Smart park safety management method and system based on AI visual identification

Abstract

The invention relates to the technical field of safety monitoring, in particular to a smart park safety management method and system based on AI visual identification, and the method comprises the following steps: obtaining a video stream captured by a monitoring camera of a smart park, extracting the movement track data of a person through AI, and carrying out the statistics of the stay duration and access times of a target object in each grid; and constructing retention heat distribution of the grid region to obtain a personnel trajectory retention heat value. According to the invention, through the data processing of the monitoring video stream and the combination of the gridding analysis of the personnel motion trail, the accurate capture of the abnormal cruise behavior is enhanced, and through the calculation of the change rate of the access times between the adjacent grids, the statistics of the round-trip condition of the target among the plurality of areas and the combination of the passing authority data, the abnormal screening is carried out. The method improves the recognition capability of abnormal movement modes, achieves the dynamic adjustment of regional safety risk levels, improves the response capability of park security, enables the recognition capability of abnormal behaviors to be enhanced, and reduces the false alarm rate.

Description

Technical Field

[0001] The present invention relates to the technical field of security monitoring, and particularly to a smart park security management method and system based on AI visual recognition. Background Art

[0002] The technical field of security monitoring involves using means such as video monitoring, sensor detection, behavior analysis, and intelligent analysis algorithms to conduct real-time monitoring, data collection, anomaly detection, and early warning of personnel, vehicles, equipment, and the environment. This field encompasses technologies such as computer vision, pattern recognition, deep learning, image processing, the Internet of Things, and cloud computing to achieve security management of specific areas. Common applications include face recognition, target tracking, boundary crossing detection, abnormal behavior recognition, intrusion detection, fire warning, etc., and are widely used in scenarios such as public security, industrial security, traffic management, and intelligent buildings to improve security protection capabilities, reduce security risks, and optimize the emergency response mechanism.

[0003] Among them, the smart park security management method is used for security prevention and control within the park. By combining technologies such as video monitoring, Internet of Things sensing, and artificial intelligence analysis, it improves the monitoring accuracy of the status of personnel, vehicles, equipment, and the environment within the park. The uses of this method include automated security patrols, abnormal event alarms, personnel access rights management, fire warnings, security risk analysis, etc., to improve the intelligent level of park management, reduce the cost of manual inspections, enhance the response speed and handling ability of emergencies, and ensure the overall security of the park.

[0004] In terms of personnel trajectory detection, traditional management methods mainly rely on access statistics of fixed paths or areas, making it difficult to effectively depict dynamic cruising patterns, resulting in some abnormal movement behaviors being difficult to detect. In most cases, abnormal behavior analysis uses pattern matching based on a single feature, making it difficult to cover complex behavior patterns and resulting in problems of false positives or false negatives. For intrusion detection, most rely on static rule settings, lacking in-depth analysis of the dynamic movement characteristics of targets and making it difficult to distinguish normal stays from abnormal intrusions, resulting in a high false alarm rate. Security risk assessment is usually based on the judgment of single-point events and fails to make dynamic adjustments in combination with the security status of the area, resulting in a lag in the classification of security levels and being unable to reflect risk changes in a timely manner, leading to limitations in the smart park security management in dealing with complex abnormal behaviors, improving the accuracy of early warnings, and optimizing security level assessments, thus affecting the overall security protection ability. Summary of the Invention

[0005] The purpose of the present invention is to solve the drawbacks existing in the prior art, and to propose a smart park security management method and system based on AI visual recognition.

[0006] To achieve the above purpose, the present invention adopts the following technical solution: A smart park security management method based on AI visual recognition, including the following steps:

[0007] S1: Obtain the video stream captured by the surveillance cameras in the smart park, use AI to extract the personnel movement trajectory data, count the residence duration and access times of the target object in each grid, construct the residence heat distribution of the grid area, and obtain the personnel trajectory residence heat value;

[0008] S2: Based on the personnel trajectory residence heat value, count the round-trip times of the target object between multiple grid cells, screen the targets with abnormal residence time and abnormal increase in access times, and determine whether the trajectory pattern of the target conforms to the characteristics of abnormal cruising behavior to obtain the abnormal cruising behavior discrimination result;

[0009] S3: Based on the video stream captured by the surveillance cameras in the smart park, detect the target personnel entering the key control area of the park, calculate the angular change rate at adjacent moments, analyze the gait, upper limb swing pattern and the change trend of the trunk tilt angle to obtain the dynamic change information of the personnel joint angles;

[0010] S4: Invoke the dynamic change information of the personnel joint angles and the abnormal cruising behavior discrimination result, determine whether the target has the characteristics of jumping in, climbing and abnormal staying, compare the access records in the park security management database, screen the high-risk targets, calculate the risk assessment value of the intrusion behavior, and obtain the key intrusion behavior risk information of the park.

[0011] The improvements of the present invention are that the personnel trajectory residence heat value includes the grid residence duration distribution, the access frequency per unit time and the spatial heat probability parameter, the abnormal cruising behavior discrimination result is specifically the abnormally high-frequency moving target, the target with the trajectory round-trip frequency threshold exceeded and the target with abnormal short-term residence, the dynamic change information of the personnel joint angles includes the gait angle change rate, the deviation of the upper limb swing pattern and the change trend of the trunk tilt angle, and the key intrusion behavior risk information of the park is specifically the target with sudden change in intrusion angle, the target crossing the warning boundary and the target with risk of illegal residence.

