An intelligent identification method for park risks based on location information

Abstract

The present invention discloses a method for intelligently identifying park risks based on location information, which relates to the technical field of park safety management. The method collects personnel and vehicle positioning data, fixed risk source coordinates, equipment status, environmental parameters and video stream data in real time; adopts the Geohash algorithm to divide the park into grids, and realizes spatiotemporal alignment of multi-source data through a sliding time window; improves and trains the YOLOv8 model to output real-time park abnormal behavior detection results; quantifies the behavior risks of mobile targets based on the entropy weight method, and superimposes the risks of equipment and environmental exceeding the standard to generate grid-level superimposed risk values; combines the real-time park abnormal behavior detection results with the superimposed risk values ​​to obtain the final risk results. The present invention breaks through the data island limitations of traditional monitoring systems, improves the risk identification rate and early warning response speed, reduces the risk false detection rate, and realizes the transformation from passive response to active prevention and control.

Description

Technical Field

[0001] The present invention belongs to the technical field of environmental management, and in particular, relates to a method for intelligently identifying park risks based on location information. Background Art

[0002] Frequent safety incidents in high-risk industrial parks, such as those in the chemical and energy sectors, urgently require intelligent prevention and control measures. The widespread adoption of the Internet of Things (IoT) and edge computing technologies has made it possible to collect real-time positioning data, equipment status, and video streams. However, the integrated analysis of multi-source, heterogeneous data remains a pain point in the industry. Existing solutions suffer from three common flaws: first, video analysis models lack adaptability to specific targets (such as flames and falls); second, risk calculations fail to incorporate the entropy weight of movement behavior; and third, alarm decisions rely on a single data source, lacking a comprehensive analysis mechanism. The rise of digital twin technology offers new insights into spatial risk modeling, but achieving a closed loop between data, model, and decision-making remains a challenge.

[0003] In traditional campus safety monitoring systems, multi-source data (such as personnel location, equipment status, environmental parameters, and video streams) is often processed independently, resulting in a fragmentation of spatiotemporal information and making dynamic risk linkage analysis difficult. Video surveillance relies on general target detection models, which have low recognition rates for fuzzy targets like smoke and flames, as well as small targets at a distance (such as people falling). Furthermore, the system lacks correlation with the location of hazardous sources in physical space, resulting in high rates of false detections and missed detections. Existing risk assessments often use static thresholding methods, which are unable to quantify the combined effects of abnormal clustering of moving targets (such as random trajectories, sudden stops and accelerations) and equipment environmental risks, leading to delayed warnings or false alarms. Limited by the computing power of edge devices, high-precision models are difficult to run in real time, hindering the response speed in complex scenarios.

[0004] Existing technologies have problems such as multi-source data fragmentation, insufficient video detection capabilities, and rigid risk assessment mechanisms. Summary of the Invention

[0005] (1) Technical problems solved

[0006] In response to the problems in the related art, the present invention provides a method for intelligently identifying park risks based on location information to overcome the above-mentioned technical problems existing in the existing related art.

[0007] (2) Technical solution

[0008] To solve the above technical problems, the present invention is achieved through the following technical solutions:

[0009] S1. Through real-time multi-source data collection operations, obtain real-time flow location data, fixed risk source location data, equipment status data, environmental parameter data and real-time video stream data;

[0010] S2: Divide the park into multiple grids, map the various data collected in S1 to the corresponding grids to obtain a grid dataset; align the data in the grid dataset to obtain an aligned grid dataset;

[0011] S3. Build a YOLOv8 model, improve the YOLOv8 model, and obtain an improved YOLOv8 model; use historical annotated video data combined with an optimization algorithm to train the improved YOLOv8 model and obtain the final YOLOv8 model;

[0012] Based on the aligned grid dataset and the final YOLOv8 model, abnormal behavior recognition in video streams was performed to obtain real-time abnormal behavior detection results in the park.

[0013] S4. Based on the aligned grid data set, calculate the park location entropy weighted risk value; superimpose the park location entropy weighted risk value with the equipment and environment data to obtain the superimposed risk value;

[0014] S5. Obtain the final risk result based on the superimposed risk value and the real-time park abnormal behavior detection results;

[0015] This invention eliminates data silos through grid-based spatiotemporal alignment, integrating personnel trajectories, equipment, environment, and video streams. Through the final YOLOv8 model, the confidence level is dynamically optimized in combination with the location of hazardous sources, which reduces the system load rate while improving the detection rate of abnormal targets and reducing the false detection rate of abnormalities. The entropy weight method integrates mobile behavior anomalies with equipment environmental risks, improving the early warning response speed. The dual-threshold mechanism triggers precise alarms, reduces false alarms, improves the accident recognition rate, and achieves a breakthrough from passive monitoring to active prevention and control.

[0016] Preferably, the S1 comprises the following steps:

[0017] S11. Set fixed risk sources within the park to obtain a set of fixed risk sources in the park; based on the set of fixed risk sources in the park, collect the tag coordinates of the fixed risk sources to obtain the location data of the fixed risk sources;

[0018] The deployed UWB positioning base station captures the coordinates of people and vehicle tags in real time to obtain real-time mobile location data;

[0019] S12: Read IoT sensor data from devices in the park to obtain device status data; collect environmental monitoring point data to obtain environmental parameter data; and call cameras to collect real-time video stream data.

