An intelligent evacuation method for urban rail transit station fires based on surveillance videos

Abstract

The present invention improves the fire evacuation method in the subway station based on surveillance videos. First, computer vision technology is used to identify fires, determine their locations and development scales; at the same time, the locations and numbers of people to be evacuated are identified to provide a data basis for evacuation route planning. Second, factors such as the fire location, crowd distribution, passing capacities of each evacuation exit, evacuation distances, and path curvatures are comprehensively considered, and multiple algorithms are integrated to quickly optimize and solve the evacuation routes, so as to maximize the overall evacuation efficiency of the station. Third, this technology can be deployed based on the existing monitoring system in the subway station, and the planned escape routes are used to guide the evacuation of station members through safety indication signs, making full use of the existing hardware facilities and keeping the application cost controllable.

Description

Technical Field

[0001] The present invention relates to the technical field of urban rail transit, and particularly relates to an intelligent evacuation method for urban rail transit station fires based on surveillance videos. Background Art

[0002] There are many internal facilities and equipment in urban rail transit stations, and the structure is complex; moreover, most stations are located underground, with relatively enclosed spaces; at the same time, as a passenger flow gathering place, the stations carry a large number of passengers. Therefore, in the event of safety accidents such as fires, a scientific and reasonable rapid evacuation plan is an important guarantee for ensuring the safety of passengers traveling in the station.

[0003] Currently, there are mainly the following problems in the fire evacuation technology of subway stations: (1) Fire information, crowd distribution information, and exit passing efficiency information are not comprehensively considered in the evacuation path planning. (2) The path planning algorithm mainly focuses on theoretical research and less on practical applications. Most of the current research on the fire evacuation path planning algorithm of subway stations stays at the theoretical level and does not fully consider the requirements of actual application scenarios. (3) The application cost of most technologies is relatively high, and the existing hardware facilities of subway stations are not fully utilized. Subway stations are usually equipped with hardware facilities such as cameras, guiding signs, and safety doors, but these facilities are rarely fully utilized in the existing intelligent evacuation technologies. Summary of the Invention

[0004] The present invention aims to provide an intelligent evacuation method for urban rail transit station fires based on surveillance videos to solve the above problems.

[0005] The technical solution of the present invention is: an intelligent evacuation method for urban rail transit station fires based on surveillance videos, including the following steps:

[0006] S1, obtaining fire scene information: using object recognition technology (You Only Look Once) to identify fire scene information, using the principle of distortion correction and camera calibration to obtain the positions of fires and obstacles, and constructing an evacuation space topology model accordingly;

[0007] S2, fusing the A* and ant colony algorithms: using the A* algorithm to adjust the initial pheromone concentration of the ant colony to make the path finding model more refined;

[0008] S3, implementing and optimizing the evacuation path fusion algorithm: introducing end point guidance, turning penalty, and exit congestion effect factors, and reconstructing the distance heuristic function, state transition probability function, and evacuation path length calculation function to make the evacuation path fusion algorithm have better global regulation and evacuation efficiency.

[0009] Preferably, the specific steps of step S1 include:

[0010] S11, Fire scene information recognition: Based on object recognition technology, obtain the fire information and personnel information in the fire scene. Among them, the process of obtaining personnel information includes: First, use the station camera to obtain video data input, perform grid processing on the video data, use a convolutional neural network to establish a prior box for crowd recognition and a target classification probability map, then perform pedestrian detection, and at the same time control the movement of the camera for pedestrian tracking;

[0011] S12, Target location acquisition: Based on the distortion correction model, obtain the spatial location information of the fire; The process of obtaining the target spatial location is to first correct the distortion generated during the video imaging process, then use the camera calibration principle to obtain the world coordinates in the world coordinate system calibrated by the camera, and finally construct an evacuation space topology model to calculate the coordinates of the target in the real scene;

[0012] S13, Construction of evacuation space topology model: Take a certain location in the real site as the coordinate origin, convert the position of the detected target in the camera calibration coordinates into the position in the real scene coordinates, and accurately express the position of the target in this scene.

[0013] Preferably, S11 specifically includes:

[0014] The specific fire scene information recognition is:

[0015] Extract the personal dynamic position and trajectory from the video data for personnel detection and tracking, select the method of head detection, and match it with the high-density crowd scene; Use the YOLO (You Only Look Once) algorithm and the Deep SORT (Simple Online and Realtime Tracking with a Deep Association Metric) algorithm to establish a crowd head detection + tracking framework;

[0016] Specifically: Use the head detection method and the YOLO algorithm for crowd detection, and then combine the Deep SORT algorithm to achieve crowd tracking, which can extract individual dynamic information from the crowd image, achieve accurate personnel detection and tracking, then use the distortion correction model to adjust the camera parameters to achieve the projection display of the crowd image, and obtain the spatial location information of the target according to the camera calibration principle; Considering that the movement of people in the video will cause the tracking ID of the same individual to change, it is necessary to calculate the trajectory vector of the pedestrian trajectory in the video to predict the next movement trajectory of the pedestrian. Take the predicted trajectory and the actual trajectory of the pedestrian in the video as input, and through the IO matching method, set the people whose feature similarity meets the predefined threshold as the matching people, so as to achieve continuous tracking of the same individual;

[0017] The detected bounding boxes and corresponding confidence information are input into the Deep SORT module. At the next time point, new detection results are input, and it is determined whether the predicted trajectory matches the new detection results. To perform the matching, the weighted similarity is obtained based on the similarity of the motion and appearance features between the predicted Kalman trajectory and the newly detected trajectory. The Hungarian algorithm, a combinatorial optimization algorithm, is used to input the weighted similarity to obtain a matching matrix, which helps improve the robustness against partial occlusion.

[0018] Preferably, S12 specifically includes:

[0019] S121, a distortion correction model;

[0020] Image distortion occurs during the video imaging process. The coordinates are corrected before converting the fire location coordinates. Considering both the tangential distortion and radial distortion of the camera, the distortion correction model is obtained as shown in Equation (1):

[0021]

[0022] where the left side of the equation is the corrected coordinates; the first half of the right side is the radial distortion model, and the second half of the right side is the tangential distortion model; X W 、Y W are the fire location coordinates, X Wcorrected 、Y Wcorrected are the corrected coordinates, The parameters k1, k2, k3, p1, p2 are from the calibrated distortion coefficients;

[0023] S122, inverse calculation of the coordinate system based on the camera calibration principle to convert the image pixel coordinates to world coordinates to obtain the target position; specifically including:

[0024] Four coordinate systems are involved in the camera calibration process: the world coordinate system Ow, the camera coordinate system Oc, the image physical coordinate system, and the image pixel coordinate system. For a pinhole camera, two parameter sets, namely the external and internal camera parameters, are used to convert a coordinate in the actual scene to a pixel coordinate in the image frame. Among them, the external parameters realize the conversion from the world coordinate system to the camera coordinate system, and the internal parameters realize the conversion from the camera coordinate to the pixel coordinate system in the image frame;

[0025] The external parameters include the rotation matrix R and the translation vector t; the rotation matrix describes the rotation information of the camera relative to the three coordinate axes of the world coordinate system; the translation vector describes the translation information of the optical center of the camera relative to the origin of the world coordinate system;

[0026] Taking the point pw(xw, yw, zw) in the world coordinate system as an example, it is converted using the following Equation (2) to the point pc(xc, yc, zc) in the camera coordinate system;

[0027]

[0028] The point pc is projected onto the two-dimensional image plane as the point p(x, y).