[0012] The improvements of the present invention are that the obtaining steps of the personnel trajectory residence heat value are specifically:

[0013] S111: Obtain the video stream captured by the surveillance cameras in the smart park, use AI to identify the personnel movement trajectory data, and establish a personnel trajectory data set;

[0014] S112: Based on the personnel trajectory data set, divide the park entrances and exits, passages and public areas into grid cells of a fixed size, call the personnel trajectory data, count the cumulative residence duration and access times of the target object in each grid cell, and obtain the grid cell residence cumulative value;

[0015] S113: Call the grid cell residence cumulative value, combine the number of stays and the residence duration per unit time, and use the formula:

[0016]

[0017] Calculate to obtain the sojourn probability value of each grid area, construct the sojourn heat distribution of the grid area, and generate the sojourn heat value of the personnel trajectory;

[0018] Among them, P represents the sojourn heat value of the personnel trajectory, T represents the cumulative sojourn duration of the grid cell, F represents the cumulative access times of the grid cell, represents the average cumulative sojourn duration of the grid cell, represents the average cumulative access times of the grid cell, D i represents the continuous sojourn duration of the target object staying in the grid cell for the i-th time, represents the average value of the continuous sojourn duration in the grid cell, T total represents the total sum of the cumulative sojourn duration of all grid cells, F total represents the total sum of the cumulative access times of all grid cells, and n represents the number of times the target object stays in the grid cell.

[0019] The improvement of the present invention is that the step of obtaining the discrimination result of the abnormal cruising behavior is specifically as follows:

[0020] S211: Based on the sojourn heat value of the personnel trajectory, calculate the ratio of the difference in access times between adjacent grid cells to the cumulative access times to obtain the grid access times change rate;

[0021] S212: Based on the grid access times change rate, call the trajectory records of the target object in multiple grid cells, and count the cumulative number of times the target travels back and forth between multiple grid cells to obtain the grid round-trip access cumulative amount;

[0022] S213: Based on the grid round-trip access cumulative amount, call the data of the park entrance and exit management system, compare the park access permission of the target object with the actual trajectory behavior, and according to the change range of the sojourn duration in the grid cell and the grid access times change rate, use the formula:

[0023]

[0024] Calculate to obtain the deviation degree of the cruising behavior of the target object, and compare it with the park cruising behavior detection threshold to determine whether the trajectory pattern of the target conforms to the characteristics of abnormal cruising behavior, and generate the discrimination result of abnormal cruising behavior;

[0025] Among them, S represents the deviation degree of the cruising behavior of the target object, V S represents the grid access times change rate, W S represents the change range of the sojourn duration, represents the average grid access times change rate of all targets, Represents the change range of the average residence time of all targets, R S Represents the total number of cumulative cruises of the target object, A S Represents the number of authorized access areas in the park for the target object, Q S Represents the actual number of cruise grid cells of the target object.

[0026] The improvement of the present invention is that the steps for obtaining the dynamic change information of the joint angles of the personnel are specifically as follows:

[0027] S311: Based on the video stream captured by the intelligent park surveillance camera, detect the target personnel entering the key control area of the park, extract the contour information of the target object, calibrate the central point coordinates of the target, determine whether the target is within the range of the control area, screen out the invalid targets outside the monitoring range, and obtain the information of the target personnel in the key control area;

[0028] S312: Based on the information of the target personnel in the key control area, extract the key point data of the target personnel, including the position coordinates of the shoulder, elbow, knee, and ankle joints, calculate the Euclidean distance between adjacent key points, calibrate the key point coordinates, and normalize them to the standard human body proportion model to obtain the standardized key point coordinate data;

[0029] S313: Based on the standardized key point coordinate data, calculate the angular change rate at adjacent times, statistically analyze the swing amplitude of the upper limb during the gait cycle, extract the gait parameters, including the swing time, support time, and the change range of the trunk tilt angle, and use the formula:

[0030]

[0031] Calculate the dynamic change rate of the gait angle to obtain the dynamic change information of the joint angles of the personnel;

[0032] Among them, B t Represents the dynamic change rate of the joint angle at time t, K represents the total number of selected joints, θ t,k Represents the angle value of joint k at time t, v t,k Represents the angular velocity of joint k at time t.

[0033] The improvement of the present invention is that the steps for obtaining the risk information of key intrusion behaviors in the park are specifically as follows:

[0034] S411: Call the dynamic change information of the joint angles of the personnel and the discriminant result of abnormal cruise behavior, compare with the set boundary intrusion angle change threshold, extract the sequence of pose angle changes of the target in the boundary area, and combine with the angle change rate at adjacent time points to screen the candidate targets with features of jumping in, climbing, and abnormal staying, and obtain the abnormal features of the boundary intrusion angle;

[0035] S412: Based on the abnormal boundary intrusion angle feature, combined with the discriminant result of abnormal cruising behavior, compare the passing records in the park security management database, and use the formula:

[0036]

[0037] Calculate the passing anomaly score of the target, compare it with the anomaly screening threshold, and obtain the screening result of high-risk targets;

[0038] Among them, S t represents the passing anomaly score of the target at time t, A t represents the cumulative passing times of the target at time t, A m represents the average passing times within the time window, W t represents the anomaly weight at time t, A s represents the standard deviation of the passing times within the time window, λ represents the attenuation coefficient of the change of passing anomaly, and e is the base of the natural logarithm;

[0039] S413: Based on the screening result of high-risk targets, use the passing anomaly score of the target and the abnormal boundary intrusion angle feature for risk weighting, combined with the distribution of intrusion events in the park, calculate the intrusion risk index of the current target, and obtain the risk information of key intrusion behaviors in the park.