[0020] The present invention uses UWB positioning base stations, IoT sensors and cameras to synchronously collect five types of key data in real time, including the flow location of people / vehicles, the coordinates of fixed risk sources, the operating status of equipment, environmental parameters and video streams; it eliminates multi-source data islands, provides a high-precision, full-dimensional real-time data base for dynamic risk analysis, and lays the foundation for coordinated prevention and control across the entire region.

[0021] Preferably, said S2 comprises the following steps:

[0022] S21, using the Geohash algorithm to perform grid division on the park to obtain a park grid set;

[0023] Mapping the real-time flow location data, fixed risk source location data, equipment status data, environmental parameter data, and real-time video stream data to corresponding grids in the park grid set to obtain a grid data set; the grid data set contains the real-time flow location data, fixed risk source location data, equipment status data, environmental parameter data, and real-time video stream data for each area of ​​the park;

[0024] S22. Setting a sliding time window, and aligning the multi-source data in the grid dataset within the window to obtain an aligned grid dataset;

[0025] The present invention divides the park into high-precision grids through the Geohash algorithm, maps multi-source data to the corresponding grids, and then achieves spatiotemporal alignment through a sliding time window; it breaks through the spatiotemporal correlation barriers between personnel trajectories, equipment status, environmental parameters and video streams, eliminates data silos, provides a unified spatiotemporal benchmark for dynamic risk analysis, and improves the accuracy of locating abnormal events.

[0026] Preferably, the step S3 includes the following steps:

[0027] S31. Build a YOLOv8 model, improve the YOLOv8 model, and obtain an improved YOLOv8 model.

[0028] S32, collecting historical video data; the historical video data includes normal situations and abnormal situations; annotating the historical video data to obtain historical annotated video data;

[0029] S33. Use the historically labeled video data to train the improved YOLOv8 model. During the training process, use an optimization algorithm to find the network parameters of the improved YOLOv8 model and obtain an optimal solution. The optimal solution is used as the network parameters of the improved YOLOv8 model to obtain a final YOLOv8 model.

[0030] S34. Input the real-time video stream of each grid in the aligned grid dataset into the final YOLOv8 model to obtain a real-time park abnormal behavior detection result; the real-time park abnormal behavior detection result includes the abnormal situation of the park location corresponding to each grid;

[0031] The present invention improves the YOLOv8 model and optimizes the parameters through historical annotated data training and algorithm optimization, outputting grid-level real-time anomaly detection results, thereby improving the anomaly recognition rate, reducing the false detection rate, and reducing resource consumption.

[0032] Preferably, the S31 includes the following steps:

[0033] S311. Optimize the model's recognition of small and unconventional targets; add a dedicated detection head for small and unconventional targets, adjust the anchor box size to match their characteristics, and increase the output of high-resolution features in the P2 layer; strengthen the use of shallow high-resolution features in the feature pyramid, and embed CBAM dual-channel attention in the Neck layer;

[0034] S312: Embed the location information of hazardous sources in the digital twin map of the park into the detection process. For each identified target, calculate the Euclidean distance between it and the nearest hazardous source in real time. When a target enters a high-risk area, dynamically increase its detection confidence and mark it as a high-risk state.

[0035] S313: Model distillation technology is used to compress the model volume; a dynamic resolution mechanism is introduced, and conventional monitoring uses low resolution to maintain a high frame rate; when the system detects an increase in regional risk value, it automatically switches to high-resolution mode for detailed analysis of anomalies;

[0036] The present invention achieves structural optimization and improves the anomaly recognition rate by adding a dedicated detection head for small targets / unconventional targets, adjusting the anchor frame size and adding a high-resolution P2 layer, combining the feature pyramid to enhance shallow features and the CBAM dual-channel attention mechanism; by embedding the coordinates of digital twin hazard sources, the Euclidean distance between the target and the nearest hazard source is calculated in real time, and the confidence level is dynamically improved when personnel enter the high-risk area, thereby achieving spatial intelligence enhancement and significantly reducing the false detection rate in complex scenarios; using model distillation to compress the parameters and designing a dynamic resolution mechanism to achieve an on-demand balance between accuracy and efficiency, and reduce the resource consumption of edge devices.

[0037] Preferably, the step of searching for network parameters of the improved YOLOv8 model in combination with an optimization algorithm to obtain an optimal solution in S33 includes the following steps:

[0038] S331. Set the recognition accuracy threshold and recognition accuracy of the improved YOLOv8 model; set the maximum number of iterations;

[0039] S332: Construct a particle swarm and set the size of the particle swarm according to the network parameters of the improved YOLOv8 model;

[0040] S333. Construct a particle position set based on the particle swarm; define a fitness function of the particle positions in the particle swarm based on the recognition accuracy threshold and the recognition accuracy;

[0041] S334, performing an iterative operation on the particle position set; in each round of iteration, calculating the fitness values ​​of the particle positions in the particle position set according to the fitness function, updating the particle position set in descending order of fitness values, and obtaining the best particle position in the particle positions and the global best particle position in each round of iteration;

[0042] S335, repeat S334, when the maximum number of optimization iterations is reached, stop the iteration, and take the global best particle position as the optimal solution;

[0043] The present invention evaluates model accuracy by defining a fitness function, dynamically updates particle positions during iteration, and ultimately converges to the globally optimal network parameters. This breaks through the local optimal limitations of traditional gradient descent, improves the recognition accuracy of the final YOLOv8 model, increases training efficiency, and avoids the risk of overfitting.

