Distributed Crowd Evacuation Route Planning Method and System Based on Federated Learning

Abstract

The present disclosure belongs to the technical field of crowd evacuation path planning, and provides a distributed crowd evacuation path planning method and system based on federated learning, including the following steps: obtaining video information data of sub-regions to construct a local potential field; calculating local potential values based on the constructed local potential field to construct a local density field; determining the collision-free movement speed of individuals according to the constructed local potential field and local density field; aggregating the local potential fields of each constructed sub-region to construct a global potential field; building a crowd evacuation simulation platform according to the constructed local potential field and global potential field, and performing simulation of crowd evacuation based on the obtained collision-free movement speed of individuals.

Description

Technical Field

[0001] The present disclosure belongs to the technical field of crowd evacuation path planning, and particularly relates to a distributed crowd evacuation path planning method and system based on federated learning. Background Art

[0002] The statements in this section merely provide background technical information related to the present disclosure and do not necessarily constitute prior art.

[0003] With the rapid development of social economy, the population base is getting larger and people's travel is more frequent. Especially in some relatively crowded places, such as schools, railway stations, shopping malls, etc., minor disturbances within the crowd will have a great impact on the crowd evacuation efficiency, posing a great safety hazard. Especially in large-scale scenarios with multiple regions that do not interact with each other, once an emergency occurs, it is easier to cause congestion and reduce the crowd evacuation efficiency. If the crowd cannot be planned and effectively controlled in advance, serious losses will be caused. Therefore, in recent years, group evacuation and computer simulation technologies have attracted extensive attention from the industrial and academic circles.

[0004] In the research process, the inventors found the following technical problems in the prior art:

[0005] Traditional crowd evacuation simulation methods are divided into two types: macroscopic methods and microscopic methods. In macroscopic methods, the most typical ones are the hydrodynamic model and the potential field-based model. In microscopic methods, the most typical method is the social force model-based method. Although these models can effectively achieve crowd simulation, the conditions are all assumptions, reducing the authenticity of crowd simulation. To improve the authenticity of crowd evacuation simulation, data-driven methods have emerged. The crowd evacuation simulation method that combines data-driven and machine learning methods can not only significantly improve the authenticity of crowd evacuation simulation, but also provide a path planning function for evacuation individuals to improve the crowd evacuation efficiency. Generally speaking, these methods can well simulate the trajectories existing in the video and improve the authenticity of crowd evacuation simulation. However, these all belong to end-to-end learning and have a high dependence on data. The video data captured by numerous cameras deployed in large-scale places can help with crowd evacuation in case of emergency. However, most surveillance videos are not allowed to be shared to avoid privacy leakage. Summary of the Invention

[0006] To solve the above problems, the present disclosure proposes a distributed crowd evacuation path planning method and system based on federated learning, which performs path planning based on the video data of the entire scene, provides a global path planning for the evacuated crowd while ensuring privacy protection, and enhances the effectiveness of path planning.

[0007] According to some embodiments, the first solution of the present disclosure provides a distributed crowd evacuation path planning method based on federated learning, adopting the following technical solutions:

[0008] A distributed crowd evacuation path planning method based on federated learning includes the following steps:

[0009] Obtain video information data of sub-regions and construct a local potential field;

[0010] Calculate local potential values based on the constructed local potential field and construct a local density field;

[0011] Determine the collision-free movement speed of individuals based on the constructed local potential field and local density field;

[0012] Aggregate the local potential fields of each constructed sub-region to construct a global potential field;

[0013] Build a crowd evacuation simulation platform according to the constructed local potential field and global potential field, and perform simulation of crowd evacuation based on the obtained collision-free movement speed of individuals.

[0014] As a further technical limitation, the video information data includes at least scene information and crowd information.

[0015] As a further technical limitation, the process of constructing the local potential field is as follows:

[0016] Divide the sub-region evenly into several grid cells, number and mark the position coordinates of each obtained grid cell;

[0017] Construct the local area scene based on the grid cells with marked position coordinates, and construct the local potential field in combination with the potential field.