[0040] The improvement of the present invention is that the method further includes:

[0041] S5: Call the discriminant result of abnormal cruising behavior and the risk information of key intrusion behaviors in the park, extract the required security weights for each area, combine the security event records and target identity levels of each area, calculate the risk level value of each area, compare it with the set trigger threshold, and notify the security management personnel to obtain the security trigger execution information;

[0042] The security trigger execution information includes the enhanced area of park monitoring, the adjustment parameter of illegal intrusion sensitivity, and the optimization factor of security patrol path.

[0043] The improvement of the present invention is that the acquisition steps of the security trigger execution information are specifically as follows:

[0044] S511: Call the discriminant result of abnormal cruising behavior and the risk information of key intrusion behaviors in the park, extract the security weight parameters of each area, count the behavior patterns of the target in the area, combine the security event records of the area, calculate the risk deviation value of the target in the area, and obtain the area risk deviation coefficient;

[0045] S512: Based on the area risk deviation coefficient, combined with the identity level of the target, use the formula:

[0046]

[0047] Calculate the risk level value of the target in the area and obtain the area risk level evaluation value;

[0048] Among them, R s represents the risk level value of the target in the area, W r represents the safety weight parameter of this area, S t represents the passing anomaly score of the target at time t, I t represents the intrusion behavior risk value of the target at time t, I m represents the average intrusion risk level of the area, L R represents the identity level of the target;

[0049] S513: Based on the area risk level evaluation value, compare with the set trigger threshold, judge whether the notification condition is met, screen the abnormal area information to be pushed to the security management personnel, calculate the urgency of the target security event, and obtain the security trigger execution information.

[0050] The intelligent park security management system based on AI visual recognition is used to execute the above-mentioned intelligent park security management method based on AI visual recognition. The system includes:

[0051] The trajectory stay analysis module obtains the video stream captured by the monitoring cameras in the intelligent park, uses AI to extract the personnel movement trajectory data, constructs the stay heat distribution of the grid area, and obtains the personnel trajectory stay heat value;

[0052] The abnormal cruise evaluation module judges whether the trajectory pattern of the target conforms to the characteristics of abnormal cruise behavior based on the personnel trajectory stay heat value, and obtains the abnormal cruise behavior discrimination result;

[0053] The joint angle monitoring module detects the target personnel entering the key control area of the park based on the video stream captured by the monitoring cameras in the intelligent park, analyzes the gait, the change trend of the upper limb swing pattern and the trunk tilt angle, and obtains the dynamic change information of the personnel joint angle;

[0054] The behavior risk assessment module calls the dynamic change information of the personnel joint angle and the abnormal cruise behavior discrimination result, judges whether the target has features such as jumping in, climbing, and abnormal staying, calculates the intrusion behavior risk assessment value, and obtains the key intrusion behavior risk information of the park;

[0055] The security management execution module calls the abnormal cruise behavior discrimination result and the key intrusion behavior risk information of the park, calculates the risk level value of each area, notifies the security management personnel, and obtains the security trigger execution information.

[0056] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0057] In the present invention, through data processing of the monitored video stream and combined with grid analysis of the movement trajectories of personnel, the quantification of residence time, access times, and heat distribution is realized, enhancing the accurate capture of abnormal cruising behaviors. By calculating the change rate of access times between adjacent grids, the round-trip situation of the target among multiple regions is statistically analyzed, and combined with access permission data for abnormal screening, the recognition ability of abnormal movement patterns is improved. In terms of behavior monitoring in key areas, key point detection and dynamic angle change calculation are adopted to obtain gait, upper limb swing, and torso tilt data, accurately evaluate the abnormal movement characteristics of the target, and combined with cruising behavior analysis, potential intrusion behaviors are screened, enhancing the accuracy of security risk assessment. Based on multi-dimensional data fusion, security events, target identity levels, and risk assessment results within the region are calculated to realize dynamic adjustment of the regional security risk level, improve the response ability of park security, enhance the recognition ability of abnormal behaviors, reduce the false alarm rate, improve the effectiveness of intrusion warnings, and strengthen the timeliness of security event handling. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is the flowchart of the method of the present invention;

[0059] Figure 2 is the flowchart of obtaining the residence heat value of the personnel trajectory in the present invention;

[0060] Figure 3 is the flowchart of obtaining the discrimination result of abnormal cruising behavior in the present invention;

[0061] Figure 4 is the flowchart of obtaining the dynamic change information of the joint angles of personnel in the present invention;

[0062] Figure 5 is the flowchart of obtaining the risk information of key intrusion behaviors in the park in the present invention;

[0063] Figure 6 is the flowchart of obtaining the security trigger execution information in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0064] In order to make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0065] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation to the present invention. In addition, in the description of the present invention, the meaning of "a plurality of" is two or more, unless otherwise specifically defined.

[0066] Please refer to Figure 1 , the present invention provides a technical solution: a smart campus security management method based on AI visual recognition, including the following steps:

[0067] S1: Obtain the video stream captured by the surveillance cameras in the smart campus, use AI to extract the personnel movement trajectory data, divide the campus entrances, exits, passages, and public areas into grid units of a fixed size, count the residence duration and access times of the target object in each grid, calculate the residence probability per unit time, construct the residence heat distribution of the grid area, and obtain the residence heat value of the personnel trajectory.