[0044] Preferably, the S4 comprises the following steps:

[0045] S41. Based on the aligned grid dataset, for each grid area in the park, calculate the aggregation risk of the mobile target and obtain the normalized real-time trajectory Shannon entropy, the normalized real-time velocity variance within the area, and the normalized real-time distance to the nearest hazard source;

[0046] S42, respectively setting weight coefficients of the trajectory Shannon entropy, the velocity variance within the region, and the normalized distance to the nearest hazard source to obtain multiple weight coefficients;

[0047] The entropy-weighted risk value of the park location is calculated based on the normalized real-time trajectory Shannon entropy, the normalized real-time speed variance within the area, the normalized distance to the nearest hazard source, and multiple weight coefficients.

[0048] S43, setting environmental standard values, equipment standard values, and risk factors; obtaining an additional risk value based on the environmental standard values, equipment standard values, and equipment status data and environmental parameter data in the aligned grid dataset;

[0049] This invention innovatively integrates the behavioral risk of moving targets, trajectory Shannon entropy to quantify movement randomness, speed variance to capture sudden stops / accelerations, and hazard source distance to assess spatial risks) with real-time data on the equipment environment, and dynamically generates a grid-level comprehensive superimposed risk value through the entropy weight method. It breaks through the limitations of static thresholds and advances the average warning response time for accidents such as leaks and fires.

[0050] Preferably, the S41 includes the following steps:

[0051] S411, obtaining a trajectory coordinate sequence of a moving target within each grid based on the real-time flow position data of each grid in the aligned grid data set;

[0052] Each grid in the aligned grid dataset is divided to obtain an aligned grid dataset matrix; based on the trajectory coordinate sequence of the moving target in each grid, the access frequency of each sub-grid is counted to obtain access frequency data;

[0053] Based on the access frequency data, the normalized real-time trajectory Shannon entropy is calculated;

[0054] S412, obtaining the instantaneous speed of all targets in each grid by aligning the real-time video stream data of each grid in the grid data set;

[0055] According to the instantaneous speed of all targets in each grid, the normalized real-time regional speed variance is calculated;

[0056] S413. Calculate the real-time normalized distance to the nearest hazard source based on the central area of ​​each grid and the fixed risk source location data;

[0057] The present invention calculates Shannon entropy based on subgrid access frequency to capture abnormal movement patterns such as wandering; calculates variance by extracting the instantaneous speed of the target in the video stream to identify sudden stop / acceleration behavior; dynamically assesses location risk by calculating the Euclidean distance between the grid center and the nearest hazard source in real time; and accurately quantifies behavioral anomalies that are difficult for traditional systems to capture, thereby improving the recognition rate of events such as falls and illegal intrusions, and increasing the speed of early warning.

[0058] Preferably, the S5 comprises the following steps:

[0059] S51. Setting an overlay risk threshold; when the overlay risk value of a grid area in the park grid set is greater than the overlay risk threshold and the real-time park abnormal behavior detection result is abnormal, the final risk result is to determine that the grid area is abnormal and issue an abnormal alarm;

[0060] When the superposition risk value of the grid area in the park grid set is less than the superposition risk threshold and the real-time park abnormal behavior detection result is abnormal, or when the superposition risk value of the grid area in the district grid set is greater than the superposition risk threshold and the real-time park abnormal behavior detection result is normal, a manual review alarm is issued, and the final risk result is obtained after manual review;

[0061] When the superposition risk value of a grid area in the park grid set is less than the superposition risk threshold and the real-time park abnormal behavior detection result is normal, the final risk result is that the grid area is judged to be normal;

[0062] The present invention adopts a dual-threshold decision-making mechanism; an automatic alarm is triggered only when the grid comprehensive risk value exceeds the threshold and the video recognition is abnormal; if a single condition is met, it is transferred to manual review; it reduces the false alarm rate, improves the accuracy of accident identification, and significantly improves the accuracy of alarms and security efficiency.

[0063] A location-based intelligent park risk identification system, used to implement the aforementioned location-based intelligent park risk identification method, comprising a multi-source data real-time acquisition module, a grid data alignment module, a video abnormal behavior intelligent identification module, a dynamic risk superposition calculation module, and a hierarchical risk decision-making and alarm module;

[0064] The multi-source data real-time acquisition module dynamically captures multi-dimensional real-time data within the park through IoT devices; periodically obtains real-time flow location data of people / vehicles through UWB positioning base stations; collects tag coordinate data of fixed risk sources; reads device status data generated by device IoT sensors; collects environmental parameter data from environmental monitoring points; and calls cameras to obtain real-time video stream data.

[0065] The grid data alignment module uniformly maps the park space and multi-source data to a standardized grid system; uses the Geohash algorithm to divide the park into high-precision grids, and maps real-time location, risk source coordinates, equipment status, environmental parameters, and video stream data to the corresponding grids; and achieves temporal alignment of multi-source data through a sliding time window to generate a spatially and temporally consistent aligned grid dataset.