[0029]

[0030] Then, using formula (4), the point p in the image coordinate system is converted to pi(u, v) with the image resolution:

[0031]

[0032] where f is the camera focal length; dx and dy are the lengths of one pixel; u0 and v0 are the optical centers in pixels;

[0033] By combining and arranging the above equations, the conversion formula from the world coordinate system to the pixel coordinate system can be obtained as the following formula (5):

[0034]

[0035] where the first term on the right side of the formula is the camera intrinsic matrix, and the second term on the right side is the camera extrinsic matrix, which is obtained through camera calibration.

[0036] Preferably, S13 specifically includes:

[0037] There are differences in scale, rotation, and translation between the world coordinate system calibrated by the camera and the coordinate system of the actual scene, which requires conversion between the coordinate systems; calculate formula (2):

[0038] B = r × A + t (6)

[0039] where A is the set of known points in the world coordinate system, B is the set of known points in the actual scene, r is the rotation matrix, t is the translation vector, and A i and B i are vectors with three rows and one column, such as [x y z] T ; the steps to find the optimal rigid transformation matrix are as in formula (3):

[0040]

[0041] Bring the two data sets to the origin, then find the rotation matrix r and calculate the translation vector t; obtain the calculation formula (4) of the covariance matrix H:

[0042] H = (A - centroid A ) × (B - centriod B ) T (8)

[0043] Then \(H\) is the covariance matrix. Using the singular value decomposition algorithm \(SVD(H)=[U,S,V]\), let \(R = VU\). T ; Considering the different scales of the two coordinate systems, the calculation method of the scale multiple \(\lambda\) between them is given by the formula (reference here):

[0044]

[0045] Formula (2) will become:

[0046]

[0047] Combined with formula (6), the rotation matrix \(r\) and translation vector \(t\) are obtained as follows:

[0048]

[0049] At this point, the actual scene coordinates are calculated.

[0050] Preferably, in step S2, the specific steps include:

[0051] S21. Establish an evacuation space topology model: Create a topology model that describes the structure of the evacuation area and the distribution of obstacles;

[0052] S22. Generate a sub - optimal evacuation path using the A* algorithm: Use the A* algorithm to generate a sub - optimal path, considering obstacle avoidance and the shortest path;

[0053] S23. Adjust the pheromone concentration of topological information elements: According to the sub - optimal path, adjust the pheromone concentration on the topology model;

[0054] S24. Initialize the parameters of the ant colony algorithm: Set the parameters of the ant colony algorithm, such as the number of ants and the pheromone evaporation rate;

[0055] S25. Start searching for the evacuation path: The ant colony starts to search for the evacuation path according to the pheromone concentration and distance;

[0056] S26. Reach the end point once and complete the search: When the ant colony reaches the end point once, complete one search.

[0057] S27. Update the pheromone concentration: According to the search results, update the pheromone concentration in the topology model; Usually, successful paths increase the pheromone concentration, while failed paths decrease the concentration;

[0058] S28. Perform iteration: If the maximum number of iterations is not reached, search again;

[0059] S29. Output the optimal path: When the maximum number of iterations is reached, output the best path as the result;

[0060] The calculation formula of the evaluation function f(n) of the A* algorithm is shown in (8):

[0061] f(n) = g(n) + h(n) (12)

[0062] The Euclidean distance is used to calculate the distance between two points. The calculation method of the distance between node i and node j is as shown in Equation (9):

[0063]

[0064] g(n) is a dissipation function used to calculate the cost from the starting point s to the current node i, as shown in Equation (10):

[0065]

[0066] h(n) is a heuristic function used to calculate the cost from the current node i to the target node t, as shown in Equation (11):

[0067]

[0068] After the spatial topology model is constructed, according to the corresponding starting and ending points and the selected cost function calculation method, the sub-optimal path planning of the A* algorithm is carried out;

[0069] In the path generated above, adjust the pheromone concentration τ ij (0), increase the pheromone concentration on the sub-optimal path, and the increment is A ij , after adjusting the initial pheromone concentration τ ij (0), that is, make the following transformation of τ ij (0) as shown in Equation (12):

[0070] τ ij (0) = τ0 + A ij (16)

[0071] Among them, τ0 is the default value of the initial pheromone concentration, and A ij is the increased value of the initial pheromone concentration in the adjustment area obtained by the A* algorithm.

[0072] Preferably, the specific steps of step S3 include:

[0073] S31, algorithm improvement of the end-point guiding effect:

[0074] Considering the guiding effect of the end point, on the premise of calculating the visibility between the current node and the next node, increase the visibility d between the next node and the end point jtThrough the calculation, the evacuation purpose is enhanced, making the passenger evacuation in the subway station fire more directional; squaring the reciprocal of the sum of two distances can shorten the running time and improve the efficiency of searching for the shortest path in the actual algorithm; based on the above improvement strategies, the distance heuristic function η is reconstructed. ij As shown in Equation (13):

[0075]

[0076] where i is the current node;

[0077] j is the next node;

[0078] t is the end point;

[0079] Idea of code writing: Define dis_1 as a list containing distance values, and traverse the elements in the dis_1 list through a loop; in the loop, calculate the heuristic function value of each distance value; the calculation method of the heuristic function is to add the distance values together and take the reciprocal of the square of the result as the value of the heuristic function; by processing the distance values, the heuristic function guides the ants to choose shorter paths in the ant colony algorithm; after the loop ends, the elements in the dis_1 list are updated to the corresponding heuristic function values.