[0068] S2: Based on the residence heat value of the personnel trajectory, calculate the change rate of access times between adjacent grids, count the number of round trips of the target object between multiple grid units, call the data of the smart campus entrance and exit management system, compare the personnel access permissions, calculate the change range of the residence time, screen the targets with abnormal residence time and abnormal increase in access times, set the detection threshold for abnormal patrol behavior in the campus, and judge whether the trajectory pattern of the target conforms to the characteristics of abnormal patrol behavior to obtain the discrimination result of abnormal patrol behavior.

[0069] S3: Based on the video stream captured by the surveillance cameras in the smart campus, detect the target personnel entering the key control area of the campus, extract the key point data, including the positions of the shoulder, elbow, knee, and ankle joints, calculate the angular change rate between adjacent moments, analyze the gait, upper limb swing pattern, and the change trend of the trunk tilt angle to obtain the dynamic change information of the personnel joint angles.

[0070] S4: Call the dynamic change information of the personnel joint angles and the discrimination result of abnormal patrol behavior, compare the set boundary intrusion angle change threshold, judge whether the target has the characteristics of jumping in, climbing, and abnormal staying, combine the discrimination result of abnormal patrol behavior, compare the access records in the campus security management database, screen the high-risk targets, calculate the risk assessment value of the intrusion behavior, and obtain the risk information of key intrusion behaviors in the campus.

[0071] S5: Invoke the discriminant result of abnormal cruise behavior and the risk information of key intrusion behaviors in the park, extract the required security weights for each area, combine the security event records and target identity levels of each area, calculate the risk level value for each area, compare it with the set trigger threshold, notify the security management personnel, and obtain the security trigger execution information;

[0072] The personnel trajectory retention heat value includes the grid residence duration distribution, the access frequency per unit time, and the spatial heat probability parameter. The discriminant result of abnormal cruise behavior is specifically an abnormally high-frequency moving target, a target with an excessive trajectory round-trip frequency threshold, and a target with abnormal short-term retention. The dynamic change information of the personnel joint angle includes the gait angle change rate, the deviation of the upper limb swing pattern, and the change trend of the trunk tilt angle. The risk information of key intrusion behaviors in the park is specifically a target with a sudden change in intrusion angle, a target crossing the warning boundary, and a target with a risk of illegal retention. The security trigger execution information includes the enhanced monitoring area in the park, the adjustment parameter of the illegal intrusion sensitivity, and the optimization factor of the security patrol path.

[0073] Please refer to Figure 2 , and the specific steps for obtaining the personnel trajectory retention heat value are as follows:

[0074] S111: Obtain the video stream captured by the intelligent park monitoring camera, use AI to identify the personnel movement trajectory data, and establish a personnel trajectory data set;

[0075] The intelligent park video stream obtained by the monitoring camera captures frames of the picture through a real-time acquisition device. The number of frames obtained per second is set according to the device performance, and the typical value is 30fps. After the camera captures the picture, the system invokes the AI model to perform personnel detection. The detection method is based on a pre-trained deep learning model, which can adopt the YOLOv5 or FasterR-CNN architecture. The AI model performs target recognition on each frame of the image, extracts the personnel features in the image, including the human body contour, walking direction, posture, etc. Subsequently, the system assigns a unique identifier to the detected personnel features, invokes the optical flow method or the DeepSORT algorithm for multi-target tracking. Whenever a target person enters the camera's field of view, the system assigns a unique number to it, calculates the position offset based on its movement trajectory within consecutive frames, combines the time interval between the previous and subsequent frames to obtain the personnel movement speed, and forms time series data. The time series data records the spatial coordinate values of the personnel at specific time points, which are obtained through monocular camera calibration or binocular camera depth ranging. For a single target, the perspective transformation parameters are used to convert the image pixel coordinates into actual physical coordinates, while the binocular camera uses parallax to calculate the spatial depth information. The data storage format adopts a five-tuple of (timestamp, personnel number, coordinate X, coordinate Y, speed V), and each five-tuple data is stored as a trajectory point in the trajectory data set. For the problem of target loss, a continuous loss frame threshold N is set. lossto determine whether a person has left the monitored area, and this threshold is set based on the walking speed V of the person person and the monitoring range R of the camera cam The specific calculation method is where T frame is the frame interval time. Set the monitoring range R cam = 5m, the walking speed V of the person person = 1.2m / s, the frame interval time T frame = 0.033s (30fps), then frames. That is, if a person with a certain number is not detected within 126 frames, it is determined that they have left the monitored area, and the recording of their trajectory points is stopped, and finally a complete personnel trajectory dataset is established.

