamically adjusts the evacuation decision and announces it by voice, guides personnel to change the evacuation route and controls the flow speed of personnel;

[0067] Step S403, continuously repeating step S401 and step S402 until all personnel are safely evacuated.

[0068] The above contents are merely examples and explanations of the structure of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the structure of the invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A method for optimizing fire evacuation paths in crowded places, characterized in that: include: Sensor modules for real-time monitoring of environmental parameters within the building; Image acquisition modules are distributed at key nodes of floor passages and use surveillance video streams to collect image information; People counting module, installed at the entrance of a single floor or the entrance of the first floor of a building, verifies personnel information and counts the actual number of people entering the floor; Voice broadcast module, installed on each floor, used to broadcast escape information; as well as The data processing module is connected with the sensor module, the image acquisition module, the people counting module and the voice broadcast module, and analyzes information including environmental parameters collected by the sensor module, personnel distribution information collected by the image acquisition module, personnel quantity information collected by the people counting module and engineering design drawings of the building, calculates and generates fire escape routes and sends them to the voice broadcast module; The method comprises: Step S01, using an algorithm based on image segmentation and target detection, the monitoring screen collected by the image acquisition module is analyzed in real time, and the human targets in the screen are identified at a speed of 30 frames per second, and the human targets in the real-time screen are identified and the position of each person is located; Step S02: for each detected person target, extract its 512-dimensional visual feature vector through the CSP-Darknet53 convolutional neural network; based on the feature vectors of the person target in 10 consecutive frames of video, combined with the spatiotemporal context information, use the long short-term memory (LSTM) recurrent neural network to model and predict the movement trajectory of the person; Step S03, obtaining environmental parameters, personnel quantity information, personnel distribution information and scene map information, and constructing a collaborative evacuation model. The collaborative evacuation model implements the global optimal evacuation strategy based on the personnel density, flow direction and evacuation route of different areas collected in real time by the monitoring video stream; Step S04, the real-time evacuation strategy obtained by the collaborative evacuation model is broadcast to people in different areas by voice until all people are safely evacuated.

2. A method for optimizing fire evacuation paths in crowded places according to claim 1, characterized in that: The algorithm based on image segmentation and target detection used in step S01 is the YOLOv5 algorithm; the YOLOv5 algorithm can use the monitoring screen to identify and obtain the personnel information and location information in the area.

3. The method for optimizing fire evacuation paths in crowded places according to claim 1, characterized in that: The modeling and prediction of the movement trajectory of the person in step S02 also includes the following steps: Step S201, using a long short-term memory (LSTM) recurrent neural network to model and predict the movement trajectory of the personnel; based on the modeling and prediction of the movement trajectory of the personnel, the K-means clustering algorithm is used to dynamically adjust the division granularity of the monitoring area, and the division granularity is increased to 5 meters in areas where the population density exceeds 2 people per square meter, and the division granularity is reduced to 2 meters in areas where the population density is less than 5 people per square meter, and the area with the most people is divided into key areas; Step S202: For each divided area, according to the position distribution of the people inside it, 1-3 anchor points are determined by the density peak clustering algorithm as the representative positions of the area; when tracking and predicting the trajectory of the people, a local coordinate system is established with the anchor point as the center, and the Kalman filter algorithm is used for motion estimation and compensation to control the positioning error within 3 meters; by dynamically adjusting the area division granularity and anchor point setting, the positioning accuracy of the people is guaranteed to reach more than 90%.

4. A method for optimizing fire evacuation paths in crowded places according to claim 3, characterized in that: Step S03 also includes the following steps: Step S301, combining the sensor module, the image acquisition module and the people counting module, obtaining the environmental parameters collected by the sensor module, the people distribution information collected in real time by the image acquisition module in combination with the target recognition algorithm, and the number of people information collected by the people counting module, and combining the scene map and historical data to build a collaborative evacuation model, and set the state space, action space and reward function of the collaborative evacuation model; Step S302, using high-definition cameras and intelligent access control devices deployed at entrances and exits, key channels and key areas, and using face recognition and identity verification technology to obtain personnel identity and entry and exit status information in real time; The DeepQ-Network algorithm is used to train the collaborative evacuation model to learn the optimal evacuation strategy; Step S303, simulate the evacuation strategy of fire escape, model and predict the division of key areas based on the movement trajectory of personnel, and give priority to the dynamic evacuation guidance of key areas when simulating the evacuation strategy of fire escape; at the same time, dynamically adjust the access control strategy according to the density and flow of personnel in the area, control the flow of personnel in and out, and provide real-time evacuation guidance information through voice broadcast module and display screen equipment, optimize evacuation routes, and improve evacuation efficiency; Step S304: During the training process, the density of people, flow direction and evacuation route in each area are comprehensively considered and used as the observation state of the collaborative evacuation model; by setting a suitable reward function, the agent is guided to learn to avoid congestion, reduce evacuation time and improve evacuation efficiency.

5. The method for optimizing fire evacuation paths in crowded places according to claim 1, characterized in that: Step S04 also includes the following steps: Step S401, during the actual evacuation process, real-time acquisition of personnel status information in each area, including location, density and flow direction as input to the collaborative evacuation model; Step S402, the collaborative evacuation model updates the evacuation status, combines the learned evacuation strategy, dynamically adjusts the evacuation decision and announces it by voice, guides personnel to change the evacuation route and controls the flow speed of personnel; Step S403, continuously repeating step S401 and step S402 until all personnel are safely evacuated.

Images