[0080] S32, turning penalty effect:

[0081] Due to the dense population, harsh conditions accompanied by smoke and the panic psychology of the masses inside the station, if the evacuation path is tortuous and the number of turns is large, the passengers waiting to be evacuated in the station will get lost due to excessive panic, making it more difficult to escape and reducing the evacuation efficiency; therefore, by calculating the turning angle θ and counting the number of turns tn, considering the turning penalty z (the larger the turning angle θ, the smaller the z value; when the path is straight and no turning is required, z = 1; as the number of turns tn increases, the p k ij (t) value becomes smaller), adjust the state transition probability p k ij (t) in the ant colony algorithm, add the turning penalty and count the number of turns to reduce the probability of turning in path search, reduce the complexity of the evacuation path, improve the smoothness of the escape path, and reduce the possibility of passengers getting lost:

[0082]

[0083] where θ is the turning angle of the ant when moving between adjacent nodes;

[0084] tn is the number of turns of the ant;

[0085] z is the turning penalty, and its value range is 0.5 to 1;

[0086] Idea for code writing: Define two empty lists prob and prob_z to store the probabilities of nodes being selected. Next, traverse the elements in the node_be_selected_1 list through a loop. It is a list containing optional nodes. Call the angle function to calculate the turning angle z to calculate the angle between the current position and the candidate node node_be_selected_1[i], and consider the turning penalty. Next, calculate the selection probability p_z of the node, which combines the distance weight dis_imp, the pheromone concentration weight pher_imp, and the number of turning times weight z turn_num ; Add the calculated node selection probability p_z to the prob_z list. Then, update the node selection probability by dividing each node selection probability by prob_sum. Subsequently, compare to obtain the maximum node transition probability. Finally, after the loop ends, the selected node is recorded as node_be_selected_1[k];

[0087] S33, Exit congestion effect:

[0088] When the passenger flow in the subway station is too large, the exits closer to the passengers are very likely to be congested. The passengers to be evacuated often need to queue up, which undoubtedly increases the evacuation time and reduces the evacuation efficiency. In this case, choosing the exit with the shortest escape route length for evacuation does not necessarily mean that this is the best way to evacuate. To meet the basic principles of rapid and safe evacuation, it is necessary and essential to optimize the above fire scene evacuation path model;

[0089] The evacuation optimization idea considering the exit congestion effect is: In the initial stage of evacuation, first evacuate the people closest to the evacuation exit quickly. During the evacuation path planning process, update the population distribution density of each evacuation exit crowd_exit, combine the congestion degree of each exit, calculate the congestion coefficient crowd_imp and the distance way_len_best from the exit to the people to be evacuated, and comprehensively consider the distribution position of the fire in the subway station to reasonably plan and flexibly adjust the evacuation path, so as to achieve the rapid and safe evacuation of the entire subway platform crowd in case of fire;

[0090] Specific implementation steps for optimizing the exit congestion effect: After projecting the personnel information identified by the YOLO algorithm onto the spatial topology model, according to the different groups of people to be evacuated peop, the distribution center points of the crowds are extracted respectively, and a two-dimensional list peop related to the groups of people to be evacuated is generated. The parameter peop_posi in the list is the position coordinates of the center point of a certain batch of people to be evacuated, and the parameter peop_density is the density of a certain batch of people to be evacuated; after evacuating the first batch of people without congestion effect, the number of the first batch of passengers to be evacuated is recorded as the crowd density at the evacuation exit crowd_exit. By introducing the congestion coefficient crowd_imp of the corresponding exit, the distance of the second batch of people to be evacuated from the exit is recalculated. After the evacuation of this batch of passengers is completed, the platform crowd density and the exit congestion coefficient are updated continuously to complete the evacuation of subsequent batches of passengers.

[0091] The beneficial effects of the present invention are as follows:

[0092] 1. Using computer vision technology to obtain the fire location and the development of the fire situation in real time. The object recognition algorithm can accurately detect the location of the fire and update the development of the fire situation in a timely manner. This real-time information acquisition enables the evacuation plan to be more accurate and flexible, and respond to the changes in the fire situation in a timely manner.

[0093] 2. This technology comprehensively considers the distribution of the evacuated crowd and the distribution of the fire source to optimize the evacuation route. The distribution of the crowd and the location information of the fire source are obtained through the object recognition algorithm. In the evacuation route planning, the system will comprehensively consider these factors and select the best evacuation route to avoid the fire source area to the greatest extent and reduce the congestion in the crowded areas.

[0094] 3. This technology realizes the optimization of multi-start and multi-end evacuation routes. In actual fire scenarios, there may be multiple starting points and multiple ending points. This technology can flexibly plan multiple evacuation routes according to the location information of the starting point and the ending point, as well as the development of the fire, to reduce the evacuation time. Description of the Drawings

[0095] Figure 1 It is a flowchart of crowd information recognition in an intelligent evacuation method for urban rail transit stations based on surveillance videos provided by an embodiment of the present invention;

[0096] Figure 2 It is a flowchart of the "crowd head detection + tracking model" in an intelligent evacuation method for urban rail transit stations based on surveillance videos provided by an embodiment of the present invention;

[0097] Figure 3 It is a schematic diagram of the transformation from the camera coordinate system to the image coordinate system in an intelligent evacuation method for urban rail transit stations based on surveillance videos provided by an embodiment of the present invention;

[0098] Figure 4 Flow chart of the A* fusion traditional ant colony algorithm in an intelligent evacuation method for urban rail transit station fires provided by an embodiment of the present invention;

[0099] Figure 5 Sub - optimal path graph generated by the A* algorithm in an intelligent evacuation method for urban rail transit station fires provided by an embodiment of the present invention;

[0100] Figure 6 Distortion map of the monitoring camera in an intelligent evacuation method for urban rail transit station fires provided by an embodiment of the present invention;

[0101] Figure 7 Restored effect diagram of the monitoring lens in an intelligent evacuation method for urban rail transit station fires provided by an embodiment of the present invention;

[0102] Figure 8 Schematic diagram of the construction effect of the evacuation space topological model in an intelligent evacuation method for urban rail transit station fires provided by an embodiment of the present invention;

[0103] Figure 9 Scene division map of crowd distribution in an intelligent evacuation method for urban rail transit station fires provided by an embodiment of the present invention;

[0104] Figure 10 Path planning diagram (crowd distribution scenario 1) under the simulated subway station fire scenario in an intelligent evacuation method for urban rail transit station fires provided by an embodiment of the present invention;

[0105] Figure 11 Path planning diagram (crowd distribution scenario 2) under the simulated subway station fire scenario in an intelligent evacuation method for urban rail transit station fires provided by an embodiment of the present invention;

[0106] Figure 12 Path planning diagram (crowd distribution scenario 3) under the simulated subway station fire scenario in an intelligent evacuation method for urban rail transit station fires provided by an embodiment of the present invention. Detailed implementation manners

[0107] The following further illustrates the present invention in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present invention and be able to implement it. The implementation manners of the present invention are not limited thereto.