[0076] S112: Based on the personnel trajectory dataset, divide the park entrances, exits, channels, and public areas into grid cells of a fixed size, call the personnel trajectory data, count the cumulative stay duration and the number of visits of the target object in each grid cell, and obtain the grid cell stay cumulative value;

[0077] Based on the personnel trajectory dataset, the system divides the entrances, exits, channels, and public areas in the park into grids. The size of the grid cells depends on the scene requirements and is commonly set to 1m×1m or 0.5m×0.5m. After division, the system traverses all trajectory points, calculates the cumulative stay duration and the number of visits in each grid cell. The traversal method uses the region query method. Set the spatial range of a certain grid cell as [x1, x2]×[y1, y2]. For each trajectory point (timestamp, personnel number, coordinate X, coordinate Y), determine whether the coordinate falls within the grid space range. If the condition is met, the number of visits to this grid is accumulated, and the stay duration is calculated. The calculation method of the stay duration is the difference between adjacent timestamps. If the consecutive trajectory points of the same numbered person are still within this grid, the stay time is continuously accumulated. If the position of adjacent trajectory points exceeds the current grid, the stay accumulation of this numbered person is stopped. If the person enters this grid again, the number of visits is re-counted. During the calculation of the stay time, to avoid the influence of abnormally long stays on the statistical results, set the upper limit threshold T max of the single stay time. This threshold is calculated based on the 95th percentile in the historical data. For example, statistically analyze the distribution of all single stay times and take its 95th percentile as T max Set the existing sample data as 12s, 15s, 18s, 20s, 25s, 30s, 50s, 120s, 200s, 600s, and the calculation result of the 95th percentile is T max = 120s. The single stay time exceeding 120s will be truncated to 120s. Finally, the cumulative stay duration and the number of visits of all grid cells are obtained, as shown in Table 1.

[0078] Table 1 Grid Cell Residence Statistics Table

[0079] Grid Number Number of Visits F (times) Total Stay Duration T (seconds) 1 50 600 2 30 400 3 80 1200

[0080] As shown in Table 1, the number of visits and residence durations of different grid cells are different, and the statistical results are used to calculate the personnel retention heat value.

[0081] S113: Call the cumulative residence value of the grid cell, combine the number of stays and residence durations within a unit time, and use the formula:

[0082]

[0083] Calculate the residence probability value of each grid area through operations, construct the retention heat distribution of the grid area, and generate the personnel trajectory retention heat value;

[0084] Among them, P represents the personnel trajectory retention heat value, T represents the cumulative residence duration of the grid cell, F represents the cumulative number of visits to the grid cell, represents the average cumulative residence duration of the grid cell, represents the average cumulative number of visits to the grid cell, D i represents the continuous residence duration of the target object's i-th stay in the grid cell, represents the average value of the continuous residence durations within the grid cell, T total represents the total sum of the cumulative residence durations of all grid cells, F total represents the total sum of the cumulative number of visits to all grid cells, and n represents the number of stays of the target object within the grid cell;

[0085] Call the residence statistical value of the grid cell, calculate the retention heat value, and set the total sum T total of the cumulative residence durations and the total sum F total of the cumulative number of visits, and the calculation method is:

[0086]

[0087] where m is the total number of grids, and calculate the average cumulative residence duration T and the average cumulative number of visits F of each grid cell:

[0088]

[0089] Calculate the retention heat value P of each grid:

[0090]

[0091] where D i represents the continuous duration of the i-th stay of the target object, represents the average continuous residence duration of the grid cell, and the data in Table 2 is used for calculation.

[0092] Table 2 Parameter Calculation Example Table

[0093] Grid Number T (seconds) F (times) <![CDATA[D1 (seconds)]]> <![CDATA[D2 (seconds)]]> <![CDATA[D3 (seconds)]]> 1 600 50 12 15 18 2 400 30 8 10 12 3 1200 80 20 22 25

[0094] Calculated based on Grid 1 Substitute into the formula:

[0095]

[0096] Calculate Grids 2 and 3 to obtain the complete heat retention distribution. The results show that the heat retention of personnel in different areas is different, which can be used for intelligent decision-making to adjust the area planning in the future.

[0097] Please refer to Figure 3 , and the specific steps for obtaining the discriminant results of abnormal cruising behavior are as follows:

[0098] S211: Based on the heat retention value of the personnel trajectory, calculate the ratio of the difference in the number of visits between adjacent grid cells to the cumulative number of visits to obtain the grid visit rate of change;

[0099] Based on the heat retention value of the personnel trajectory, the system calculates the number of visits to adjacent grid cells. First, obtain the number of visits F of each grid cell i and record the number of visits F of adjacent grid cells j , calculate the difference in the number of visits between the two by taking the difference ΔF = |F i -F j |, then obtain the cumulative number of visits ratio or where the larger value is taken as the standard ratio. If the number of visits to a certain grid cell is zero, the ratio is set to the maximum threshold, such as 10 or higher, to prevent data abnormal overflow. Subsequently, traverse all grid cells, calculate the change in the number of visits to their adjacent cells, and organize them in a table format, as shown in Table 3. The setting basis of the maximum threshold of 10 is to reasonably control the ratio of the number of visits to grid cells. If the value is too large, it may cause abnormal points to affect the overall statistical results. Therefore, a fixed value upper limit is set so that the ratio will not be infinitely amplified. This value is set according to the normal distribution of the number of visits to a single grid in the park. For example, if the number of visits to most grids is between 30 and 100 times, the maximum ratio is set to 10 to prevent the statistical data from deviating abnormally.

[0100] Table 3 Grid Visit Rate of Change Calculation Table

[0101]

[0102] As shown in Table 3, the ratio and difference in the number of visits to different grid cells are used for subsequent behavior analysis.

