Emergency Site Data Sensing System Based on Cloud-Edge-Terminal Processing

Abstract

The present invention discloses an emergency site data perception system based on cloud-edge-end processing, which relates to the technical field of emergency data perception. It includes an identification model construction unit, an object box detection unit, a global position analysis unit, an impact degree judgment unit, an evacuation path generation unit, and an emergency processing unit. In the present invention, the global position analysis unit analyzes the local position information contained in the image data frames corresponding to each sub-region in the edge server, and uses a coordinate optimization algorithm in the cloud server to convert the local position information involved in each sub-region to obtain the global position information within the specified scenario. Such a design can, on the one hand, obtain the accurate position of the target, on the other hand, greatly improve the computing speed of the system, and reduce the requirements for the server performance. The impact degree judgment unit and the evacuation path generation unit are used to calculate the optimal evacuation path for each group of people at the emergency site, avoiding congestion when people evacuate.

Description

Technical Field

[0001] The present invention relates to the technical field of emergency data perception, and specifically to an emergency site data perception system based on cloud-edge-end processing. Background Art

[0002] The cloud-edge-end architecture refers to a distributed computing architecture, including three parts: cloud, edge, and end. The emergency site data perception system organically combines functions such as information perception, command and dispatch, resource allocation, monitoring and early warning, etc., to improve the response speed and disposal efficiency of emergency events, and reduce the losses caused by the events. In the invention patent with the application number 202310361774.6, "An emergency management method based on dynamic perception and fusion of disaster site information" is disclosed. Disaster site emergency management refers to the management activities of organizing and coordinating various emergency resources and implementing emergency rescue and post-disaster recovery work when emergencies such as natural disasters and accident disasters occur, which helps to improve the efficiency of rescue and recovery at the disaster site and minimize the casualties and property losses caused by the disaster. However, the current limited sensor technology for disaster site information perception and the limited resources of monitoring equipment, combined with the shortage of personnel at the disaster site, result in low efficiency of disaster site information collection and processing, making it difficult to meet the needs of disaster emergency management. Therefore, the present invention proposes an emergency management method based on dynamic perception and fusion of disaster site information, which comprehensively uses technologies such as the Internet of Things, big data, and artificial intelligence to perform real-time, dynamic, and interactive perception and fusion of data to generate a disaster site situation perception map, and conducts data analysis and timely early warning based on the disaster site situation perception map, which helps to achieve precise disaster disposal, intelligent rescue command, and efficient dispatching work.

[0003] The above-mentioned existing technology solves problems such as the inability to obtain key information in the disaster area in a timely manner. However, during the operation of the system, due to the lack of a target recognition model, it is impossible to obtain the position information of people, obstacles, dangerous objects, and exit signs in a specified scene, resulting in the system being unable to provide a passage plan for the people in the scene in a short time, leading to congestion of a large number of people during the evacuation process, and being unable to determine the urgency of each dangerous object. Therefore, the dispatching of professional personnel requires a long time. Summary of the Invention

[0004] The purpose of the present invention is to provide an emergency site data perception system based on cloud-edge-end processing to solve the problems raised in the above background art.

[0005] To achieve the above purpose, the present invention provides the following technical solution: An emergency site data perception system based on cloud-edge-end processing, including an evacuation path generation unit and an emergency processing unit;

[0006] An identification model construction unit, which obtains a sample data set of crowds, obstacles, dangerous objects, and exit signs. Since each sample information in the data set contains multiple image data frames, the pixel point analysis algorithm is used to calculate the gradient magnitude and direction of each pixel point in the image data frame, and the gradient feature vector of the image data frame is determined according to these data. The color feature vector is fused with the gradient feature vector, and the kernel correlation filter and the regression model are used to construct the target identification model. All the fusion results contained in each sample information are input into the target identification model for analysis to obtain the response map corresponding to each frame;

[0007] A target box detection unit, which scans the image information contained in the frame according to the peak position of each frame's response map and the size of the target box in the previous frame, outputs the position and size of the target box corresponding to the current image data frame, and determines the similarity between the position and size of the target box of the image data frame contained in each sample information and the position and size of the identification box. If the similarity is lower than the preset value, the peak value of the response map corresponding to the image data frame is recalculated and the detection parameters are updated. Otherwise, the similarity between the position and size of the target box in the current frame and the previous frame is calculated;

[0008] A global position analysis unit, which divides the specified scene into multiple sub-regions, uses a camera to take pictures of the sub-regions, transmits the obtained video data to the edge server, sorts the data frames in each video data according to the time sequence, and stores them in the sampling set. The target identification model is used to analyze the local position information of crowds, obstacles, dangerous objects, and exits contained in the image data frame corresponding to each sub-region, and the coordinate optimization algorithm is used to convert the local position information involved in each sub-region to obtain the global position information of each crowd, obstacle, dangerous object, and exit in the specified scene;

[0009] An influence degree judgment unit, which divides the sub-region to generate multiple grid units. After determining the position coordinates of each obstacle and dangerous object in the sub-region, the obstacle influence total value of each grid unit is calculated using the influence value determination algorithm. The exit influence total value of each grid unit is calculated according to the exit influence coefficient and the Euclidean distance between the center point of the grid unit and the exit. The crowd influence total value of each grid unit is calculated using the crowd influence coefficient and the Euclidean distance between the center point of the grid unit and the crowd. After adding the three, the field value of each grid unit is obtained, and the moving direction corresponding to any center point is determined using these data.