[0108] Embodiment 1

[0109] The technical solution of the present invention is divided into four steps: 1. Obtaining fire scene information, using object recognition technology (You Only Look Once) to identify fire scene information, using the principles of distortion correction and camera calibration to obtain the positions of fires and obstacles, and constructing an evacuation space topology model accordingly; 2. Fusion of A* and ant colony algorithms, using the A* algorithm to adjust the initial pheromone concentration of the ant colony to make the pathfinding model more reasonable and precise; 3. Implementation and optimization of the evacuation path fusion algorithm, introducing end-point orientation, turning penalty, and exit congestion effect factors, reconstructing the distance heuristic function, state transition probability function, and evacuation path length calculation function to make the evacuation path fusion algorithm have better global regulation and evacuation efficiency; 4. Verifying the technical feasibility by combining actual cases.

[0110] 1 Fire information acquisition

[0111] 1.1 Fire scene information recognition

[0112] Based on object recognition technology, the acquisition of fire information and personnel information in the fire scene is realized. The process of personnel information acquisition is as Figure 1 shown, and the process of fire information recognition is the same.

[0113] 1.1.1 Recognition method

[0114] Extract personal dynamic positions and trajectories from video data for personnel detection and tracking. Considering the matching with high-density crowd scenes, the method of head detection is selected. Using the YOLO (You Only Look Once) algorithm and the Deep SORT (Simple Online and Realtime Tracking with a Deep Association Metric) algorithm, a crowd head detection + tracking framework is established, as Figure 2 shown. Among them, the YOLO algorithm is used for crowd detection. The Deep SORT algorithm is used for crowd tracking.

[0115] Input information such as the detected bounding boxes and corresponding confidence levels into the Deep SORT module. At the next time point, new detection results are input, and it is judged whether the predicted trajectory matches the new detection results. For matching, the weighted similarity is obtained according to the similarity of the motion and appearance features between the predicted Kalman trajectory and the newly detected trajectory. Using the Hungarian algorithm, a combinatorial optimization algorithm, input the weighted similarity to obtain a matching matrix, which helps to improve the robustness weakened by partial occlusion. Such a detection + tracking type framework can obtain individual trajectories in high-density crowd scenes.

[0116] 1.2 Target position acquisition

[0117] 1.2.1 Distortion correction model

[0118] During the video imaging process, image distortion will occur. The coordinates need to be corrected before calculating the fire location coordinates. Considering the tangential distortion and radial distortion of the camera comprehensively, the distortion correction model is obtained as shown in the formula:

[0119]

[0120] Among them, the left side of the equation is the corrected coordinate; the first half of the right side is the radial distortion model, and the second half of the right side is the tangential distortion model. X W , Y W are the fire location coordinates, X Wcorrected , Y Wcorrected are the corrected coordinates. The parameters k1, k2, k3, p1, p2 come from the calibrated distortion coefficients.

[0121] 1.2.2 Camera Calibration Principle

[0122] Based on the camera calibration principle, the coordinate system is calculated inversely to convert the image pixel coordinates into world coordinates to obtain the target position.

[0123] 1.3 Construction of Evacuation Space Topological Model

[0124] Taking a certain location in the real site as the coordinate origin, the position of the detection target in the camera calibration coordinates is converted into the position in the real scene coordinates to accurately express the position of the target in this scene.

[0125] As Figure 3 shown, there are differences in scale, rotation, and translation between the world coordinate system of camera calibration and the coordinate system of the actual scene. This requires conversion between coordinate systems. The only solution requires at least 3 points. The basic calculation formula:

[0126] B = r × A + t (21)

[0127] Among them, A is the set of known points in the world coordinate system, B is the set of known points in the actual scene, r is the rotation matrix, t is the translation vector, A i and B i are vectors with three rows and one column, such as [x y z] T . The steps to find the optimal rigid transformation matrix are as shown in the formula:

[0128]

[0129] Bring the two data sets to the origin, then find the rotation matrix r and calculate the translation vector t. The calculation formula for the covariance matrix H:

[0130] H = (A - centroid A)×(B - centroid B ) T (23)

[0131] Then \(H\) is the covariance matrix. Using the singular value decomposition algorithm \(SVD(H)=[U, S, V]\), let \(R = VU\). T . Considering that the scales of the two coordinate systems are different, the calculation method of the scale multiple \(\lambda\) between them is given by the formula (refer to here for the formula):

[0132]

[0133] The basic formula (21) will become:

[0134]

[0135] Combined with formula (25), the rotation matrix \(r\) and the translation vector \(t\) can be obtained respectively as:

[0136]

[0137] So far, the actual scene coordinates can be calculated.

[0138] Fusion of \(2A*\) and ant colony algorithm

[0139] 2.1 Fusion of \(A*\) algorithm

[0140] To solve the problems such as the relatively long calculation time and the tendency to generate local optimal solutions easily in the traditional ACO algorithm for fire evacuation route planning, the present invention adjusts by re - assigning the initial information of the ant colony with the \(A*\) algorithm, which can improve the search efficiency in the initial stage of the ACO algorithm, make up for its shortcomings, optimize the traditional ACO algorithm, and generate the best evacuation path.

[0141] Use the \(A*\) algorithm to plan the path first and adjust the pheromone concentration on this path. Select an appropriate pheromone concentration increment and increase the pheromone concentration on the sub - optimal path generated by the \(A*\) algorithm, which can have a certain guiding effect on the movement of ants, thus shortening the calculation time and improving the search efficiency of the algorithm.

[0142] 2.2 Implementation of the fusion algorithm

[0143] The fusion process is as Figure 4 shown:

[0144] The calculation formula of the evaluation function \(f(n)\) of the \(A*\) algorithm is shown in:

[0145] \(f(n)=g(n)+h(n)\) (27)

[0146] The present invention uses the Euclidean distance to calculate the distance between two points. The calculation method of the distance between node i and node j is as follows:

[0147]

[0148] g(n) is a dissipation function used to calculate the cost from the starting point s to the current node i, as follows:

[0149]

[0150] h(n) is a heuristic function used to calculate the cost from the current node i to the target node t, as follows:

[0151]

[0152] After the spatial topology model is constructed, according to the corresponding starting and ending points and the selected cost function calculation method, the sub-optimal path planning of the A* algorithm is carried out, and the generated path is as Figure 5 shown.

[0153] In the path generated above, the pheromone concentration τ ij (0) is adjusted to increase the pheromone concentration on the sub-optimal path, and the increment is A ij , after adjusting the initial pheromone concentration τ ij (0), that is, making a transformation on τ ij (0):

[0154] τ ij (0) = τ0 + A ij (31)

[0155] where τ0 is the default value of the initial pheromone concentration, and A ij is the increased value of the initial pheromone concentration in the adjustment area obtained by the A* algorithm.

[0156] Initialize the parameters of the ant colony algorithm, and set the pheromone concentration increment A ij on the sub-optimal path to 8.