[0018] As a further technical limitation, the local potential values include at least the scene discomfort value, the final potential value affected by all exits, and the total crowd discomfort value affected by all individuals.

[0019] As a further technical limitation, the collision-free movement speed of individuals includes the speed magnitude and the speed direction; the local density field affects the speed magnitude, and the local potential field determines the speed direction.

[0020] As a further technical limitation, in the process of constructing the global potential field, a distributed framework of federated learning is adopted, the mutual guardianship between sub-regions is considered, the modeling of emergency evacuation planning is carried out based on the local potential field, and the global potential field is obtained through global aggregation to plan the global optimal path for individuals.

[0021] Further, in the process of obtaining the global potential field through global aggregation, the object position information uploaded by each sub-region is concatenated, and the position information uploaded by each sub-region is converted to global coordinates to obtain global scene information, and a global potential field is constructed.

[0022] According to some embodiments, the second solution of the present disclosure provides a distributed crowd evacuation path planning system based on federated learning, adopting the following technical solutions:

[0023] A distributed crowd evacuation path planning system based on federated learning, comprising:

[0024] A local potential field construction module, configured to obtain video information data of a sub-region and construct a local potential field;

[0025] A local density field construction module, configured to calculate local potential values based on the constructed local potential field and construct a local density field;

[0026] An individual velocity determination module, configured to determine the collision-free movement velocity of an individual according to the constructed local potential field and local density field;

[0027] A global potential field construction module, configured to aggregate the constructed local potential fields of each sub-region to construct a global potential field;

[0028] A crowd evacuation simulation module, configured to build a crowd evacuation simulation platform according to the constructed local potential field and global potential field, and perform simulation of crowd evacuation based on the obtained collision-free movement velocity of the individual.

[0029] According to some embodiments, the third solution of the present disclosure provides a computer-readable storage medium, adopting the following technical solutions:

[0030] A computer-readable storage medium, on which a program is stored, and when the program is executed by a processor, it implements the steps in the distributed crowd evacuation path planning method based on federated learning as described in the first aspect of the present disclosure.

[0031] According to some embodiments, the fourth solution of the present disclosure provides an electronic device, adopting the following technical solutions:

[0032] An electronic device, comprising a memory, a processor, and a program stored on the memory and executable on the processor, and when the processor executes the program, it implements the steps in the distributed crowd evacuation path planning method based on federated learning as described in the first aspect of the present disclosure.

[0033] Compared with the prior art, the beneficial effects of the present disclosure are:

[0034] (1) The present disclosure constructs a local potential field model for planning local evacuation paths for a crowd. First, a scene information model is constructed, and then the movement of the crowd is incorporated to obtain potential field information in combination with the scene information for modeling sub-regions.

[0035] (2) The present disclosure constructs a global potential field model for planning a globally optimal evacuation path for a crowd. This model uses the global aggregation step in federated learning to integrate the potential field information from each sub-region to construct a global potential field to guide the evacuation of the crowd.

[0036] (3) The present disclosure proposes a method for crowd evacuation path planning based on federated learning, and introduces the potential field model into the federated learning framework. By uploading the potential field information of each local region instead of video information, privacy protection is achieved.

[0037] (4) The present disclosure constructs a simulation and visualization platform to simulate the crowd movement process based on federated learning, and more realistically demonstrates the simulation effect. Brief Description of the Drawings

[0038] The accompanying drawings forming a part of this disclosure are used to provide a further understanding of the present disclosure. The schematic embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation to the present disclosure.