[0066] The video abnormal behavior intelligent recognition module is used to improve the YOLOv8 model to obtain an improved YOLOv8 model; and the improved YOLOv8 model is trained using historical annotated video data, combined with the particle swarm optimization algorithm to adjust parameters, and output grid-level real-time abnormal behavior detection results;

[0067] The dynamic risk overlay calculation module is used to integrate mobile trajectory, environment and equipment data to quantify regional risks. By calculating the normalized real-time Shannon entropy based on trajectory coordinates, the normalized intra-regional variance of target velocity, and the normalized distance to the nearest hazard source, the mobile target behavior risk index within the grid is obtained. The entropy weight method is combined with the weight coefficient to generate the position entropy weighted risk value. The equipment and environmental risks are then added to finally output the overlay risk value.

[0068] The hierarchical risk decision and alarm module formulates a response strategy based on the video analysis results and the superimposed risk value. If the grid area meets the superimposed risk value exceeding the threshold and the video detects abnormal behavior, it is directly determined to be abnormal and an alarm is triggered. If only a single condition is met, a manual review alarm is triggered, and the final risk status is confirmed after manual review. When both indicators are normal, the area is determined to be safe, realizing hierarchical intelligent decision-making.

[0069] (3) Beneficial effects

[0070] The present invention has the following beneficial effects:

[0071] This invention eliminates data silos through grid-based spatiotemporal alignment, integrating personnel trajectories, equipment, environment, and video streams. Through the final YOLOv8 model, the confidence level is dynamically optimized in combination with the location of hazardous sources, reducing the system load rate while improving the detection rate of abnormal targets and lowering the false detection rate of abnormalities. The entropy weight method integrates mobile behavior anomalies with equipment environmental risks, improving the early warning response speed. The dual-threshold mechanism triggers precise alarms, reducing false alarms and improving the accident recognition rate.

[0072] The present invention divides the park into grids through the Geohash algorithm and adopts a sliding window to achieve spatiotemporal alignment, breaking down the barriers associated with personnel trajectories, equipment data, environmental parameters and video streams, and solving the response lag problem caused by data silos in traditional systems; spatiotemporal alignment improves the accuracy of locating abnormal events.

[0073] The present invention embeds the CBAM dual-channel attention mechanism in the Neck layer of the YOLOv8 model, significantly enhancing the response capability to fuzzy textures through spatial-channel joint weight distribution; adds a new high-resolution P2 feature layer to improve the detection rate of long-distance and small targets; embeds the coordinates of hazard sources in the digital twin map into the detection process to dynamically improve the confidence level, and combines the dedicated detection head and anchor frame optimization to reduce the abnormal false detection rate and missed detection rate; adopts model distillation technology to compress the YOLOv8 parameters, and uses a dynamic resolution mechanism to improve the recognition accuracy of high-threat scenes while reducing the overall resource consumption of the system.

[0074] The present invention quantifies the randomness of movement based on the trajectory Shannon entropy, captures abnormal acceleration / sudden stop through speed variance, calculates the position entropy weight risk value in combination with the normalized distance to the nearest hazard source, and obtains a risk assessment model that integrates the entropy weight of mobile behavior. The model then outputs a comprehensive risk value after superimposing it with the behavioral risk, thereby improving the time for early warning identification and response.

[0075] The present invention establishes a dual-trigger mechanism of "video anomaly recognition + superimposed risk threshold". This strategy reduces false alarms, improves recognition accuracy, and reduces the burden on security personnel.

[0076] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the technical solutions of the embodiments of the invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the invention. For ordinary technicians in this field, they can also obtain drawings based on these drawings without paying any creative work.

[0078] Figure 1 This is a flow chart of a method for intelligently identifying park risks based on location information according to the present invention;

[0079] Figure 2 This is a flow chart of obtaining the final risk analysis result in a method for intelligently identifying park risks based on location information according to the present invention.

[0080] Figure 3 This is a module schematic diagram of a park risk intelligent identification system based on location information of the present invention. DETAILED DESCRIPTION

[0081] The following will clearly and completely describe the technical solutions in the embodiments of the invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only part of the embodiments of the invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0082] In the description of the present invention, it should be understood that the terms "opening", "upper", "lower", "top", "middle", "inside" and the like indicating orientation or positional relationship are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the components or elements referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the invention.