[0157] From the perspective of algorithm search and running time analysis, the path search efficiency is compressed from 18.69 s of the A* algorithm and 17.55 s of the traditional algorithm to 16.93 s of the integrated A* ant colony algorithm. The search time is significantly shortened, and the evacuation efficiency can be improved in the actual scenario, which is applicable to the actual evacuation scenario.

[0158] 3 Optimization of the evacuation path algorithm

[0159] Considering the influencing factors of the comprehensive evacuation cost, introduce the corresponding effect factors to correct the function, obtain the optimized evacuation cost function, and use the model that integrates the A* and ant colony algorithms to carry out path planning.

[0160] 3.1 Algorithm Improvement of End - Oriented Effect

[0161] Considering the guiding effect of the end, on the premise of calculating the visibility between the current node and the next node, the present invention adds the calculation of the visibility d jt between the next node and the end, enhancing the evacuation purpose and making the passenger evacuation in subway station fires more directional; squaring the reciprocal of the sum of the two distances, this processing can shorten the running time and improve the efficiency of searching for the shortest path in the actual algorithm. Based on the above improvement strategies, the present invention reconstructs the distance heuristic function η ij As the formula:

[0162]

[0163] where i is the current node;

[0164] j is the next node;

[0165] t is the end point.

[0166] Idea of code writing: Define dis_1 as a list containing distance values, and traverse the elements in the dis_1 list through a loop. In the loop, calculate the heuristic function value of each distance value. The calculation method of the heuristic function is to add the distance values together and take the reciprocal of the square of the result as the value of the heuristic function. By processing the distance values, the heuristic function can guide the ants to choose shorter paths in the ant colony algorithm. After the loop ends, the elements in the dis_1 list are updated to the corresponding heuristic function values.

[0167] 3.2 Turning Penalty Effect

[0168] The interior of the station is densely populated, accompanied by harsh conditions such as smoke and the panic psychology of the masses. If the evacuation path is too tortuous and the number of turns is too large, the passengers waiting to be evacuated in the station will get lost due to excessive panic, making it more difficult to escape and reducing the evacuation efficiency. Therefore, the present invention calculates the turning angle θ and counts the number of turns tn, considering the turning penalty z (the larger the turning angle θ, the smaller the z value. When the path is straight and no turning is required, z = 1. As the number of turns tn increases, p k ij (t) value becomes smaller), adjusts the state transition probability p k ij (t) in the ant colony algorithm, adds a turning penalty coefficient and counts the number of turns to reduce the probability of turning in path search, reduce the complexity of the evacuation path, improve the smoothness of the escape path, and reduce the possibility of passengers getting lost:

[0169]

[0170] Among them, θ is the turning angle of the ant when traveling between adjacent nodes;

[0171] tn is the number of turns of the ant;

[0172] z is the turning penalty coefficient, and its value range is 0.5 to 1.

[0173] Idea for code writing: Define two empty lists prob and prob_z to store the probabilities of nodes being selected. Next, traverse the elements in the node_be_selected_1 list through a loop. It is a list containing optional nodes. Call the angle function to calculate the turning angle z to calculate the angle between the current position and the candidate node node_be_selected_1[i], and consider the turning penalty. Next, calculate the selection probability p_z of the node, where the distance weight dis_imp, the pheromone concentration weight pher_imp, and the number of turns weight z are combined turn_num . Add the calculated node selection probability p_z to the prob_z list. Then, update the node selection probability by dividing each node selection probability by prob_sum. Subsequently, compare to obtain the maximum node transfer probability. Finally, after the loop ends, the selected node is recorded as node_be_selected_1[k].

[0174] 3.3 Exit congestion effect

[0175] When the passenger flow in the subway station is too large, the exits closer to the passengers are very likely to be congested. The passengers to be evacuated often need to queue up, which undoubtedly increases the evacuation time and reduces the evacuation efficiency. In this case, choosing the exit with the shortest escape route length for evacuation does not necessarily mean that this is the best way to evacuate. To meet the basic principles of rapid and safe evacuation, it is necessary and essential to optimize the above evacuation path model for the fire scenario.

[0176] The evacuation optimization idea considering the exit congestion effect is: In the initial stage of evacuation, first evacuate the people closest to the evacuation exits quickly. During the evacuation path planning process, update the population distribution density of each evacuation exit crowd_exit, combine the congestion degree of each exit, calculate the congestion coefficient crowd_imp and the distance way_len_best from the exit to the people to be evacuated, and comprehensively consider the distribution position of the fire in the subway station to reasonably plan and flexibly adjust the evacuation path, so as to achieve the rapid and safe evacuation of the entire subway platform crowd in case of fire.

[0177] Specific implementation steps for optimizing the exit congestion effect: After projecting the personnel information identified by YOLO onto the spatial topology model, according to the different groups of people to be evacuated peop, the distribution center points of the crowds are extracted respectively, and a two-dimensional list peop related to the crowds is generated. The parameter peop_posi in the list is the position coordinates of the center point of a certain group of people to be evacuated, and the parameter peop_density is the density of a certain group of people to be evacuated. After evacuating the first batch of people without congestion effect, the number of the first batch of passengers to be evacuated is recorded as the crowd density at the evacuation exit crowd_exit. By introducing the congestion coefficient crowd_imp of the corresponding exit, the distance of the second batch of people to be evacuated from the exit is recalculated. After the evacuation of this batch of passengers is completed, the platform crowd density and the exit congestion coefficient are updated continuously to complete the evacuation of subsequent batches of passengers.

[0178] 4 Case Study

[0179] 4.1 Monitoring Distortion Repair

[0180] Taking the platform floor plan of a certain station on the Beijing Subway as an example, an evacuation topology network model is constructed. The distorted image of the monitoring camera is as Figure 6 shown. Using the offline video data to identify the distribution and quantity of passengers on the platform, assuming the location of the fire, the camera parameters are calibrated using the floor image. The camera matrix and lens distortion coefficients finally obtained through experimental test data are shown in Table 1.

[0181] Table 1 Camera Calibration Parameters

[0182]

[0183] After obtaining the parameters through camera calibration in the present invention, the distorted monitoring images are repaired, realizing the de-distortion of the camera, improving the accuracy of the model position detection, so as to project the passenger personnel information and fire information onto the grid model subsequently. The monitoring camera distortion repair effect is as Figure 7 shown.

[0184] 4.2 Evacuation Route Planning

[0185] 4.2.1 Constructing the Evacuation Space Topology Model

[0186] According to the station plane design drawing, an evacuation topology network model is constructed. Using the offline video data to identify the distribution and quantity of passengers on the platform, assuming the fire location to construct the evacuation space topology model, as Figure 8 shown.