[0010] Preferably, the recognition model construction unit includes a sample information acquisition module and a feature vector determination module. After the sample information acquisition module obtains the sample data sets of people, obstacles, dangerous objects, and exit signs, it sets corresponding identification frames for the sample information in each sample data set. Each sample information contains multiple image data frames. The pixel point analysis algorithm is used to calculate the amplitude and direction of the gradient of each pixel point in the image data frame. The gradient feature vector determination module divides the area involved in the image data frame into multiple small units, obtains the corresponding nine-level gradient histogram according to the amplitude and direction of the pixel point gradient in each small unit, and statistically analyzes the histograms corresponding to adjacent areas. After analyzing the corresponding four-dimensional feature vector for each small unit, the histograms of all small units are aggregated to obtain a nine-dimensional feature vector, and the corresponding eighteen-level gradient histogram is analyzed using the amplitude and direction of the pixel point gradient in each small unit, and it is aggregated to obtain an eighteen-dimensional feature vector. All feature vectors are associated to determine the final gradient feature vector of the image data frame. The pixel point analysis algorithm is specifically:

[0011]

[0012] θ(x,y) = arctan(Gy(x,y) / Gx(x,y))

[0013] where G(x,y) represents the pixel point gradient amplitude at (x,y), θ(x,y) represents the pixel point gradient direction at (x,y), G x (x,y) represents the pixel point horizontal gradient at (x,y), G y (x,y) represents the pixel point vertical gradient at (x,y), and (x,y) represents the specific position corresponding to the pixel point.

[0014] Preferably, the recognition model construction unit further includes a parameter update module. After the parameter update module extracts the color feature vector of the image data frame, it fuses it with the gradient feature vector, constructs a target recognition model using a kernel correlation filter and a regression model, inputs all the fusion results included in each sample information into the target recognition model for analysis, obtains the response map corresponding to each frame, calculates the accuracy corresponding to the model result according to the difference between the peak value of the response map of each frame and other points in the map. If the accuracy is lower than the threshold, it is determined that the tracking is invalid, otherwise it is determined that the tracking is valid. The filter parameters and the target template are optimized using a parameter update algorithm. The kernel correlation filter is specifically a correlation filter based on the KCF algorithm, and the regression model is specifically a ridge regression model. The parameter update algorithm is specifically:

[0015]

[0016] Among them, α represents the updated filter parameter, x represents the updated target template, and x t-1 represents the target template of the previous frame, and x t represents the target template of the current frame. η represents the linear interpolation factor, and α t represents the filter parameter of the current frame, and α t-1 represents the filter parameter of the previous frame.

[0017] Preferably, the target box detection unit includes a position analysis module, a similarity calculation module, and a sample information deletion module. If the tracking is effective, the position analysis module uses the target recognition model to scan the image information contained in each frame according to the peak position of the response map of each frame and the size of the target box in the previous frame, and outputs the position and size of the target box corresponding to the current image data frame. If the tracking is ineffective, it scans each piece of image information contained in the frame one by one and outputs the position and size of the target box corresponding to the current image data frame. The similarity calculation module determines the similarity between the position and size of the target box of the image data frame contained in each sample information and the position and size of the identification box. If the similarity is lower than the preset value, it recalculates the peak value of the response map corresponding to the image data frame and updates the detection parameters. Otherwise, it calculates the similarity between the position and size of the target box of the current frame and the previous frame. If the similarity between the current image data frame and the previous frame is greater than the average value, the sample information deletion module deletes the current frame. If this value is less than or equal to the average value, it retains the current frame. After determining the similarity of all image data frames in the sample information, it counts the number of image data frames contained in the current sample information. If the number is lower than the threshold, it deletes the sample information. Otherwise, it does not perform any operation.

[0018] Preferably, the global position analysis unit includes a data frame sampling module, an exit position parsing module, and a coordinate conversion module. The data frame sampling module divides a specified scene into multiple sub-regions, numbers each sub-region, uses a camera to capture each sub-region, and transmits the obtained video data to an edge server. A sampling interval is set, and according to the sampling interval, the image data frames included in each video data are collected. After sorting the data frames in each video data according to the time sequence, they are stored in a sampling set. The exit position parsing module uses a target recognition model to sequentially scan the image data frames in the sampling set, and respectively obtains the local position information of the crowd, obstacles, dangerous objects, and exit signs included in the image data frame corresponding to each sub-region. Using deep learning technology, according to the content of the exit sign and the local position information, the local position information of the exit included in the image data frame corresponding to each sub-region is analyzed. The coordinate conversion module transmits the local position information related to each sub-region to a cloud server, and uses a coordinate optimization algorithm to convert the local position information related to each sub-region to obtain the global position information of each crowd, obstacle, dangerous object, and exit in the specified scene.