[0039] Figure 1 is a flowchart of the distributed crowd evacuation path planning method based on federated learning in Embodiment 1 of the present disclosure;

[0040] Figure 2(a) is a schematic diagram of the local potential field network result in Embodiment 1 of the present disclosure;

[0041] Figure 2(b) is a schematic diagram of the local potential field threshold in Embodiment 1 of the present disclosure;

[0042] Figure 3 is a schematic diagram of the crowd density grid in Embodiment 1 of the present disclosure;

[0043] Figure 4 is a density-velocity relationship diagram in Embodiment 1 of the present disclosure;

[0044] Figure 5(a) is a schematic diagram of the real scene in Embodiment 1 of the present disclosure;

[0045] Figure 5(b) is a diagram demonstrating the potential energy value of the static potential field in Embodiment 1 of the present disclosure;

[0046] Figure 6(a) is a diagram demonstrating the potential energy value of the local dynamic potential field at t = 0 in Embodiment 1 of the present disclosure;

[0047] Figure 6(b) is a diagram demonstrating the potential energy value of the local dynamic potential field at t = 10 in Embodiment 1 of the present disclosure;

[0048] Figure 6(c) is a demonstration diagram of the potential energy value of the local dynamic potential field at t = 20 in the first embodiment of the present disclosure;

[0049] Figure 6(d) is a demonstration diagram of the potential energy value of the local dynamic potential field at t = 30 in the first embodiment of the present disclosure;

[0050] Figure 7(a) is a demonstration diagram of the potential energy value of the global potential field at t = 0 in the first embodiment of the present disclosure;

[0051] Figure 7(b) is a demonstration diagram of the potential energy value of the global potential field at t = 10 in the first embodiment of the present disclosure;

[0052] Figure 7(c) is a demonstration diagram of the potential energy value of the global potential field at t = 20 in the first embodiment of the present disclosure;

[0053] Figure 7(d) is a demonstration diagram of the potential energy value of the global potential field at t = 30 in the first embodiment of the present disclosure;

[0054] Figure 8(a) is a demonstration diagram of the potential energy value of the global potential field in the case of multiple regions at t = 0 in the first embodiment of the present disclosure;

[0055] Figure 8(b) is a demonstration diagram of the potential energy value of the global potential field in the case of multiple regions at t = 10 in the first embodiment of the present disclosure;

[0056] Figure 8(c) is a demonstration diagram of the potential energy value of the global potential field in the case of multiple regions at t = 20 in the first embodiment of the present disclosure;

[0057] Figure 8(d) is a demonstration diagram of the potential energy value of the global potential field in the case of multiple regions at t = 30 in the first embodiment of the present disclosure;

[0058] Figure 9 is a structural block diagram of a distributed crowd evacuation path planning system based on federated learning in the first embodiment of the present disclosure. Detailed implementation manners

[0059] The present disclosure will be further described below in conjunction with the accompanying drawings and embodiments.

[0060] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanations of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present disclosure belongs.

[0061] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0062] Without conflict, the embodiments in the present disclosure and the features in the embodiments can be combined with each other.

[0063] Embodiment 1

[0064] Embodiment 1 of the present disclosure introduces a distributed crowd evacuation path planning method based on federated learning.

[0065] As Figure 1 shown, a distributed crowd evacuation path planning method based on federated learning includes the following steps:

[0066] Step S01: Obtain video information data of sub-regions and construct a local potential field;

[0067] Step S02: Calculate local potential values based on the constructed local potential field and construct a local density field;

[0068] Step S03: Determine the collision-free movement speed of individuals according to the constructed local potential field and local density field;

[0069] Step S04: Aggregate the local potential fields of each constructed sub-region to construct a global potential field;

[0070] Step S05: Build a crowd evacuation simulation platform according to the constructed local potential field and global potential field, and perform simulation of crowd evacuation according to the obtained collision-free movement speed of individuals.

[0071] As one or more embodiments, in step S01, the specific process of constructing the local potential field is as follows:

[0072] Step S101: Model each sub-region according to the potential field and evenly divide it into a number of grid cells according to the size of different sub-regions;

[0073] Step S102: Each grid cell has a corresponding number (i, j) (for example, the number of the first grid cell in the first row is (1, 1)), and also has a corresponding position coordinate (x, y). Usually, the position coordinate of the grid cell is selected as the central position coordinate of the grid cell, as shown in Fig. 2(a).