[0187] 4.2.2 Simulating the Crowd Distribution Scenario

[0188] As Figure 9As shown, different population densities and distribution locations will have different impacts on the evacuation process. Therefore, dividing the population distribution into different scenarios for analysis can better understand and evaluate the performance of the algorithm.

[0189] The first scenario: The density differences of the population to be evacuated are not significant and the distribution locations are relatively uniform. In this case, there is enough space between the people to move freely, and the impact of the congestion coefficient on the evacuation process is relatively small. The path planning mainly considers the principle of being closest to the exit, and the algorithm can effectively meet the evacuation requirements.

[0190] The second scenario: Among the population to be evacuated, a part has a relatively high density and is relatively concentrated in distribution location, while the density of the other part of the population is relatively low. In this case, it is necessary to consider the congestion situation and optimize the path planning to avoid congestion and queuing. The algorithm not only considers the principle of being closest to the exit, but also needs to be flexibly allocated to select an exit that is relatively close but without congestion for evacuation.

[0191] The third scenario: The distribution locations of the population to be evacuated are relatively uniform, but the population density is relatively high on both sides and relatively sparse in the middle. In this case, it is necessary to reasonably arrange each exit, consider the congestion situation and select an unobstructed path for evacuation. The algorithm flexibly allocates the population according to the actual situation and preferentially selects the relatively sparse exit for evacuation.

[0192] Through the analysis of the above three scenarios, the performance of the algorithm under different population distributions can be comprehensively understood, its adaptability and effectiveness can be evaluated, and the evacuation path planning algorithm can be further improved and optimized to meet different evacuation requirements. The simulation effects of the three scenarios are as follows:

[0193] For the population distribution scenario 1, the density differences of the population to be evacuated are not significant (100 people, 150 people, 100 people, 200 people). From the simulation results Figure 10 It can be seen that in the case of relatively uniform personnel distribution, the algorithm mainly conducts evacuation path planning based on the principle of being closest to the exit. The path finding has good directionality, the path is smooth and straight, and the number of turns is small. In this scenario, the density of the population to be evacuated is not large, and the impact of the congestion coefficient on the evacuation process is not significant, and its role will not be significantly reflected, which is in line with the actual evacuation situation.

[0194] In the population distribution scenario 2 (100 people, 150 people, 100 people, 300 people), the density of the population to be evacuated 4 is relatively high, and the population distribution location is relatively concentrated on the left. From the simulation results Figure 11It can be seen that when the personnel distribution is relatively concentrated, the algorithm does not simply plan the evacuation route according to the principle of the nearest exit. For example, in the evacuation route of the crowd to be evacuated No. 3, it takes into account the congestion at the left exit caused by the crowds to be evacuated No. 1 and No. 2, so it does not choose the nearest left exit for evacuation. Instead, it chooses the relatively near middle exit that does not require queuing for escape, and the role of the congestion coefficient is specifically and effectively realized. At the same time, the overall evacuation route planning also has good directivity. Based on the characteristic that the crowd to be evacuated is distributed on the left, the left exit is preferentially selected for evacuation to better adapt to the actual situation. Such route planning results in smooth paths and fewer turning times. That is, for the situation where the density of people is large and the distribution position is relatively concentrated on the left in the crowd distribution scenario 2, the evacuation algorithm does not simply plan the route according to the nearest exit principle. It takes into account the congestion situation and selects a suitable exit to realize the role of the congestion coefficient. In addition, the algorithm also takes into account the characteristic that the crowd to be evacuated is distributed on the left, and preferentially selects the left exit for evacuation, so that the path is smooth and the number of turning times is less.

[0195] In the crowd distribution scenario 3 (250 people, 100 people, 100 people, 200 people), the distribution positions of the crowds to be evacuated are relatively uniform, and the crowd density shows a situation of sparse in the middle (the densities of the crowds to be evacuated No. 2 and No. 3 are small) and dense on both sides (the densities of the crowds to be evacuated No. 1 and No. 4 are large). From the simulation results Figure 12 It can be seen that when the distribution positions of the crowds to be evacuated are relatively uniform, overall, according to the characteristic of uniform crowd distribution, corresponding crowds are arranged for evacuation at each exit. The algorithm flexibly allocates according to the congestion situation of the exits and plans the evacuation route. For example, in the evacuation route of the crowd to be evacuated No. 2, it takes into account the congestion at the left exit caused by the crowd to be evacuated No. 1 with high density, so it does not choose the nearest left exit for evacuation. Instead, it chooses the relatively near middle exit that is not congested for escape; in the evacuation route of the crowd to be evacuated No. 3, it takes into account the congestion at the right exit caused by the crowd to be evacuated No. 4 with high density, so it does not choose the nearest right exit for evacuation. Instead, it also chooses the relatively near middle exit that is not congested for escape, which conforms to the actual evacuation situation.

[0196] Those of ordinary skill in the art can understand that the drawings are only schematic diagrams of an embodiment, and the processes in the drawings are not necessarily essential for implementing the present invention.

Claims

1. An intelligent evacuation method for urban rail transit station fires based on surveillance videos, characterized in that, It includes the following steps: S1, Fire scene information acquisition: Use object recognition technology to identify fire scene information, use the principles of distortion correction and camera calibration to obtain the positions of fires and obstacles, and construct an evacuation space topology model based on this; S2, Fusion of A* and ant colony algorithms: Use the A* algorithm to adjust the initial pheromone concentration of the ant colony to refine the pathfinding model; S3, Implementation and optimization of the evacuation path fusion algorithm: Introduce end-point orientation, turning penalty, and exit congestion effect factors, and reconstruct the distance heuristic function, state transition probability function, and evacuation path length calculation function to make the evacuation path fusion algorithm have better global regulation and evacuation efficiency; Among them, The specific steps of step S3 include: S31, Algorithm improvement of the end-point orientation effect; including: Based on increasing the calculation of the visibility between the next node and the end point and squaring the reciprocal of the sum of two distances, reconstruct the distance heuristic function; S32, Turning penalty effect; including: By calculating the turning angle and counting the number of turns, consider the turning penalty; Adjust the state transition probability in the ant colony algorithm, add a turning penalty coefficient and count the number of turns to reduce the probability of turning in path search; S33, Exit congestion effect; including: Consider the evacuation optimization idea of the exit congestion effect, and optimize the exit congestion effect; Among them, the optimization idea includes: In the initial stage of evacuation, first target the people closest to the evacuation exit and evacuate them quickly; During the evacuation path planning process, update the crowd distribution density at each evacuation exit crowd_exit, combine the congestion degree of each exit, calculate the congestion coefficient crowd_imp and the distance from the exit to the people to be evacuated way_len_best, comprehensively consider the distribution position of the fire in the subway station, and reasonably plan and flexibly adjust the evacuation path, so as to achieve the rapid and safe evacuation of the entire subway platform crowd in case of fire.