[0019] Preferably, the influence degree judgment unit includes an obstacle influence value calculation module, an exit influence value calculation module, a crowd influence value calculation module, and a field value calculation module. The obstacle influence value calculation module analyzes all the passage paths according to all the local position information of the exits in the sub-region, divides the regions involved in the passage paths, generates multiple grid cells, determines the position coordinates of each obstacle and dangerous object in the sub-region, and then uses an influence value determination algorithm to statistically analyze the influence degree of each obstacle and dangerous object on each grid cell, and calculates the total obstacle influence value of each grid cell according to the weights corresponding to the dangerous objects and obstacles. After the exit influence value calculation module determines the local position coordinates of each exit in the sub-region, it calculates the influence degree of each grid cell affected by each exit according to the exit influence coefficient and the Euclidean distance between the center point of the grid cell and the exit, and takes the maximum value of the exit influence degree as the total exit influence value of the corresponding grid cell. The crowd influence value calculation module obtains the local position coordinates of each crowd in the sub-region, calculates the influence degree of each grid cell affected by each crowd according to the crowd influence coefficient and the Euclidean distance between the center point of the grid cell and the crowd, and accumulates them to obtain the total crowd influence value of the corresponding grid cell. The field value calculation module statistically analyzes the total obstacle influence value, the total exit influence value, and the total crowd influence value corresponding to each grid cell, accumulates the three, obtains the field value of each grid cell, calculates the passage path loss value from the center point of each grid cell to the exit position according to the field value size, and uses these data to determine the moving direction corresponding to any center point.

[0020] Preferably, the evacuation path generation unit includes a movement direction optimization module, a speed analysis module, and a movement path simulation module. The movement direction optimization module transmits the field values, passage path loss values, and movement directions of the grid cells in each sub-region to the cloud server, calculates the global field value of each grid cell using the global coordinate information of the center points of the grid cells in each sub-region, adjusts the movement direction of the center points, and returns this data to the edge server. The speed analysis module calculates the density value corresponding to each grid cell based on the local position information of the crowd in the sub-region, analyzes the movement speed when passing through each center point using the density value, transmits it to the cloud server, analyzes the movement paths and speeds of the crowd evacuations at each location through an aggregation analysis algorithm to obtain the global passage path, and returns it to the edge server. The movement path simulation module builds a real-time scenario model through simulation software based on all the global position information and the local position information of each sub-region in the specified scenario, receives the movement paths and speeds corresponding to the crowd evacuations at each location for simulation, and outputs the obtained results through a visualization interface. The aggregation analysis algorithm is specifically the aggregation optimization artificial bee colony algorithm.

[0021] Preferably, the emergency handling unit includes an emergency value determination module and a personnel scheduling module. The emergency value determination module transmits the local position information and the collected data of the temperature and humidity sensors, gas sensors, and smoke sensors in each sub-region to the edge server, delimits multiple abnormal ranges for each type of data, sets emergency values for them, calculates the distances between each dangerous object and each sensor based on the local position information of the dangerous objects and sensors in each sub-region, determines the weights of the emergency values corresponding to each sensor in different dangerous objects according to the distances, and calculates the actual emergency value of each dangerous object in the sub-region. After receiving the actual emergency value of the dangerous object in each sub-region, the personnel scheduling module transmits it to the cloud server, sorts the dangerous objects in the specified scenario according to the size of the actual emergency value, and the control center schedules and deploys personnel based on the global position information and the sorting order of the dangerous objects.

[0022] Compared with the prior art, the beneficial effects of the present invention are:

[0023] 1. The present invention analyzes the image data frames in the sample data through the recognition model construction unit to obtain the multi-dimensional gradient feature vectors corresponding to each data frame. Such feature vectors can more accurately extract the details in the image data frames, reduce the number of mistakes of the model during the recognition process. At the same time, each frame will generate a corresponding response map, and the accuracy corresponding to the model result is calculated based on the difference between the peak value of the response map and other points in the map. If the accuracy is high, it indicates that the model can estimate the target range in the image data frame, and updating the corresponding parameters can further improve the performance of the filter. If the accuracy is low, it indicates that the model cannot estimate the target range, and avoiding parameter update can reduce the number of calculations and optimize the operation speed of the overall system. The target box detection unit scans the collected sample information and deletes the redundant sample information, increasing the storage space size without affecting the model training.

[0024] 2. The present invention analyzes the local position information of people, obstacles, dangerous objects, and exit signs contained in the image data frames corresponding to each sub-region in the edge server through the global position analysis unit, and uses the coordinate optimization algorithm to convert the local position information involved in each sub-region in the cloud server to obtain the global position information in the specified scenario. Such a design can, on the one hand, obtain the accurate position of the target, on the other hand, greatly improve the calculation speed of the system and reduce the requirements for the server performance. The influence degree judgment unit and the evacuation path generation unit calculate the optimal evacuation path of each group of people at the emergency site to avoid congestion when people evacuate. The emergency handling unit clarifies the urgency and location of dangerous objects, facilitating the scheduling and deployment of professional personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the overall system process provided by an embodiment of the present invention;

[0026] Figure 2 It is an internal module block diagram of the recognition model construction unit provided by an embodiment of the present invention;

[0027] Figure 3 It is an internal module block diagram of the target box detection unit provided by an embodiment of the present invention;

[0028] Figure 4 It is an internal module block diagram of the global position analysis unit provided by an embodiment of the present invention;

[0029] Figure 5 It is an internal module block diagram of the influence degree judgment unit provided by an embodiment of the present invention;

[0030] Figure 6 It is an internal module block diagram of the evacuation path generation unit provided by an embodiment of the present invention.