[0103] S212: Based on the change rate of grid access times, call the trajectory records of the target object in multiple grid cells, count the cumulative number of times the target travels between multiple grid cells, and obtain the cumulative grid round-trip access volume;

[0104] Based on the change rate of grid access times, the system calls the trajectory data of the target object, records the access situation of the target object between multiple grids, defines the round-trip behavior of each target object between adjacent grid cells, calculates its round-trip access times, sets that when the target object continuously accesses two grids at least twice within a certain time window, it is recognized as one round-trip, counts all the round-trip paths of the target object, obtains the cumulative round-trip volume, and records it in the format of target object number, starting grid number, ending grid number, and round-trip times, as shown in Table 4. The setting of the time window is based on the rationality of the target object frequently entering and leaving different grids in a short period of time. Usually, the average residence time of personnel activities in the park can be obtained through historical statistics. If the statistical data shows that the average residence time of most targets in a certain area is 5 minutes, the time window can be set to 5 minutes to ensure that the trajectory data within this time range is representative.

[0105] Table 4 Cumulative Table of Target Object Grid Round-Trip Access

[0106] Target Number Starting Grid Ending Grid Number of Round Trips A1 1 2 5 A2 2 3 8 A3 1 3 3

[0107] As shown in Table 4, the round-trip times of the target object between multiple grid cells are different, and this data is used for the subsequent calculation of the deviation degree of the cruising behavior.

[0108] S213: Based on the cumulative grid round-trip access volume, call the data of the park entrance and exit management system, compare the park access permission of the target object with the actual trajectory behavior, and according to the change range of the residence time in the grid cell and the change rate of the grid access times, use the formula:

[0109]

[0110] Calculate to obtain the deviation degree of the cruising behavior of the target object, compare it with the detection threshold of the park cruising behavior, judge whether the trajectory pattern of the target conforms to the characteristics of abnormal cruising behavior, and generate a discriminant result of abnormal cruising behavior;

[0111] Among them, S represents the deviation degree of the cruising behavior of the target object, V S represents the change rate of grid access times, W S represents the change range of residence time, represents the average change rate of grid access times of all targets, represents the average change range of residence time of all targets, R S represents the total cumulative cruising times of the target object, A SRepresents the number of authorized access areas in the target object's park, Q S Represents the actual number of patrolled grid cells of the target object;

[0112] Based on the cumulative round-trip access volume of the grid, the system calls the data of the park entrance and exit management system, extracts the authorized access areas of the target object, and obtains the total number of authorized access grid cells A in the park S , and at the same time counts the actual number of patrolled grid cells Q of the target object S , calculates the deviation degree of the target object's patrolling behavior. In the calculation process, it is necessary to first obtain the cumulative number of patrols R of the target object S , calculates the change rate V of the grid access times S and the change range W of the stay duration S , and then calculates the average change rate of the grid access times of all targets and the average change range of the stay duration Finally, substitute into the formula:

[0113]

[0114] Taking the target object A1 as an example, set its number of patrols R S = 10, the change rate V of the grid access times S = 1.67, the change range W of the stay duration S = 1.2, the average change rate of the grid access times of all targets The average change range of the stay duration The number of authorized access grid cells A in the target object's park S = 5, the actual number of patrolled grid cells Q S = 8, the calculation is as follows:

[0115]

[0116] Among them, log(R S ) reflects the patrolling frequency of the target object. The calculation of this item is based on the statistical data of the historical patrolling times of different targets. If the number of patrols is less than 5 times, then log(R S ) is close to 0. If the number of patrols is relatively high, then this value is close to 2. Generally, for targets with short-time and high-frequency patrolling, their number of patrols may be greater than 10 times. Therefore, the value range of this item is generally between [0, 2], and It represents the average statistical value of all targets. When calculating, it is necessary to ensure that the sample size is sufficient to reduce the influence of extreme values. The specific value can be obtained by taking the average of the data of all targets. The result shows that the deviation degree of the cruising behavior of target object A1 is 0.0653. Subsequently, abnormal behavior can be determined according to the cruising behavior detection threshold. Among them, the setting basis of the cruising behavior detection threshold lies in the safety requirements of the park. If the park has relatively strict requirements for cruising behavior, a lower threshold can be set, such as 0.05. If more free cruising behavior is allowed, the threshold can be set above 0.1. In practical applications, this threshold is usually set after statistical analysis of historical normal cruising data.

[0117] Please refer to Figure 4 , and the steps for obtaining the dynamic change information of the joint angles of personnel are specifically as follows:

[0118] S311: Based on the video stream captured by the intelligent park monitoring camera, detect the target personnel entering the key control area of the park, extract the contour information of the target object, calibrate the center point coordinates of the target, determine whether the target is within the range of the control area, screen out the invalid targets outside the monitoring range, and obtain the information of the target personnel in the key control area;

[0119] Based on the video stream captured by the intelligent park monitoring camera, continuously obtain image frames and perform target detection on each frame. First, call the frame decomposition process and set the time interval (0.2 seconds or 0.5 seconds, depending on the frame rate of the monitoring camera. If the camera frame rate is 30 FPS, the 0.2 - second interval corresponds to 6 frames, and the 0.5 - second interval corresponds to 15 frames, ensuring smooth trajectories and no data redundancy). In each frame, use morphological operations to remove background interference and only retain the moving target area. Extract the outer contour of the target personnel through the contour detection algorithm and use the centroid calculation method to obtain the center point coordinates of the target object. Set the boundary coordinate range of the control area and store it in the form of an array of polygon area coordinates. Call the center point coordinates of the target object and calculate whether it falls within the coordinate range of the control area. If it is not within the range, directly screen out the target. If it is within the range, record the information of the target personnel. In the specific screening process, use the region judgment function to match the coordinates of the target object. For example, assume the control area is a rectangle with the upper - left coordinate (10, 10) and the lower - right coordinate (100, 100), and the center point coordinates of the target object are (50, 50). The judgment is as follows:

[0120] 10 ≤ 50 ≤ 100, 10 ≤ 50 ≤ 100;

[0121] This target meets the region determination condition, so record the position information of this target. If the center point coordinates of a certain target are (150, 80), the judgment is as follows:

[0122] 10 ≤ 150 ≤ 100 (not satisfied);

[0123] The target is screened out, and finally the information of the target personnel within the key control area is obtained. The core threshold for screening invalid targets is the boundary coordinate value of the control area. The setting of this value is based on the actual physical boundary of the park planning map, such as the entrance and exit, the fence or the demarcation of the functional area. The boundary value is usually provided by the park management system and calibrated through the GIS system or manual measurement data. Its range determines whether the target is valid.

[0124] S312: Based on the information of the target personnel within the key control area, extract the key point data of the target personnel, including the position coordinates of the shoulder, elbow, knee and ankle joints, calculate the Euclidean distance between adjacent key points, calibrate the key point coordinates, and normalize them to the standard human proportion model to obtain the standardized key point coordinate data;

[0125] Based on the information of the target personnel within the key control area, extract the key point data of the target personnel. In the video stream, the detection of the skeletal key points of each target object adopts the human pose estimation method. For the target object detected in each frame of the image, identify the pixel coordinates of its shoulder, elbow, knee and ankle, and calculate the Euclidean distance between adjacent key points. Let the shoulder coordinates be (x1, y1) and the elbow coordinates be (x2, y2). The Euclidean distance from the shoulder to the elbow is calculated as follows:

[0126]

[0127] Perform the same calculation for all key points, and further calibrate the key point coordinates. Based on the proportion of the human bone model, normalize the coordinates of each key point to the standard human proportion model. For example, if the shoulder-knee distance of the actual human body is 90 cm and the pixel distance measured in the video image is 180 pixels, the normalization ratio is calculated as follows:

[0128]

[0129] Perform the normalization calculation for all key point coordinates to obtain the standardized key point coordinate data. The scale factor in the normalization process depends on the standard human bone model, and its reference value can be set based on the average value of human anatomy. For example, the shoulder width of an adult is between 40 cm and 50 cm, and the elbow-knee distance is between 85 cm and 95 cm. When calculating the scale factor, the average value is usually selected for conversion. The closer the scale factor is to 1, the less the original data needs to be scaled, and the farther it is from 1, the greater the scale adjustment required.

[0130] S313: Based on the standardized key point coordinate data, calculate the angular change rate at adjacent times, statistically analyze the swing amplitude of the upper limb during the gait cycle, and extract the gait parameters, including the swing time, support time and the range of change of the trunk tilt angle, using the formula:

[0131]

[0132] Calculate the dynamic change rate of the gait angle to obtain the dynamic change information of the joint angles of the person;

[0133] Among them, B t represents the dynamic change rate of the joint angle at time t, K represents the total number of selected joints, and θ t,k represents the angle value of joint k at time t, and v t,k represents the angular velocity of joint k at time t;

[0134] Based on the standardized key point coordinate data, calculate the angle change rate at adjacent times, obtain the amplitude of the upper limb swing during the gait cycle, call the gait cycle division rule, set the gait time window (such as 1 second or 2 seconds, this time window is set according to the conventional walking gait cycle, generally the gait cycle is between 0.8 seconds and 1.2 seconds, and the time window should cover at least one complete gait cycle to ensure accurate calculation of the swing amplitude), calculate the swing time, support time and the change range of the trunk tilt angle during the gait cycle. During the calculation of gait parameters, for the key joint points, obtain the angle changes at their adjacent time points. Let the angle of a certain joint k at time t be θ t,k , and the angle at time t - 1 be θ t-1,k , calculate the angle change rate as follows:

[0135]

[0136] Among them, let K = 4 (i.e., the shoulder, elbow, knee, and ankle), and set the gait data of a person as follows: θ t,1 = 30°, θ t-1,1 = 25°; θ t,2 = 45°, θ t-1,2 = 40°; θ t,3 = 60°, θ t-1,3 = 55°; θ t,4 = 50°, θ t-1,4 = 48°;

[0137] Calculate the angle change:

[0138]

[0139] Set the angular velocity change: v t,1 = 5° / s, v t-1,1 = 4° / s; v t,2 = 6° / s, v t-1,2 = 5° / s; v t,3 = 7° / s, v t-1,3 = 6° / s; v t,4 = 5° / s, v t-1,4 = 4° / s;

[0140] Calculate the sum of squares of speed changes:

[0141]

[0142] Calculate the dynamic change rate of the final gait angle:

[0143] B t = 4.25 + 2 = 6.25;

[0144] This result indicates that the dynamic change rate of the gait angle of the target person is 6.25, which can be used for gait feature analysis in the future. Among them, the setting of the gait time window is based on the human walking frequency data. The walking frequency range is between 1.2 Hz and 1.5 Hz, and the corresponding gait cycle is about 0.8 seconds to 1.2 seconds. Therefore, the time window should cover at least 1.5 times the maximum gait cycle, that is, 1.2 seconds × 1.5 = 1.8 seconds, and the integer value is set to 2 seconds to ensure the integrity of the gait cycle characteristics and avoid data loss affecting the dynamic angle calculation.