[0074] Step S103: As shown in Fig. 2(b), the constructed local area scene includes individuals, obstacles, hazard sources, and exits. Since the exits generate attractive potential energy for all individuals within the area, all individuals in the area tend to move towards the exits. However, the influence ranges of hazard sources, obstacles, and individuals are limited. Therefore, we use γ, η, and μ to represent the influence distance thresholds of hazard sources, obstacles, and individuals respectively. Only when the distance between an individual and them is less than or equal to their corresponding thresholds will they have an impact on the individual. Otherwise, there will be no impact.

[0075] As one or more implementation manners, in step S02, first calculate the local potential energy value.

[0076] (1) Scene discomfort

[0077] During the process of individuals moving towards the exits, it is necessary to avoid hazard sources and obstacles as much as possible. We represent this process by establishing a scene discomfort g.

[0078] 1) Discomfort value generated by hazard sources:

[0079] Suppose there are p hazard sources D (D 1 , …, D a , …, D p ) in this sub - area. The influence of the discomfort value generated by the hazard sources on each grid cell is inversely proportional to the distance between them. The discomfort value function is as follows:

[0080]

[0081] Among them, m is the discomfort coefficient of the hazard source, l((i, j), D a ) is the Euclidean distance between the center position of the grid cell (i, j) and the hazard source D a , and γ is the influence distance threshold of the hazard source.

[0082] The discomfort value of each grid cell is affected by multiple hazard sources. The total discomfort value generated by all hazard sources in the area for each grid cell can be expressed as:

[0083]

[0084] Among them, g dangerous (i, j) represents the total discomfort value of the grid cell (i, j) affected by all hazard sources.

[0085] 2) Discomfort value generated by obstacles:

[0086] Suppose there are q obstacles O (O 1 , …, O b , …, O q), the influence of the obstacle on the discomfort value generated by each grid cell is inversely proportional to its distance, and the discomfort value function is as follows:

[0087]

[0088] where n is the discomfort coefficient of the obstacle, and l((i,j),O b ) is the Euclidean distance between the center position of the grid cell (i,j) and the obstacle O b , and η is the obstacle influence distance threshold.

[0089] The discomfort value of each grid cell is affected by multiple obstacles. The total discomfort value generated by all obstacles in the area for each grid cell can be expressed as:

[0090]

[0091] where g obstacle (i,j) represents the total discomfort value of the grid cell (i,j) affected by all obstacles.

[0092] 3) Total discomfort value:

[0093] Both obstacles and hazard sources will generate discomfort values for individuals. However, compared with obstacles, when the distance from an individual is the same, the hazard source generates a higher discomfort value than the obstacle. The total discomfort value of each grid cell can be expressed as:

[0094] g total (i,j) = α·g dangerous (i,j) + β·g obstacle (i,j)

[0095] where α and β are the weights of the discomfort values generated by the hazard source and the obstacle for individuals, respectively, and α > β.

[0096] (2) Distance field

[0097] The exit will generate an attractive potential field for individuals, attracting them to move towards the exit. We establish a distance field d to represent this process.

[0098] Assume that there are r exits E(E 1 , …, E c , …, E r ) in this sub-region. The influence of the exit on the potential energy value generated by each grid cell is proportional to the distance between them. The potential energy value function is as follows:

[0099]

[0100] where k is the attraction coefficient, and l((i,j),E c ) is the Euclidean distance between the center position of the grid cell (i,j) and the exit Ec Euclidean distance

[0101] However, in general, individuals will always choose the exit closest to themselves for evacuation. We can choose to define the final potential energy value generated by all exits in the area for each grid cell by taking the minimum potential energy value:

[0102]

[0103] where d total (i,j) represents the final potential energy value of the grid cell (i,j) affected by all exits.

[0104] (3) Crowd discomfort field

[0105] During the movement process, individuals will change their movement trajectories by perceiving the crowd discomfort field, so as to avoid areas with higher discomfort levels and move towards areas with lower discomfort levels. We represent this process by establishing a crowd discomfort field c.