[0031] In the figure: 1. Recognition model construction unit; 101. Sample information acquisition module; 102. Feature vector determination module; 103. Parameter update module; 2. Target box detection unit; 201. Location analysis module; 202. Similarity calculation module; 203. Sample information deletion module; 3. Global location analysis unit; 301. Data frame sampling module; 302. Exit location parsing module; 303. Coordinate conversion module; 4. Influence degree judgment unit; 401. Obstacle influence value calculation module; 402. Exit influence value calculation module; 403. Crowd influence value calculation module; 404. Field value calculation module; 5. Evacuation path generation unit; 501. Moving direction optimization module; 502. Speed analysis module; 503. Moving path simulation module; 6. Emergency handling unit; 601. Emergency value determination module; 602. Personnel scheduling module. Specific implementation manners

[0032] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0033] Please refer to Figure 1-6 , the present invention provides a technical solution: an emergency site data perception system based on cloud-edge-end processing, including an evacuation path generation unit 5 and an emergency handling unit 6;

[0034] The recognition model construction unit 1. The recognition model construction unit 1 obtains a sample data set of crowds, obstacles, dangerous objects, and exit signs. Since each sample information in the data set contains multiple image data frames, the gradient amplitude and direction of each pixel point in the image data frame are calculated using the pixel point analysis algorithm. According to these data, the gradient feature vector of the image data frame is determined. The color feature vector is fused with the gradient feature vector, and a target recognition model is constructed using a kernel correlation filter and a regression model. All the fusion results included in each sample information are input into the target recognition model for analysis to obtain a response map corresponding to each frame;

[0035] The target box detection unit 2 scans the image information contained in each frame according to the peak position of each frame's response map and the size of the target box in the previous frame, outputs the position and size of the target box corresponding to the current image data frame, determines the similarity between the position and size of the target box of the image data frame contained in each sample information and the position and size of the identification box. If the similarity is lower than the preset value, recalculate the peak value of the response map corresponding to the image data frame and update the detection parameters. Otherwise, calculate the similarity between the position and size of the target box in the current frame and the previous frame;

[0036] The global position analysis unit 3 divides the specified scene into multiple sub-regions, uses a camera to take pictures of the sub-regions, transmits the obtained video data to the edge server, sorts the data frames in each video data according to the time sequence, stores them in the sampling set, uses the target recognition model to analyze the local position information of people, obstacles, dangerous objects, and exits contained in the image data frames corresponding to each sub-region, and uses the coordinate optimization algorithm to convert the local position information involved in each sub-region to obtain the global position information of people, obstacles, dangerous objects, and exits at each location in the specified scene;

[0037] The influence degree judgment unit 4 divides the sub-region to generate multiple grid units. After determining the position coordinates of each obstacle and dangerous object in the sub-region, it uses the influence value determination algorithm to calculate the total obstacle influence value of each grid unit, calculates the total exit influence value of each grid unit according to the exit influence coefficient and the Euclidean distance between the center point of the grid unit and the exit, calculates the total population influence value of each grid unit according to the population influence coefficient and the Euclidean distance between the center point of the grid unit and the population, adds the three together to obtain the field value of each grid unit, and uses this data to determine the moving direction corresponding to any center point.

[0038] The recognition model construction unit 1 includes a sample information acquisition module 101 and a feature vector determination module 102. After the sample information acquisition module 101 obtains the sample data sets of people, obstacles, dangerous objects, and exit signs, it sets corresponding identification frames for the sample information in each sample data set. Each sample information contains multiple image data frames. Using the pixel point analysis algorithm, it calculates the amplitude and direction of the gradient of each pixel point in the image data frame. The gradient feature vector determination module 102 divides the area involved in the image data frame into multiple small units, obtains the corresponding nine-level gradient histograms based on the gradient amplitude and direction of the pixel points in each small unit, and statistically analyzes the histograms corresponding to adjacent areas. After analyzing the corresponding four-dimensional feature vectors for each small unit, it aggregates the histograms of all small units to obtain a nine-dimensional feature vector, and also analyzes the corresponding eighteen-level gradient histograms using the gradient amplitude and direction of the pixel points in each small unit, aggregates them to obtain an eighteen-dimensional feature vector, and correlates all the feature vectors to determine the final gradient feature vector of the image data frame. The pixel point analysis algorithm is specifically:

[0039]

[0040] θ(x,y) = arctan(Gy(x,y) / Gx(x,y))

[0041] where G(x,y) represents the gradient amplitude of the pixel point at (x,y), θ(x,y) represents the gradient direction of the pixel point at (x,y), G x (x,y) represents the horizontal gradient of the pixel point at (x,y), G y (x,y) represents the vertical gradient of the pixel point at (x,y), and (x,y) represents the specific position corresponding to the pixel point;