[0083] Example 1:

[0084] See also Figure 1 The present invention discloses a method for intelligently identifying park risks based on location information, comprising the following steps:

[0085] S1. Through real-time multi-source data collection operations, obtain real-time flow location data, fixed risk source location data, equipment status data, environmental parameter data and real-time video stream data;

[0086] Said S1 comprises the following steps:

[0087] S11. Set fixed risk sources within the park to obtain a set of fixed risk sources within the park; based on the set of fixed risk sources within the park, collect the tag coordinates of the fixed risk sources to obtain the location data of the fixed risk sources (e.g., the center point of a storage tank);

[0088] Set the collection cycle; in each collection cycle, the UWB positioning base station deployed captures the coordinates of people and vehicles in real time to obtain real-time mobile location data;

[0089] S12: Read IoT sensor data (temperature, pressure, vibration) from devices within the park during each collection cycle to obtain device status data; collect environmental monitoring point data (temperature, humidity, gas concentration, wind speed and direction) to obtain environmental parameter data;

[0090] Call the camera to collect real-time video stream data;

[0091] S2: Divide the park into multiple grids, map the various data collected in S1 to the corresponding grids to obtain a grid dataset; align the data in the grid dataset to obtain an aligned grid dataset;

[0092] The S2 comprises the following steps:

[0093] S21. Use the Geohash algorithm to divide the park into 0.5m×0.5m grids (accuracy level 12) to obtain the park grid set;

[0094] Mapping the real-time flow location data, fixed risk source location data, equipment status data, environmental parameter data, and real-time video stream data to corresponding grids in the park grid set to obtain a grid data set; the grid data set contains the real-time flow location data, fixed risk source location data, equipment status data, environmental parameter data, and real-time video stream data for each area of ​​the park;

[0095] S22. Set a sliding time window (window size 5s, step size 1s), align the multi-source data in the grid dataset within the window, and obtain an aligned grid dataset; multi-source data stream (e.g., t = 12:00:00 ~ 12:00:05);

[0096] S3. Build a YOLOv8 model, improve the YOLOv8 model, and obtain an improved YOLOv8 model; use historical annotated video data combined with an optimization algorithm to train the improved YOLOv8 model and obtain the final YOLOv8 model;

[0097] Based on the aligned grid dataset and the final YOLOv8 model, abnormal behavior recognition in video streams was performed to obtain real-time abnormal behavior detection results in the park.

[0098] The S3 includes the following steps:

[0099] S31. Build a YOLOv8 model, improve the YOLOv8 model, and obtain an improved YOLOv8 model.

[0100] The S31 includes the following steps:

[0101] S311: Embed CBAM dual-channel attention (spatial + channel) in the Neck layer to improve the response to texture-blurred targets such as smoke / fire; add high-resolution feature output (160×160) in the P2 layer to improve the detection rate of small targets (such as people falling at a distance);

[0102] S312: Optimize the model specifically for recognizing small and unconventional targets (such as flames and smoke). Add a dedicated detection head for small and unconventional targets and adjust the anchor box size to match their characteristics. Strengthen the use of shallow high-resolution features in the feature pyramid and introduce attention mechanisms (such as CBAM) to significantly improve the detection rate of fuzzy textured targets such as smoke and the recognition accuracy of people falling.

[0103] S313: Embed hazard source location information (e.g., storage tank coordinates, high-voltage zone ranges) from the digital twin map of the park into the detection process. For identified targets (e.g., personnel and vehicles), calculate the Euclidean distance to the nearest hazard source in real time. When a target enters a high-risk area, dynamically increase its detection confidence (e.g., +50% confidence when a person is near high-temperature equipment) and mark it as high-risk. This spatial location constraint significantly reduces false detections and missed detections in complex environments.

[0104] S314. Model distillation technology is used to compress the model size (e.g., the number of parameters is compressed from 68M to 3.2M), ensuring real-time operation on edge devices. A dynamic resolution mechanism is introduced: routine monitoring uses low resolution (640x640) to maintain a high frame rate; when the system detects an increased risk value in an area (e.g., Risk > 0.6), it automatically switches to high-resolution mode (1280x1280) for detailed analysis of fire or leaks, achieving an on-demand balance between accuracy and efficiency.

[0105] S32. Collect historical video data; the historical video data includes normal situations and abnormal situations; the abnormal situations include people falling, objects falling, and fire;

[0106] Label the historical video data to obtain historical labeled video data; that is, mark the time and location of the abnormal situation in the video;

[0107] S33. Use the historically labeled video data to train the improved YOLOv8 model. During the training process, use an optimization algorithm to find the network parameters of the improved YOLOv8 model and obtain an optimal solution. The optimal solution is used as the network parameters of the improved YOLOv8 model to obtain a final YOLOv8 model.

[0108] In S33, the network parameters of the improved YOLOv8 model are found by combining the optimization algorithm to obtain the optimal solution, which includes the following steps:

[0109] S331. Set the recognition accuracy threshold of the improved YOLOv8 model to f1 and the recognition accuracy to f2; set the maximum number of iterations;

[0110] S332, construct a particle swarm, and set the size of the particle swarm to b according to the network parameters of the improved YOLOv8 model. Then the particle swarm is expressed as , where ui represents the i-th particle in the particle swarm;

[0111] S333. Construct a particle position set based on the particle swarm; define a fitness function for the particle positions in the particle swarm based on the recognition accuracy threshold f1 and the recognition accuracy f2. The fitness function formula is as follows:

[0112] ;

[0113] F represents the fitness function, l represents the bias of the fitness function, which is used to adjust the output of the fitness function;

[0114] S334, performing an iterative operation on the particle position set, wherein the higher the fitness value, the better the particle position; in each round of iteration, calculating the fitness values ​​of the particle positions in the particle position set according to the fitness function, updating the particle position set in descending order of fitness values, and obtaining the best particle position in the particle positions and the global best particle position in each round of iteration;