[0106] We stipulate an influence threshold μ, and the target individual will only contribute to the crowd discomfort value within the circle centered at its position with a radius of μ.

[0107] Suppose there are s individuals I (I 1 ,…,I d ,…,I s ) in the grid cell. The crowd discomfort value generated by any individual for the grid cell is inversely proportional to the distance between it and the center coordinates of the grid cell. The contribution of individual I d to the crowd discomfort value of the grid cell (i,j) can be expressed as:

[0108]

[0109] where u is the crowd discomfort coefficient, and l((i,j),I d ) is the Euclidean distance between the center position of the grid cell (i,j) and individual I d , and μ is the individual influence distance threshold.

[0110] The total crowd discomfort value of the grid cell affected by all individuals within the threshold range can be expressed as:

[0111]

[0112] where c total (i,j) represents the total crowd discomfort value of the grid cell (i,j) affected by all individuals, and s is the total number of individuals in the grid cell.

[0113] Therefore, the potential energy value can be expressed as:

[0114] P(i,j) = gtotal (i, j) + d total (i, j) + c total (i, j)

[0115] Among them, g total (i, j) is the scene discomfort value, d total (i, j) is the final potential energy value affected by all exits, c total (i, j) represents the total crowd discomfort value affected by all individuals.

[0116] As one or more implementation manners, in step S02, it is also necessary to construct a local density field.

[0117] Convert each individual (including obstacles and hazard sources) into a separate density field, which reaches a peak at the location of the individual and then decreases radially. The specific form of this function is not important as long as it is not less than outside the boundary disk and not greater than inside the boundary disk with a radius of r. When establishing the density field, "tile" the crowd particles in the area onto the density grid, and the density contribution value of an individual at any position to the grid cell where it is located can be calculated, as Figure 3 shown.

[0118] As Figure 3 shown, the density contribution value of this individual in the C grid cell is added to other adjacent grids according to the following formula:

[0119] ρ A = min(1 - Δx, 1 - Δy) λ ρ B = min(Δx, 1 - Δy) λ

[0120] ρ C = min(Δx, Δy) λ ρ D = min(1 - Δx, Δy) λ

[0121] Among them, λ is the density exponent, which determines the speed of density attenuation.

[0122] From this, the density field generated by each individual can be constructed, and the density contribution value of any individual to the grid cell where it is located and its adjacent grids can be calculated.

[0123] The crowd density ρ of each grid cell is the sum of the density values contributed by all individuals in this grid cell and its adjacent grid cells, and the formula is as follows:

[0124]

[0125] Among them, there are a total of m individuals (including hazard sources and obstacles) in the grid cell and its adjacent grid cells.

[0126] As one or more embodiments, in step S03, the collision-free movement speed of the individual is determined.

[0127] (1) Direction of individual movement speed

[0128] The potential energy field determines the speed direction. Once the potential energy field is constructed, the potential energy value of any grid cell can be calculated:

[0129]

[0130] Among them, is the potential energy value gradient at (x, y), is the magnitude of the potential energy value gradient at (x, y).

[0131] In summary, the potential energy value at any position of the grid cell is obtained according to quadratic linear interpolation, so as to determine the movement direction of the individual at any position.

[0132] (2) Magnitude of individual movement speed

[0133] The density field affects the speed magnitude. Armin Seyfried et al. have further studied the Fundamental Diagram of Pedestrian on the basis of previous work and given the required density-speed correspondence diagram, as Figure 4 shown.

[0134] According to Figure 4 analysis, when the density ρ < 0.7, the individual can move at the desired speed; when the density ρ > 5.1, people can hardly move forward. When the crowd density is between 0.46 / m 2 and 5.1 / m 2 the evacuation speed of the individual satisfies the formula:

[0135] V = k - akρ

[0136] Among them, V is the evacuation speed of the individual, k is a parameter and k is a random number between [-1, 1], ρ is the crowd density, and a is a coefficient and a = 0.189.

[0137] The above steps solve the evacuation path planning problem of each individual in the local area. The local potential energy field is constructed according to the positions of the crowd, obstacles and hazard sources extracted from the real video data, so as to determine a local optimal evacuation path for each individual.