[0042] The recognition model construction unit 1 also includes a parameter update module 103. After the parameter update module 103 extracts the color feature vector of the image data frame, it fuses it with the gradient feature vector, constructs a target recognition model using a kernel correlation filter and a regression model, inputs all the fusion results included in each sample information into the target recognition model for analysis, obtains the response map corresponding to each frame, calculates the accuracy corresponding to the model result based on the difference between the peak value of the response map of each frame and other points in the map. If the accuracy is lower than the threshold, it is determined that the tracking is invalid, otherwise it is determined that the tracking is valid, and the filter parameters and the target template are optimized using the parameter update algorithm. The kernel correlation filter is specifically a correlation filter based on the KCF algorithm, the regression model is specifically a ridge regression model, and the parameter update algorithm is specifically:

[0043]

[0044] Among them, α represents the filter parameter after update, x represents the target template after update, and x t-1 represents the target template of the previous frame, and x t represents the target template of the current frame, η represents the linear interpolation factor, and α t represents the filter parameter of the current frame, and α t-1 represents the filter parameter of the previous frame;

[0045] The target box detection unit 2 includes a position analysis module 201, a similarity calculation module 202, and a sample information deletion module 203. If the tracking is effective, the position analysis module 201 uses the target recognition model to scan the image information contained in each frame according to the peak position of the response map of each frame and the size of the target box in the previous frame, and outputs the position and size of the target box corresponding to the current image data frame. If the tracking is ineffective, it scans each piece of image information in the frame and outputs the position and size of the target box corresponding to the current image data frame. The similarity calculation module 202 determines the similarity between the position and size of the target box of the image data frame contained in each sample information and the position and size of the identification box. If the similarity is lower than the preset value, it recalculates the peak value of the response map corresponding to the image data frame and updates the detection parameters. Otherwise, it calculates the similarity between the position and size of the target box of the current frame and the previous frame. The sample information deletion module 203 deletes the current frame if the similarity between the current image data frame and the previous frame is greater than the average value, and retains the current frame if this value is less than or equal to the average value. After determining the similarity of all image data frames in the sample information, it counts the number of image data frames contained in the current sample information. If the number is lower than the threshold, it deletes the sample information. Otherwise, it does not perform any operation;

[0046] The global position analysis unit 3 includes a data frame sampling module 301, an exit position parsing module 302, and a coordinate conversion module 303. The data frame sampling module 301 divides a specified scene into multiple sub-regions, numbers each sub-region, uses a camera to take pictures of the sub-regions, transmits the obtained video data to an edge server, sets a sampling interval, collects the image data frames included in each video data according to the sampling interval, sorts the data frames in each video data according to the time series, and stores them in a sampling set. The exit position parsing module 302 uses a target recognition model to sequentially scan the image data frames in the sampling set, respectively obtains the local position information of the crowd, obstacles, dangerous objects, and exit signs included in the image data frames corresponding to each sub-region, and uses deep learning technology to analyze the local position information of the exits included in the image data frames corresponding to each sub-region according to the content of the exit signs and the local position information. The coordinate conversion module 303 transmits the local position information involved in each sub-region to a cloud server, and uses a coordinate optimization algorithm to convert the local position information involved in each sub-region to obtain the global position information of each crowd, obstacle, dangerous object, and exit in the specified scene. The coordinate optimization algorithm is specifically:

[0047]

[0048] Among them, represents the global position information of the individual O in the scene, a represents the length corresponding to each sub-region, b represents the width corresponding to each sub-region, O represents the number information corresponding to the crowd, obstacles, dangerous objects, and exits, (i, j) represents the number information of each sub-region, and (x, y) represents the local position information of the individual O in the sub-region where it is located;

[0049] The impact degree judgment unit 4 includes an obstacle impact value calculation module 401, an exit impact value calculation module 402, a crowd impact value calculation module 403, and a field value calculation module 404. The obstacle impact value calculation module 401 analyzes all the passage paths based on all the local position information of the exits in the sub-region, divides the regions involved in the passage paths, generates multiple grid cells, determines the position coordinates of each obstacle and dangerous object in the sub-region, and then uses the impact value determination algorithm to statistically calculate the impact degree of each obstacle and dangerous object on each grid cell. According to the weights corresponding to the dangerous objects and obstacles, the total obstacle impact value of each grid cell is calculated. After the exit impact value calculation module 402 determines the local position coordinates of each exit in the sub-region, it calculates the degree of influence of each grid cell by each exit according to the exit impact coefficient and the Euclidean distance between the center point of the grid cell and the exit, and takes the maximum value of the exit impact degree as the total exit impact value of the corresponding grid cell. The crowd impact value calculation module 403 obtains the local position coordinates of each crowd in the sub-region, calculates the degree of influence of each grid cell by each crowd according to the crowd impact coefficient and the Euclidean distance between the center point of the grid cell and the crowd, and accumulates them to obtain the total crowd impact value of the corresponding grid cell. The field value calculation module 404 statistically calculates the total obstacle impact value, the total exit impact value, and the total crowd impact value corresponding to each grid cell, adds the three together to obtain the field value of each grid cell, calculates the passage path loss value from the center point of each grid cell to the exit position according to the field value size, and uses these data to determine the moving direction corresponding to any center point. The impact value determination algorithm is specifically as follows:

[0050]

[0051] Among them, (i, j) represents the coordinate position of the center point of the grid cell. If D a represents the position coordinates of the dangerous object, then c a (i, j) represents the impact value of the dangerous object a on the grid cell with the position coordinates (i, j), m represents the impact coefficient of the dangerous object, l((i, j), D a ) represents the Euclidean distance between the center point of the grid cell and the dangerous object a, γ represents the dangerous object impact distance threshold, and l represents the distance value. If D a represents the position coordinates of the obstacle, then c a (i, j) represents the impact value of the obstacle a on the grid cell with the position coordinates (i, j), m represents the impact coefficient of the obstacle, l((i, j), D a ) represents the Euclidean distance between the center point of the grid cell and the obstacle a, γ represents the obstacle impact distance threshold, and l represents the distance value;

[0052] The evacuation path generation unit 5 includes a moving direction optimization module 501, a speed analysis module 502, and a moving path simulation module 503. The moving direction optimization module 501 transmits the field values, passage path loss values, and moving directions of the grid cells in each sub-region to the cloud server, calculates the global field value of each grid cell using the global coordinate information of the center points of the grid cells in each sub-region, adjusts the moving direction of the center points, and returns this data to the edge server. The speed analysis module 502 calculates the density value corresponding to each grid cell based on the local position information of the crowd in the sub-region, analyzes the moving speed when passing through each center point using the density value, and transmits it to the cloud server. The moving path and speed of each crowd evacuation are analyzed through an aggregation analysis algorithm to obtain the global passage path, which is returned to the edge server. The moving path simulation module 503 builds a real-time scene model through simulation software according to all the global position information and the local position information of each sub-region in the specified scene, receives the moving path and speed corresponding to each crowd evacuation and performs simulation, and outputs the obtained results through a visualization interface. The aggregation analysis algorithm is specifically the aggregation optimization artificial bee colony algorithm, the cloud server is specifically the Tencent cloud server, and the edge server is specifically the Atlas 500Pro edge server;

[0053] The emergency handling unit 6 includes an emergency value determination module 601 and a personnel scheduling module 602. The emergency value determination module 601 transmits the local position information and the collected data of the temperature and humidity sensors, gas sensors, and smoke sensors in each sub-region to the edge server, delimits multiple abnormal ranges for each type of data, sets emergency values for them, calculates the distance between each dangerous object and each sensor according to the local position information of the dangerous objects and sensors in each sub-region, calculates the actual emergency value of each dangerous object in the sub-region after determining the weight of the emergency value corresponding to each sensor in different dangerous objects according to the distance, and the personnel scheduling module 602 receives the actual emergency value of the dangerous objects in each sub-region and transmits it to the cloud server, sorts the dangerous objects in the specified scene according to the size of the actual emergency value, and the control center schedules and deploys personnel according to the global position information and the sorting order of the dangerous objects.

[0054] Working principle: In the present invention, the sample information acquisition module 101 in the recognition model construction unit 1 analyzes the gradient amplitude and direction of each pixel point in the image data frame included in the sample information, the feature vector determination module 102 determines the gradient feature vector of the image data frame, the parameter update module 103 optimizes the filter parameters and the target template, the position analysis module 201 in the target box detection unit 2 outputs the position and size of the target box corresponding to the current image data frame, the similarity calculation module 202 determines the similarity between the position and size of the target box of the image data frame and the position and size of the identification box, the sample information deletion module 203 counts the number of image data frames included in the current sample information, and if the number is lower than the threshold, the sample information is deleted. After the data frame sampling module 301 in the global position analysis unit 3 sorts the data frames related to the sub-region according to the time series, it stores them in the sampling set. The exit position analysis module 302 analyzes the local exit position information included in the image data frame corresponding to each sub-region, and the coordinate conversion module 303 converts the local position information involved in each sub-region to obtain the corresponding global position information in the specified scenario. The obstacle influence value calculation module 401 in the influence degree judgment unit 4 calculates the total obstacle influence value of each grid cell, the exit influence value calculation module 402 determines the total exit influence value of the corresponding grid cell, and the crowd influence value calculation module 403 obtains the total crowd influence value of the corresponding grid cell. After the field value calculation module 404 accumulates the total obstacle influence value, the total exit influence value, and the total crowd influence value, it obtains the field value of each grid cell. The movement direction optimization module 501 in the evacuation path generation unit 5 adjusts the movement direction of the center point, the speed analysis module 502 analyzes the movement path and speed of each crowd evacuation to obtain the global passage path, and the movement path simulation module 503 builds a real-time scenario model for simulation and outputs the obtained results through the visualization interface. The actual emergency value determination module 601 in the emergency processing unit 6 calculates the actual emergency value of each dangerous object in the sub-region, and the personnel scheduling module 602 sorts the dangerous objects in the specified scenario according to the actual emergency value size to facilitate the scheduling and deployment of personnel.