[0115] S3335, repeat S336, when the maximum number of optimization iterations is reached, stop the iteration, and take the global best particle position as the optimal solution;

[0116] S34. Input the real-time video stream of each grid in the aligned grid dataset into the final YOLOv8 model to obtain a real-time park abnormal behavior detection result; the real-time park abnormal behavior detection result includes the abnormal situation of the park location corresponding to each grid;

[0117] S4. Based on the aligned grid data set, calculate the park location entropy weighted risk value; superimpose the park location entropy weighted risk value with the equipment and environment data to obtain the superimposed risk value;

[0118] The S4 comprises the following steps:

[0119] S41. Based on the aligned grid dataset, for each grid area in the park, calculate the aggregation risk of mobile targets, and obtain the normalized real-time trajectory Shannon entropy (reflecting the randomness of movement), the normalized real-time speed variance within the area (abnormal acceleration / sudden stop), and the normalized real-time distance to the nearest hazard source;

[0120] The S41 includes the following steps:

[0121] S411, obtaining a trajectory coordinate sequence of a moving target within each grid based on the real-time flow position data of each grid in the aligned grid data set;

[0122] Each grid in the aligned grid dataset is divided into N subgrids to obtain the aligned grid dataset matrix. Based on the trajectory coordinate sequence of the moving object in each grid, the access frequency of each subgrid is counted as pn (n=1,2,...,N) to obtain the access frequency data.

[0123] According to the access frequency data, the normalized real-time trajectory Shannon entropy is calculated; the calculation formula is as follows:

[0124] ;

[0125] Where E represents the normalized real-time trajectory Shannon entropy, N represents the total number of subgrids, pn represents the frequency of each subgrid being visited, and log2N represents the normalization of the real-time trajectory Shannon entropy;

[0126] S412, obtaining the instantaneous speed of all targets in each grid by aligning the real-time video stream data of each grid in the grid data set;

[0127] According to the instantaneous speed of all targets in each grid, the normalized real-time area speed variance is calculated; the calculation formula is as follows:

[0128] ;

[0129] Where U represents the normalized real-time area speed variance, m represents the total number of targets in the grid, i represents the i-th target in the grid, vi represents the instantaneous speed of the i-th target in the grid, and vmax represents the maximum instantaneous speed of the target in the grid.

[0130] S413. Based on the central area of ​​each grid and the fixed risk source location data, calculate the real-time normalized distance to the nearest hazard source; the calculation formula is as follows:

[0131] ;

[0132] Where d represents the normalized distance to the nearest hazard source in real time, xr and yr represent the coordinates of the central area of ​​each grid, and xj and yj represent the coordinates of the fixed risk source respectively;

[0133] S42, respectively setting weight coefficients of the trajectory Shannon entropy, the velocity variance within the region, and the normalized distance to the nearest hazard source to obtain multiple weight coefficients;

[0134] Based on the normalized real-time trajectory Shannon entropy, the normalized real-time speed variance within the area, the normalized distance to the nearest hazard source in real time, and a combination of multiple weight coefficients, the entropy-weighted risk value of the park location is calculated; the calculation formula is as follows:

[0135] ;

[0136] Among them, g represents the entropy-weighted risk value of the park location, α1, α2, and α3 represent the weight coefficients of the trajectory Shannon entropy, the speed variance within the area, and the normalized distance to the nearest hazard source, respectively;

[0137] S44. Set environmental standard values, equipment standard values, and risk coefficients. Obtain an additional risk value based on the environmental standard values, equipment standard values, and the equipment status data and environmental parameter data in the aligned grid dataset. For example, if the entropy weighted risk value of the park location is 0.5, the detected equipment temperature is 120 degrees, the equipment standard temperature value is 100 degrees, and the risk coefficient is 0.8, then the additional risk value is (120 / 100)2*0.8=1.152, and the superimposed risk value is 1.662.

[0138] S5. Obtain the final risk result based on the superimposed risk value and the real-time park abnormal behavior detection results;

[0139] The S5 comprises the following steps:

[0140] S51. Setting an overlay risk threshold; when the overlay risk value of a grid area in the park grid set is greater than the overlay risk threshold and the real-time park abnormal behavior detection result is abnormal, the final risk result is to determine that the grid area is abnormal and issue an abnormal alarm;

[0141] When the superposition risk value of the grid area in the park grid set is less than the superposition risk threshold and the real-time park abnormal behavior detection result is abnormal, or when the superposition risk value of the grid area in the district grid set is greater than the superposition risk threshold and the real-time park abnormal behavior detection result is normal, a manual review alarm is issued, and the final risk result is obtained after manual review;

[0142] When the superposition risk value of a grid area in the park grid set is less than the superposition risk threshold and the real-time park abnormal behavior detection result is normal, the final risk result is to determine that the grid area is normal.

[0143] Example 2:

[0144] See also Figure 3 A location-based intelligent park risk identification system is used to implement the above-mentioned location-based intelligent park risk identification method, including a multi-source data real-time acquisition module, a grid data alignment module, a video abnormal behavior intelligent identification module, a dynamic risk superposition calculation module, and a hierarchical risk decision and alarm module;

[0145] The multi-source data real-time acquisition module dynamically captures multi-dimensional real-time data within the park through IoT devices; periodically obtains real-time flow location data of people / vehicles through UWB positioning base stations; collects tag coordinate data of fixed risk sources; reads device status data generated by device IoT sensors; collects environmental parameter data from environmental monitoring points; and calls cameras to obtain real-time video stream data.