[0138] As one or more embodiments, in step S04, the global optimal path planning based on the dynamic potential field is performed.

[0139] (1) Global Aggregation Model Update Method

[0140] The local potential field information consists of local scene information and crowd information. The central server collects the potential field information from each local area through the global aggregation step to construct a global potential field, thereby planning the global optimal evacuation path for the crowd.

[0141] Due to the limitation of the camera coverage, the cameras in each sub-region can only collect the video information in their own regions. In case of an emergency, if only the situation in its own region is considered while ignoring the congestion in other sub-regions, it is very likely to cause greater congestion, thus affecting the crowd evacuation efficiency. Therefore, referring to the distributed framework of federated learning, considering the interaction between each sub-region, the potential field method is used to model the large-scale emergency evacuation planning scheme, and the global optimal path is planned for individuals by performing the global aggregation step.

[0142] Through the real video data captured by the cameras in each sub-region, the scene information and crowd information in the region can be obtained. According to the scene information, the position coordinates of obstacles, hazard sources, and exits in the region are extracted respectively. Since different objects in the region have different effects on the potential energy value, the coordinates of different types of objects are marked.

[0143] For example, the position coordinates of obstacles, hazard sources, and exits in sub-region 1 are respectively recorded as According to the crowd information, the position coordinates of each individual in each region are extracted. For example, the position coordinate of individual i in sub-region 1 is recorded as P i 1 .

[0144] (2) Global Coordinate Transformation Method

[0145] When performing the global aggregation step, the uploaded parameters are the position coordinates of obstacles, hazard sources, and exits in each sub-region. Through global aggregation, the object position information uploaded by each sub-region is spliced together. However, the position information uploaded by each sub-region is only its relative position coordinate, and global coordinate transformation is required to obtain the global scene information, thereby constructing the global potential field.

[0146] Given the size of the entire scene area and dividing the entire scene evenly into n rectangular sub-regions, the length and width of each sub-region are a and b respectively, and the lower left corner of the entire scene is taken as the origin. From this, the position coordinates of any individual in any region in the global scene can be determined.

[0147] After obtaining the global scene information, a global potential field is constructed. The constructed global potential field is sent to each sub-region for the next round of iteration. Each sub-region will re-plan the evacuation route for individuals according to the updated global potential field situation.

[0148] Since the movement of the crowd is dynamically changing, the potential energy value of each grid cell in the global potential field is also constantly changing. By each sub-region uploading the position coordinates of the crowd and obstacles to the central server in real time, and then the central server performs global aggregation, a dynamic global potential field is constructed to ensure that the evacuation route of individuals is optimal at each moment. At the same time, based on the potential field modeling, individuals automatically move towards the position with a lower potential energy value, thus realizing the collision avoidance function.

[0149] First, construct its corresponding static potential field through the given scene information, then add crowd movement, demonstrate the change of the potential field during the crowd movement process, and verify whether the collision avoidance function can be realized and the crowd can be driven to the destination in our improved potential field. Then construct the global potential field through the spliced global scene information to guide the crowd evacuation, and verify whether individuals can find the optimal evacuation route under the influence of the global potential field.

[0150] As one or more embodiments, in step S05, a simulation of crowd evacuation is performed.

[0151] (1) Demonstration of crowd movement in the local potential field

[0152] 1.1) Local static potential field

[0153] To verify that the distributed crowd evacuation route planning method based on federated learning proposed in this embodiment can enable the crowd to realize the collision avoidance function and drive the crowd to evacuate towards the target point. First, extract the positions of obstacles, hazard sources, and exits from the real video data to construct a static potential field. The experimental results are shown in Figure 5. Figure 5(a) is a schematic diagram of the real scene of the static potential field, and Figure 5(b) is a schematic diagram of the static potential field.