2. A method for intelligent evacuation of urban rail transit station fires based on surveillance video according to claim 1, characterized in that, The specific steps of step S1 include: S11, Fire scene information recognition: Based on object recognition technology, obtain the fire information and personnel information in the fire scene. Among them, the personnel information acquisition process includes: First, obtain video data input using the station camera, perform grid processing on the video data, use a convolutional neural network to establish a prior box for crowd recognition and a target classification probability map, and then perform pedestrian detection, and at the same time control the movement of the camera to perform pedestrian tracking; S12, Target position acquisition: Based on the distortion correction model, obtain the spatial position information of the fire; The target spatial position acquisition process is to first correct the distortion generated during the video imaging process, then use the camera calibration principle to obtain the world coordinates in the world coordinate system calibrated by the camera, and finally construct an evacuation space topology model to calculate the coordinates of the target in the real scene; S13, Construction of the evacuation space topology model: Use a certain location in the real site as the coordinate origin, convert the position of the detected target in the camera calibration coordinates into the position in the real scene coordinates, and accurately express the position of the target in this scene.

3. The intelligent evacuation method for urban rail transit station fires based on surveillance videos according to claim 2, characterized in that, S11 specifically includes: The specific fire scene information is identified as follows: Extract the personal dynamic positions and trajectories from video data for personnel detection and tracking. Select the method of head detection and match it with the high-density crowd scene. Use the YOLO algorithm and the Deep SORT algorithm to establish a crowd head detection + tracking framework; Specifically: Use the head detection method and the YOLO algorithm for crowd detection, and then combine the Deep SORT algorithm to achieve crowd tracking. It can extract individual dynamic information from crowd images, realize accurate personnel detection and tracking. Then, adopt a distortion correction model, adjust the camera parameters, realize the projection display of crowd images, and obtain the spatial position information of the target according to the camera calibration principle. Considering that the movement of people in the video will cause the tracking ID of the same individual to change, it is necessary to calculate the trajectory vector of the pedestrian trajectory in the video to predict the next movement trajectory of the pedestrian. Use the predicted trajectory and the actual trajectory of the pedestrian in the video as inputs, and through the IO matching method, set the person whose feature similarity meets the predefined threshold as the matching person, so as to achieve continuous tracking of the same individual; Input the detected bounding boxes and the corresponding confidence information into the module corresponding to the Deep SORT algorithm. At the next time point, new detection results are input, and it is judged whether the predicted trajectory matches the new detection results; To perform the matching, according to the similarity of the motion situation and appearance features between the predicted Kalman trajectory and the newly detected trajectory, obtain the weighted similarity; Use the Hungarian algorithm, a combinatorial optimization algorithm, and input the weighted similarity to obtain a matching matrix, which helps to improve the robustness weakened by partial occlusion.

4. The intelligent evacuation method for urban rail transit station fires based on surveillance videos according to claim 2, characterized in that S12 specifically includes: S121, distortion correction model; Image distortion will occur during the video imaging process. The coordinates are corrected before converting the fire location coordinates. Considering the tangential distortion and radial distortion of the camera comprehensively, the distortion correction model is obtained as shown in Equation (1): Among them, the left side of the equation is the corrected coordinate; the first half of the right side is the radial distortion model, and the second half of the right side is the tangential distortion model; X W , Y W are the fire location coordinates, X Wcorrected , Y Wcorrected are the corrected coordinates, The parameters k1, k2, k3, p1, p2 are from the calibrated distortion coefficients; S122, inverse calculation of the coordinate system based on the camera calibration principle to convert the image pixel coordinates into world coordinates to obtain the target position; Specifically including: Four coordinate systems are involved in the camera calibration process: the world coordinate system Ow, the camera coordinate system Oc, the image physical coordinate system, and the image pixel coordinate system; For a pinhole camera, two parameter sets are used, namely the camera external internal parameters and the camera internal external parameters, to convert a coordinate in the actual scene into a pixel coordinate in the image frame; Among them, the external parameters realize the conversion from the world coordinate system to the camera coordinate system, and the internal parameters realize the conversion from the camera coordinate to the pixel coordinate system in the image frame; The external parameters include the rotation matrix R and the translation vector t; The rotation matrix describes the rotation information of the camera relative to the three coordinate axes of the world coordinate system; The translation vector describes the translation information of the optical center of the camera relative to the origin of the world coordinate system; Specify the point p in the world coordinate system w (x w , y w , z w ), and perform the conversion using the following formula (2). In the camera coordinate system, it becomes the point p c (x c , y c , z c ); Point p c When projected onto the two-dimensional image plane, it becomes point p(x, y). Then, use Equation (4) to convert the point p in the image coordinate system into pi(u, v) with the image resolution: Among them, f is the camera focal length; dx, dy are the lengths of one pixel; u0, v0 are the optical centers in pixels; By combining and arranging the above equations, the conversion formula from the world coordinate system to the pixel coordinate system can be obtained as shown in Equation (5) below: Among them, the first term on the right side of the formula is the camera internal parameter matrix, and the second term on the right side is the camera external parameter matrix, which is obtained through camera calibration.

5. The intelligent evacuation method for urban rail transit station fires based on surveillance videos according to claim 2, characterized in that S13 specifically includes: There are differences in scale, rotation, and translation between the world coordinate system calibrated by the camera and the coordinate system of the actual scene, which requires conversion between coordinate systems; calculate Formula (6): B = r × A + t (6) where A is the set of known points in the world coordinate system, B is the set of known points in the actual scene, r is the rotation matrix, t is the translation vector, A i and B i are vectors with three rows and one column, expressed as [x y z] T ; the steps to find the optimal rigid transformation matrix are as shown in Equation (7): Bring the two data sets to the origin, then find the rotation matrix r and calculate the translation vector t; obtain the calculation formula (8) of the covariance matrix H: H = (A - centroid A ) × (B - centriod B ) T (8) Then H is the covariance matrix. Using the singular value decomposition algorithm SVD(H) = [U, S, V], let R = VU T ; Considering that the scales of the two coordinate systems are different, the calculation method of the scale multiple λ between them is shown in the formula (reference here): Formula (6) will become: Combined with Formula (10), the rotation matrix r and the translation vector t are respectively: At this point, calculate the actual scene coordinates.