[0146] The grid data alignment module uniformly maps the park space and multi-source data to a standardized grid system; uses the Geohash algorithm to divide the park into high-precision grids, and maps real-time location, risk source coordinates, equipment status, environmental parameters, and video stream data to the corresponding grids; and achieves temporal alignment of multi-source data through a sliding time window to generate a spatially and temporally consistent aligned grid dataset.

[0147] The video abnormal behavior intelligent recognition module is used to improve the YOLOv8 model to obtain an improved YOLOv8 model; and the improved YOLOv8 model is trained using historical annotated video data, combined with the particle swarm optimization algorithm to adjust parameters, and output grid-level real-time abnormal behavior detection results;

[0148] The dynamic risk overlay calculation module is used to integrate mobile trajectory, environment and equipment data to quantify regional risks. By calculating the normalized real-time Shannon entropy based on trajectory coordinates, the normalized intra-regional variance of target velocity, and the normalized distance to the nearest hazard source, the mobile target behavior risk index within the grid is obtained. The entropy weight method is combined with the weight coefficient to generate the position entropy weighted risk value. The equipment and environmental risks are then added to finally output the overlay risk value.

[0149] The hierarchical risk decision and alarm module formulates a response strategy based on the video analysis results and the superimposed risk value. If the grid area meets the superimposed risk value exceeding the threshold and the video detects abnormal behavior, it is directly determined to be abnormal and an alarm is triggered. If only a single condition is met, a manual review alarm is triggered, and the final risk status is confirmed after manual review. When both indicators are normal, the area is determined to be safe, realizing hierarchical intelligent decision-making.

[0150] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0151] The preferred embodiments of the invention disclosed above are intended only to help illustrate the invention. These preferred embodiments do not exhaust all details, nor do they limit the invention to the specific embodiments described. Obviously, many modifications and variations are possible based on the content of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for intelligently identifying park risks based on location information, characterized in that: The following steps are involved: S1. Through real-time multi-source data collection operations, obtain real-time flow location data, fixed risk source location data, equipment status data, environmental parameter data and real-time video stream data; S2, divide the park into multiple grids, map the various data collected in S1 to the corresponding grids, and obtain a grid data set; Align the data in the grid dataset to obtain an aligned grid dataset; S3. Build a YOLOv8 model, improve the YOLOv8 model, and obtain an improved YOLOv8 model; use historical annotated video data combined with an optimization algorithm to train the improved YOLOv8 model and obtain the final YOLOv8 model; Based on the aligned grid dataset and the final YOLOv8 model, abnormal behavior recognition in video streams was performed to obtain real-time abnormal behavior detection results in the park. In S3, the YOLOv8 model is improved to obtain an improved YOLOv8 model, which includes the following steps: Optimize the model's recognition of small and unconventional targets; Embed the hazard source location information from the park's digital twin map into the inspection process; Model distillation technology is used to compress the model volume, and a dynamic resolution mechanism is introduced; S4. Based on the aligned grid data set, calculate the park location entropy weighted risk value; superimpose the park location entropy weighted risk value with the equipment and environment data to obtain the superimposed risk value; S5. Based on the superimposed risk value and the real-time park abnormal behavior detection results, the final risk result is obtained.

2. The method for intelligently identifying park risks based on location information according to claim 1 is characterized in that: Said S1 comprises the following steps: S11. Set fixed risk sources within the park to obtain a set of fixed risk sources in the park; based on the set of fixed risk sources in the park, collect the tag coordinates of the fixed risk sources to obtain the location data of the fixed risk sources; The deployed UWB positioning base station captures the coordinates of people and vehicle tags in real time to obtain real-time mobile location data; S12. Read IoT sensor data from devices in the park to obtain device status data; collect environmental monitoring point data to obtain environmental parameter data; and call cameras to collect real-time video stream data.

3. The method for intelligently identifying park risks based on location information according to claim 1 is characterized in that: The S2 comprises the following steps: S21, using the Geohash algorithm to perform grid division on the park to obtain a park grid set; Mapping the real-time flow location data, fixed risk source location data, equipment status data, environmental parameter data, and real-time video stream data to corresponding grids in the park grid set to obtain a grid dataset; the grid dataset contains the real-time flow location data, fixed risk source location data, equipment status data, environmental parameter data, and real-time video stream data for each area of ​​the park; S22. Set a sliding time window, and align the multi-source data in the grid dataset within the window to obtain an aligned grid dataset.

4. The method for intelligently identifying park risks based on location information according to claim 1 is characterized in that: The S3 includes the following steps: S31. Build a YOLOv8 model, improve the YOLOv8 model, and obtain an improved YOLOv8 model. S32, collecting historical video data; the historical video data includes normal situations and abnormal situations; annotating the historical video data to obtain historical annotated video data; S33. Use the historically labeled video data to train the improved YOLOv8 model. During the training process, use an optimization algorithm to find the network parameters of the improved YOLOv8 model and obtain an optimal solution. The optimal solution is used as the network parameters of the improved YOLOv8 model to obtain a final YOLOv8 model. S34. Input the real-time video stream of each grid in the aligned grid dataset into the final YOLOv8 model to obtain real-time park abnormal behavior detection results; the real-time park abnormal behavior detection results include the abnormal situation of the park location corresponding to each grid.