[0154] 1.2) Local dynamic potential field

[0155] After the static potential field is constructed, according to the crowd information collected from the real video data, crowd movement is added. To verify whether individuals can avoid obstacles as expected and evacuate towards the exit position during the crowd movement process, our experiment simulated the movement of 30 people. The experimental results are shown in Figures 6(a), 6(b), 6(c), and 6(d) respectively.

[0156] (2) Demonstration of crowd movement in the global potential field

[0157] 2.1) Comparative experiment between the global potential field and the local potential field

[0158] Each sub-region uploads its respective scenario information and population information to the central server. The central server integrates the information to complete a global aggregation step, constructs a global potential field to guide the evacuation of the crowd. To verify whether the optimal evacuation path can be found through the global aggregation step, a comparative experiment was first conducted. That is, whether the evacuation path of individuals in the sub-region would change if they considered the global potential field. The experimental results are shown in Figures 7(a), 7(b), 7(c), and 7(d). The experimental results prove that under the influence of the global potential field, the crowd will re-select the current optimal path for evacuation.

[0159] (3) Demonstration of crowd movement in the global potential field in the case of multiple regions

[0160] Considering the situation where there are multiple total exits in the evacuation area, whether the crowd will automatically select the optimal evacuation path according to the global potential field. To verify that our framework can be applied to different situations, we simulated a three-region scenario, set the positions of the total exits. The experimental results are shown in Figures 8(a), 8(b), 8(c), and 8(d). The distributed data-driven crowd evacuation framework based on federated learning we proposed can flexibly adapt to the dynamic changes of the scenario.

[0161] In this embodiment, in order to find the global optimal path for evacuation individuals and guide the crowd evacuation on the premise of ensuring the privacy of camera users, the potential field model is introduced into the federated learning framework. By uploading the potential field information of each local region to the central server instead of uploading video information, the problem of privacy protection of each camera user during the interaction is solved; the central server integrates the potential field information of each local region through the global aggregation step, thereby constructing a global potential field to find the global optimal evacuation path for each evacuation user.

[0162] Embodiment Two

[0163] Embodiment Two of the present disclosure introduces a distributed crowd evacuation path planning system based on federated learning.

[0164] As Figure 9 shown, a distributed crowd evacuation path planning system based on federated learning includes:

[0165] A local potential field construction module, which is configured to obtain video information data of a sub-region and construct a local potential field;

[0166] A local density field construction module, which is configured to calculate local potential values based on the constructed local potential field and construct a local density field;

[0167] An individual speed determination module, which is configured to determine the collision-free movement speed of an individual based on the constructed local potential field and local density field;

[0168] A global potential energy field construction module, which is configured to aggregate the local potential energy fields of the constructed sub-regions to construct a global potential energy field;

[0169] A crowd evacuation simulation module, which is configured to build a crowd evacuation simulation platform according to the constructed local potential energy field and global potential energy field, and perform simulation of crowd evacuation based on the collision-free movement speed of the obtained individuals.

[0170] The detailed steps are the same as those of the distributed crowd evacuation path planning method based on federated learning provided in the first embodiment, and will not be described in detail here.

[0171] Embodiment III

[0172] Embodiment III of the present disclosure provides a computer-readable storage medium.

[0173] A computer-readable storage medium, on which a program is stored, and when the program is executed by a processor, the steps in the distributed crowd evacuation path planning method based on federated learning as described in the first embodiment of the present disclosure are implemented.

[0174] The detailed steps are the same as those of the distributed crowd evacuation path planning method based on federated learning provided in the first embodiment, and will not be described in detail here.

[0175] Embodiment IV

[0176] Embodiment IV of the present disclosure provides an electronic device.

[0177] An electronic device, including a memory, a processor, and a program stored on the memory and executable on the processor, and when the processor executes the program, the steps in the distributed crowd evacuation path planning method based on federated learning as described in the first embodiment of the present disclosure are implemented.

[0178] The detailed steps are the same as those of the distributed crowd evacuation path planning method based on federated learning provided in the first embodiment, and will not be described in detail here.

[0179] Although the specific implementation manners of the present disclosure are described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that, based on the technical solutions of the present disclosure, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present disclosure.