6. The intelligent evacuation method for urban rail transit station fires based on surveillance videos according to claim 1, characterized in that, In step S2, the specific steps include: S21. Establish an evacuation space topology model: Create a topology model that describes the structure of the evacuation area and the distribution of obstacles; S22. Generate a sub-optimal evacuation path using the A* algorithm: Use the A* algorithm to generate a sub-optimal path, considering obstacle avoidance and the shortest path; S23. Adjust the pheromone concentration of topological information elements: According to the sub-optimal path, adjust the pheromone concentration on the topology model; S24. Initialize the parameters of the ant colony algorithm: Set the parameters of the ant colony algorithm, the number of ants, and the pheromone evaporation rate; S25. Start searching for the evacuation path: The ant colony starts to search for the evacuation path according to the pheromone concentration and distance; S26. Reach the end point once and complete the search: When the ant colony reaches the end point once, complete a search S27. Update the pheromone concentration: According to the search results, update the pheromone concentration in the topology model; Usually, the successful path will increase the pheromone concentration, while the failed path will decrease the concentration; S28. Perform iteration: If the maximum number of iterations is not reached, search again; S29. Output the optimal path: When the maximum number of iterations is reached, output the best path as the result; The calculation formula of the evaluation function f(n) of the A* algorithm is shown in (12): f(n) = g(n) + h(n) (12) The Euclidean distance is used to calculate the distance between two points, and the calculation method of the distance between node i and node j is as shown in Equation (13): g(n) is a dissipation function used to calculate the cost from the starting point s to the current node i, as shown in Equation (14): h(n) is a heuristic function used to calculate the cost from the current node i to the target node t, as shown in Equation (15): After the spatial topology model is constructed, according to the corresponding starting and ending points and the selected cost function calculation method, perform sub-optimal path planning using the A* algorithm; In the generated path above, adjust the pheromone concentration τ ij (0), increase the pheromone concentration on the sub-optimal path, and the increment is A ij . After adjusting the initial pheromone concentration τ ij (0), that is, perform the transformation on τ ij (0) as shown in Equation (16): τ ij τ(0) = τ0 + A ij (16) Among them, τ0 is the default value of the initial pheromone concentration, and A ij is the increase value of the initial pheromone concentration in the adjustment area obtained by the A* algorithm.

7. The intelligent evacuation method for urban rail transit station fires based on surveillance videos according to claim 1, wherein The specific steps of step S3 include: S31, Algorithm improvement of the end-point guiding effect: Considering the guiding effect of the end point, on the premise of calculating the visibility between the current node and the next node, the calculation of the visibility d jt between the next node and the end point is added, and the evacuation purpose is enhanced, making the passenger evacuation in the subway station fire more directional; the reciprocal of the sum of the two distances is squared, and this processing can shorten the running time and improve the efficiency of searching for the shortest path in the actual algorithm; based on the above improvement strategies, the distance heuristic function η ij is reconstructed as shown in Equation (17): Among them, i is the current node; j is the next node; t is the end point; Idea of code writing: Define dis_1 as a list containing distance values, and traverse the elements in the dis_1 list through a loop; in the loop, calculate the heuristic function value of each distance value; the calculation method of the heuristic function is to add the distance values together and take the reciprocal of the square of the result as the value of the heuristic function; by processing the distance values, the heuristic function guides ants to choose shorter paths in the ant colony algorithm; after the loop ends, the elements in the dis_1 list are updated to the corresponding heuristic function values; S32, turning penalty effect: Inside the station, there are dense crowds, harsh conditions accompanied by smoke, and the panic psychology of the masses. If the evacuation route is tortuous with many turns, the passengers waiting to be evacuated inside the station will lose their way due to excessive panic, making it more difficult to escape and reducing the evacuation efficiency. Therefore, by calculating the turning angle θ and counting the number of turns tn, considering the turning penalty z, where the larger the turning angle θ, the smaller the z value; when the path is straight and no turning is required, z = 1; as the number of turns tn increases, the p k ij (t) value becomes smaller; adjust the state transition probability p k ij (t) in the ant colony algorithm, add the turning penalty coefficient and count the number of turns to reduce the probability of turning in path search, reduce the complexity of the evacuation path, improve the smoothness of the escape path, and reduce the possibility of passengers losing their way: Among them, θ is the turning angle of the ant when traveling between adjacent nodes; tn is the number of turns of the ant; z is the turning penalty, and its value range is 0.5 to 1; Idea for code writing: Define two empty lists prob and prob_z to store the probabilities of nodes being selected. Next, traverse the elements in the node_be_selected_1 list through a loop. It is a list containing optional nodes. Call the angle function to calculate the turning angle θ to calculate the angle between the current position and the candidate node node_be_selected_1[i], and consider the turning penalty. Next, calculate the selection probability p_z of the node, where the distance weight dis_imp, pheromone concentration weight pher_imp, and number of turns weight z are combined turn_num ; Add the calculated node selection probability p_z to the prob_z list; then, update the node selection probability by dividing each node selection probability by prob_sum; subsequently, compare to obtain the maximum node transition probability; finally, after the loop ends, the selected node is recorded as node_be_selected_1[k]; S33, exit congestion effect: When the passenger flow in the subway station is too large, the exits closer to the passengers are very likely to be congested. The passengers to be evacuated often need to queue up, which undoubtedly increases the evacuation time and reduces the evacuation efficiency; in this case, choosing the exit with the shortest escape route length for evacuation does not necessarily mean that this is the best way to evacuate; in order to meet the basic principles of rapid and safe evacuation, it is necessary and essential to optimize the above fire scene evacuation path model; The evacuation optimization idea considering the exit congestion effect is: in the initial stage of evacuation, first evacuate the people closest to the evacuation exit quickly; during the evacuation path planning process, update the population distribution density of each evacuation exit crowd_exit, combine the congestion degree of each exit, calculate the congestion coefficient crowd_imp and the distance way_len_best from the exit to the people to be evacuated, and comprehensively consider the distribution position of the fire in the subway station, reasonably plan and flexibly adjust the evacuation path, so as to achieve the rapid and safe evacuation of the entire subway platform crowd in case of fire; Specific implementation steps for optimizing the exit congestion effect: After projecting the personnel information identified by the YOLO algorithm onto the spatial topology model, according to the different crowds to be evacuated, respectively extract the distribution center points of the crowds and generate a two-dimensional list peop related to the crowds to be evacuated. The parameter peop_posi in the list is the position coordinates of the center point of a certain batch of crowds to be evacuated, and the parameter peop_density is the density of a certain batch of crowds to be evacuated; after evacuating the first batch of crowds without congestion effect, the number of the first batch of passengers to be evacuated is recorded as the population density of the evacuation exit crowd_exit. By introducing the congestion coefficient crowd_imp of the corresponding exit, recalculate the distance from the second batch of crowds to be evacuated to the exit. After the evacuation of this batch of passengers is completed, continue to update the platform population density and the exit congestion coefficient to complete the evacuation of subsequent batches of passengers.

Images