5. The method for intelligently identifying park risks based on location information according to claim 4 is characterized in that: The S31 further comprises the following steps: A new dedicated detection head for small and unconventional targets is added, and the anchor box size is adjusted to match their characteristics. High-resolution feature output is added to the P2 layer. The utilization of shallow high-resolution features is strengthened in the feature pyramid, and CBAM dual-channel attention is embedded in the Neck layer. The location information of hazardous sources in the digital twin map of the park is embedded in the detection process. For each identified target, the Euclidean distance between it and the nearest hazardous source is calculated in real time. When a target enters a high-risk area, its detection confidence is dynamically increased and the target is marked as high-risk. Conventional monitoring uses low resolution to maintain a high frame rate; when the system detects an increase in regional risk, it automatically switches to high-resolution mode to analyze the anomaly in detail.

6. The method for intelligently identifying park risks based on location information according to claim 4 is characterized in that: In S33, the network parameters of the improved YOLOv8 model are found by combining the optimization algorithm to obtain the optimal solution, which includes the following steps: S331. Set the recognition accuracy threshold and recognition accuracy of the improved YOLOv8 model; set the maximum number of iterations; S332: Construct a particle swarm and set the size of the particle swarm according to the network parameters of the improved YOLOv8 model; S333. Construct a particle position set based on the particle swarm; define a fitness function of the particle positions in the particle swarm based on the recognition accuracy threshold and the recognition accuracy; S334, performing an iterative operation on the particle position set; in each round of iteration, calculating the fitness values ​​of the particle positions in the particle position set according to the fitness function, updating the particle position set in descending order of fitness values, and obtaining the best particle position in the particle positions and the global best particle position in each round of iteration; S335. Repeat S334. When the maximum number of optimization iterations is reached, stop the iteration and take the global best particle position as the optimal solution.

7. The method for intelligently identifying park risks based on location information according to claim 1 is characterized in that: The S4 comprises the following steps: S41. Based on the aligned grid dataset, for each grid area in the park, calculate the aggregation risk of the mobile target and obtain the normalized real-time trajectory Shannon entropy, the normalized real-time velocity variance within the area, and the normalized real-time distance to the nearest hazard source; S42, respectively setting weight coefficients of the trajectory Shannon entropy, the velocity variance within the region, and the normalized distance to the nearest hazard source to obtain multiple weight coefficients; The entropy-weighted risk value of the park location is calculated based on the normalized real-time trajectory Shannon entropy, the normalized real-time speed variance within the area, the normalized distance to the nearest hazard source, and multiple weight coefficients. S43. Setting environmental standard values, equipment standard values, and risk factors; obtaining additional risk values ​​based on the environmental standard values, equipment standard values, and equipment status data and environmental parameter data in the aligned grid data set.

8. The method for intelligently identifying park risks based on location information according to claim 7 is characterized in that: The S41 includes the following steps: S411, obtaining a trajectory coordinate sequence of a moving target within each grid based on the real-time flow position data of each grid in the aligned grid data set; Each grid in the aligned grid dataset is divided to obtain an aligned grid dataset matrix; based on the trajectory coordinate sequence of the moving target in each grid, the access frequency of each sub-grid is counted to obtain access frequency data; Based on the access frequency data, the normalized real-time trajectory Shannon entropy is calculated; S412, obtaining the instantaneous speed of all targets in each grid by aligning the real-time video stream data of each grid in the grid data set; According to the instantaneous speed of all targets in each grid, the normalized real-time regional speed variance is calculated; S413. Based on the central area of ​​each grid and the fixed risk source location data, calculate the real-time normalized distance to the nearest hazard source.

9. The method for intelligently identifying park risks based on location information according to claim 1, characterized in that: The S5 comprises the following steps: S51. Setting an overlay risk threshold; when the overlay risk value of a grid area in the park grid set is greater than the overlay risk threshold and the real-time park abnormal behavior detection result is abnormal, the final risk result is to determine that the grid area is abnormal and issue an abnormal alarm; When the superposition risk value of the grid area in the park grid set is less than the superposition risk threshold and the real-time park abnormal behavior detection result is abnormal, or when the superposition risk value of the grid area in the district grid set is greater than the superposition risk threshold and the real-time park abnormal behavior detection result is normal, a manual review alarm is issued, and the final risk result is obtained after manual review; When the superposition risk value of a grid area in the park grid set is less than the superposition risk threshold and the real-time park abnormal behavior detection result is normal, the final risk result is to determine that the grid area is normal.

10. A park risk intelligent identification system based on location information, characterized in that: A method for intelligently identifying park risks based on location information as described in any one of claims 1 to 9 is implemented, wherein the system includes a multi-source data real-time acquisition module, a grid data alignment module, a video abnormal behavior intelligent identification module, a dynamic risk superposition calculation module, and a hierarchical risk decision and alarm module.

Images