[0055] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0056] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An emergency scene data perception system based on cloud-edge-end processing, comprising an evacuation path generation unit (5) and an emergency processing unit (6), characterized in that: A recognition model construction unit (1), wherein the recognition model construction unit (1) obtains a sample data set of a crowd, an obstacle, a dangerous object, and an exit sign. Since each sample information in the data set contains a plurality of image data frames, a pixel point analysis algorithm is used to calculate the gradient amplitude and direction of each pixel point in the image data frame, a gradient feature vector of the image data frame is determined based on these data, a color feature vector is fused with a gradient feature vector, a target recognition model is constructed using a kernel correlation filter and a regression model, and all fusion results contained in each sample information are input into the target recognition model for analysis to obtain a response map corresponding to each frame; A target frame detection unit (2), wherein the target frame detection unit (2) scans the image information contained in the frame according to the peak position of the response graph of each frame and the size of the target frame of the previous frame, outputs the target frame position and size corresponding to the current image data frame, determines the similarity between the target frame position and size of the image data frame contained in each sample information and the position and size of the identification frame, and if the similarity is lower than a preset value, recalculates the peak value of the response graph corresponding to the image data frame and updates the detection parameters, otherwise calculates the similarity between the target frame position and size of the current frame and the previous frame; A global position analysis unit (3), wherein the global position analysis unit (3) divides a designated scene into a plurality of sub-areas, uses a camera to shoot the sub-areas, transmits the obtained video data to an edge server, sorts the data frames in each video data according to a time series, stores the data frames in the video data in a sampling set, uses a target recognition model to analyze the local position information of the crowd, obstacles, dangerous objects and exits contained in the image data frames corresponding to each sub-area, uses a coordinate optimization algorithm to convert the local position information involved in each sub-area, and obtains the global position information of each crowd, obstacle, dangerous object and exit in the designated scene; The impact degree judgment unit (4) divides the sub-area to generate a plurality of grid units, determines the position coordinates of each obstacle and dangerous object in the sub-area, and then uses an impact value determination algorithm to calculate the total obstacle impact value of each grid unit, calculates the total exit impact value of each grid unit based on the exit impact coefficient and the Euclidean distance between the center point of the grid unit and the exit, and calculates the total crowd impact value of each grid unit based on the crowd impact coefficient and the Euclidean distance between the center point of the grid unit and the crowd. After accumulating the three, the field value of each grid unit is obtained, and the moving direction corresponding to any center point is determined using these data.

2. The emergency scene data perception system based on cloud-edge processing according to claim 1 is characterized in that: The recognition model building unit (1) comprises a sample information acquisition module (101) and a feature vector determination module (102). After acquiring a sample data set of a crowd, an obstacle, a dangerous object and an exit sign, the sample information acquisition module (101) sets a corresponding identification frame for the sample information in each sample data set. Each sample information includes a plurality of image data frames. The pixel point analysis algorithm is used to calculate the gradient amplitude and direction of each pixel point in the image data frame. The gradient feature vector determination module (102) divides the area involved in the image data frame to obtain a plurality of small units. According to the gradient amplitude and direction of the pixel point in each small unit, a corresponding nine-level gradient histogram is analyzed, and the histograms corresponding to the adjacent areas are counted. After the four-dimensional feature vector corresponding to each small unit is analyzed according to these data, the histograms of all the small units are aggregated to obtain a nine-dimensional feature vector. The gradient amplitude and direction of the pixel point in each small unit are analyzed to obtain a corresponding eighteen-level gradient histogram, which is aggregated to obtain an eighteen-dimensional feature vector. All the feature vectors are associated to determine the final gradient feature vector of the image data frame.

3. The emergency scene data perception system based on cloud-edge-end processing according to claim 2 is characterized in that: The recognition model construction unit (1) also includes a parameter updating module (103). After extracting the color feature vector of the image data frame, the parameter updating module (103) fuses the color feature vector with the gradient feature vector, constructs a target recognition model using a kernel correlation filter and a regression model, inputs all fusion results contained in each sample information into the target recognition model for analysis, obtains a response graph corresponding to each frame, calculates the accuracy corresponding to the model result based on the difference between the peak value of the response graph of each frame and other points in the graph, and determines that the tracking is invalid if the accuracy is lower than a threshold value, otherwise it is determined that the tracking is valid, and optimizes the filter parameters and the target template using a parameter updating algorithm.

4. The emergency scene data perception system based on cloud-edge processing according to claim 1 is characterized in that: The target frame detection unit (2) comprises a position analysis module (201), a similarity calculation module (202) and a sample information deletion module (203). If the tracking is effective, the position analysis module (201) uses the target recognition model to scan the image information contained in the frame according to the peak position of each frame response graph and the size of the target frame of the previous frame, and outputs the position and size of the target frame corresponding to the current image data frame. If the tracking is invalid, the image information contained in the frame is scanned one by one, and the position and size of the target frame corresponding to the current image data frame are output. The similarity calculation module (202) determines the target frame of each sample information contained in the image data frame. The similarity between the position and size of the target frame and the position and size of the identification frame is calculated. If the similarity is lower than a preset value, the peak value of the response graph corresponding to the image data frame is recalculated and the detection parameters are updated. Otherwise, the similarity between the position and size of the target frame of the current frame and the previous frame is calculated. If the similarity between the current image data frame and the previous frame is greater than the average value, the current frame is deleted. If the value is less than or equal to the average value, the current frame is retained. After the similarity of all image data frames in the sample information is determined, the number of image data frames included in the current sample information is counted. If the number is lower than a threshold, the sample information is deleted. Otherwise, no operation is performed.