Claims

1. A distributed crowd evacuation path planning method based on federated learning, characterized in that, it includes the following steps: Obtain video information data of sub-regions and construct a local potential field; Calculate local potential values based on the constructed local potential field and construct a local density field; The local potential values at least include scene discomfort values, the final potential values after being affected by all exits, and the total crowd discomfort values after being affected by all individuals; among them, the scene discomfort values include the discomfort values generated by obstacles and dangerous sources to individuals; select the smallest potential value to define the final potential values generated by all exits in the region for each grid cell; the total crowd discomfort value refers to the crowd discomfort value of the grid cells affected by all individuals within the threshold range; Determine the collision-free movement speed of individuals according to the constructed local potential field and local density field; Aggregate the local potential fields of each constructed sub-region to construct a global potential field; During the construction of the global potential field, adopt a distributed framework of federated learning, consider the mutual guardianship between sub-regions, model emergency evacuation planning based on the local potential field, and obtain the global potential field through global aggregation to plan the global optimal path for individuals; Build a crowd evacuation simulation platform according to the constructed local potential field and global potential field, and perform simulation of crowd evacuation according to the obtained collision-free movement speed of individuals.

2. A distributed crowd evacuation path planning method based on federated learning as described in claim 1, characterized in that, the video information data at least includes scene information and crowd information.

3. A distributed crowd evacuation path planning method based on federated learning as described in claim 1, characterized in that, the process of constructing the local potential field is: Divide the sub-region evenly into several grid cells, number and mark the position coordinates of each obtained grid cell; Construct the local area scene based on the grid cells with marked position coordinates, and construct the local potential field in combination with the potential field.

4. A distributed crowd evacuation path planning method based on federated learning as described in claim 1, characterized in that, the collision-free movement speed of the individual includes the speed magnitude and the speed direction; the local density field affects the speed magnitude, and the local potential field determines the speed direction.

5. A distributed crowd evacuation path planning method based on federated learning as described in claim 1, characterized in that, During the process of obtaining the global potential field through global aggregation, splice the object position information uploaded by each sub-region, perform global coordinate conversion on the position information uploaded by each sub-region to obtain global scene information, and construct the global potential field.

6. A distributed crowd evacuation path planning system based on federated learning, characterized in that, it includes: A local potential field construction module, which is configured to obtain video information data of sub-regions and construct a local potential field; A local density field construction module, which is configured to calculate local potential values based on the constructed local potential field and construct a local density field; The local potential value at least includes a scene discomfort value, a final potential value after being affected by all exits, and a total population discomfort value after being affected by all individuals; wherein, the scene discomfort value includes the discomfort value generated by obstacles and hazard sources to individuals; the final potential value generated by all exits in the area for each grid cell is defined by selecting the minimum potential value; the total population discomfort value refers to the population discomfort value of the grid cells affected by all individuals within the threshold range; An individual speed determination module, which is configured to determine the collision-free movement speed of an individual based on the constructed local potential field and local density field; A global potential field construction module, which is configured to aggregate the local potential fields of the constructed sub-regions to construct a global potential field; During the construction of the global potential field, a distributed framework of federated learning is adopted, considering the mutual guardianship between sub-regions, modeling emergency evacuation planning based on the local potential field, and obtaining the global potential field through global aggregation to plan the global optimal path for individuals; A crowd evacuation simulation module, which is configured to build a crowd evacuation simulation platform according to the constructed local potential field and global potential field, and perform simulation of crowd evacuation based on the obtained collision-free movement speed of individuals.

7. A computer-readable storage medium, on which a program is stored, characterized in that, when the program is executed by a processor, it implements the steps in the distributed crowd evacuation path planning method based on federated learning as described in any one of claims 1-5.

8. An electronic device, including a memory, a processor, and a program stored on the memory and executable on the processor, characterized in that, when the processor executes the program, it implements the steps in the distributed crowd evacuation path planning method based on federated learning as described in any one of claims 1-5.

Images