5. The emergency scene data perception system based on cloud-edge processing according to claim 1 is characterized in that: The global position analysis unit (3) comprises a data frame sampling module (301), an exit position analysis module (302) and a coordinate conversion module (303). The data frame sampling module (301) divides a specified scene into a plurality of sub-areas, numbers each sub-area, shoots the sub-area with a camera, transmits the obtained video data to an edge server, sets a sampling interval, collects image data frames contained in each video data according to the sampling interval, sorts the data frames in each video data according to a time series, and stores them in a sampling set. The exit position analysis module (302) uses a target recognition model to obtain the image data frames contained in each video data. The image data frames in the sampling set are scanned in sequence to obtain the local position information of the crowd, obstacles, dangerous objects and exit signs contained in the image data frames corresponding to each sub-region respectively. The local position information of the exit contained in the image data frames corresponding to each sub-region is analyzed according to the content and local position information of the exit sign by using deep learning technology. The coordinate conversion module (303) transmits the local position information involved in each sub-region to the cloud server, and uses the coordinate optimization algorithm to convert the local position information involved in each sub-region to obtain the global position information of each crowd, obstacle, dangerous object and exit in the specified scene.

6. The emergency scene data perception system based on cloud-edge processing according to claim 1 is characterized in that: The impact degree judgment unit (4) comprises an obstacle impact value calculation module (401), an exit impact value calculation module (402), a crowd impact value calculation module (403) and a field value calculation module (404). The obstacle impact value calculation module (401) analyzes all the passage paths according to the local position information of all the exits in the sub-area, divides the area involved in the passage path, generates a plurality of grid units, determines the position coordinates of each obstacle and dangerous object in the sub-area, and uses an impact value determination algorithm to count the degree of influence of each obstacle and dangerous object on each grid unit, and calculates the total obstacle impact value of each grid unit according to the weights corresponding to the dangerous objects and obstacles. After the exit impact value calculation module (402) determines the local position coordinates of each exit in the sub-area, it calculates the total obstacle impact value of each grid unit according to the exit impact coefficient and the grid unit. The Euclidean distance between the center point of the sub-area and the exit is used to calculate the degree to which each grid unit is affected by each exit, and the maximum exit influence degree is taken as the total exit influence value of the corresponding grid unit. The crowd influence value calculation module (403) obtains the local position coordinates of each crowd in the sub-area, and calculates the degree to which each grid unit is affected by each crowd according to the crowd influence coefficient and the Euclidean distance between the center point of the grid unit and the crowd, and accumulates them to obtain the total crowd influence value of the corresponding grid unit. The field value calculation module (404) counts the total obstacle influence value, the total exit influence value and the total crowd influence value corresponding to each grid unit, and accumulates the three to obtain the field value of each grid unit. The path loss value from the center point of each grid unit to the exit position is calculated according to the field value, and the moving direction corresponding to any center point is determined using these data.

7. The emergency scene data perception system based on cloud-edge processing according to claim 1 is characterized in that: The evacuation path generation unit (5) comprises a moving direction optimization module (501), a speed analysis module (502) and a moving path simulation module (503). The moving direction optimization module (501) transmits the field value, the passage path loss value and the moving direction of the grid unit in each sub-area to the cloud server, calculates the global field value of each grid unit using the global coordinate information of the center point of the grid unit in each sub-area, adjusts the moving direction of the center point, and returns these data to the edge server. The speed analysis module (502) calculates the density value corresponding to each grid unit according to the local position information of the crowd in the sub-area, analyzes the moving speed when passing through each center point using the density value, transmits it to the cloud server, analyzes the moving path and speed of each crowd evacuation by using the aggregation analysis algorithm, obtains the global passage path, and returns it to the edge server. The moving path simulation module (503) builds a real-time scene model according to all the global position information and the local position information of each sub-area in the specified scene through simulation software, simulates after receiving the moving path and speed corresponding to the evacuation of each crowd, and outputs the obtained results through a visual interface.

8. The emergency scene data perception system based on cloud-edge processing according to claim 1 is characterized in that: The emergency handling unit (6) comprises an emergency value determination module (601) and a personnel dispatching module (602). The emergency value determination module (601) transmits the local position information of the temperature and humidity sensor, the gas sensor and the smoke sensor in each sub-area and the collected data to the edge server, defines a plurality of abnormal ranges for each data, and sets an emergency value for it. The distance between each dangerous object and each sensor is calculated according to the local position information of the dangerous object and the sensor in each sub-area. After the weight of the emergency value corresponding to each sensor in different dangerous objects is determined according to the distance, the actual emergency value of each dangerous object in the sub-area is calculated. After receiving the actual emergency value of the dangerous object in each sub-area, the personnel dispatching module (602) transmits it to the cloud server, sorts the dangerous objects in the specified scene according to the actual emergency value, and the control center dispatches and deploys personnel according to the global position information and arrangement order of the dangerous objects.

Images