[0145] Please refer to Figure 5 , the specific steps for obtaining the risk information of key intrusion behaviors in the park are as follows:

[0146] S411: Invoke the dynamic change information of the joint angles of the personnel and the discrimination results of abnormal cruising behaviors, compare with the set boundary intrusion angle change threshold, extract the sequence of posture angle changes of the target in the boundary area, and combine the angle change rate at adjacent time points to screen the candidate targets with features of jumping in, climbing, and abnormal staying, and obtain the abnormal features of the boundary intrusion angle;

[0147] Call the dynamic change information of the joint angles of the personnel and the discrimination result of abnormal cruising behavior, extract the sequence of attitude angle changes of the target in the boundary area. First, select the joint angle change data within the boundary area from the historical gait data, set a time interval (such as 0.5 seconds or 1 second) to extract the key joint points (shoulder, elbow, knee, ankle) of the target object, calculate the angle change rate between adjacent time points. The setting of the boundary intrusion angle change threshold is based on the common angle change range in human kinematics research, referring to the joint angle change rate within the normal gait cycle. Generally, the angle change rate of the shoulder and elbow during normal walking is within the range of 10° / s to 20° / s, while during a jumping-in or climbing action, this angle change rate can reach 30° / s or even higher. Therefore, set the boundary intrusion angle change threshold to 15° / s to 30° / s, and the specific value can be adjusted according to the historical abnormal behavior data of the monitoring area. Within the set time window (such as 5 seconds or 10 seconds), count the trend of the attitude angle change of the target in the boundary area. By comparing the set boundary intrusion angle change threshold, screen out the targets with a large angle change rate in a short period of time. If the target shows a large-angle rapid change in a short period of time (such as the trunk forward inclination angle changes from 15° to 60° within 0.5 seconds), then mark this target as a jumping-in candidate target. If the angle change of the target at low-height joints (such as the knee and ankle) is abnormal (such as the knee angle changes from 30° to 120°), then mark this target as a climbing candidate target. If the angle change rate of the target between adjacent time points is close to zero and the duration in the boundary area exceeds the set threshold (such as 10 seconds), then mark this target as an abnormal stay candidate target. Finally, obtain the abnormal characteristics of the boundary intrusion angle.

[0148] S412: Based on the abnormal characteristics of the boundary intrusion angle, combined with the discrimination result of abnormal cruising behavior, compare the access records in the park security management database, and use the formula:

[0149]

[0150] Calculate the access anomaly score of the target, compare it with the anomaly screening threshold, and obtain the screening result of high-risk targets;

[0151] Among them, S t represents the access anomaly score of the target at time t, A t represents the cumulative access times of the target at time t, A m represents the average access times within the time window, W t represents the anomaly weight at time t, A s represents the standard deviation of the access times within the time window, λ represents the attenuation coefficient of the access anomaly change, and e is the base of the natural logarithm;

[0152] Based on the abnormal characteristics from the perspective of boundary intrusion, combined with the discriminant results of abnormal cruising behavior, compare the access records in the park security management database, set a time window (such as 30 minutes or 1 hour) to extract the historical access times A of the target t , calculate the average access times A within the time window m and the standard deviation of access times A s , call the abnormal cruising behavior weight W of the target t , combined with the attenuation coefficient λ of the abnormal change in access, calculate the access anomaly score of the target, and the calculation formula is as follows:

[0153]

[0154] Among them, the selection of the time window refers to the time span of high-frequency access behavior in the park. For example, in an office park, people mainly access during the morning rush hour (8:00 - 10:00) and the evening rush hour (17:00 - 19:00), so the time window is often set to 1 hour. In low-frequency access scenarios (such as at night), the time window can be extended to 2 hours. Set the parameters of a certain target as follows: A t = 12 (the access times of the target within the current time window); A m = 5 (the average access times of people in the park within this time window); A s = 2 (the standard deviation of access times in the park within this time window) - the abnormal cruising behavior weight W t The setting of the weight W is based on the trajectory pattern of the target in the park. If the target frequently crosses multiple grids in a short time and has a low matching degree with the authorized area of the park, a higher weight (such as 1.5 to 2.0) is given. If the target only repeats back and forth within some grids, a lower weight (such as 0.8 to 1.2) is given. The current target's trajectory behavior shows that it cruises between multiple grids, and the weight is set to 1.2; - the attenuation coefficient λ of the abnormal change in access is set within the range of 0.3 to 0.7. This value depends on the persistence of the abnormal behavior. Generally speaking, the sudden abnormal access behavior in a short time has little impact on λ (such as 0.3), while the long-term abnormal access behavior increases its attenuation coefficient (such as 0.7). The current target's access behavior suddenly increases in a short time, so λ = 0.5.

[0155] Calculate the access anomaly score of the target:

[0156]

[0157] Set the abnormal screening threshold to 2.0. This threshold is set by comparing the score distribution of historical normal access behavior, usually set as the mean of the access anomaly scores of normal people plus 1.5 times the standard deviation. If the target's access anomaly score exceeds this threshold, it is determined as a high-risk target. The current target's access anomaly score is lower than the abnormal screening threshold, so it is not determined as a high-risk target.

[0158] S413: Based on the screening results of high-risk targets, use the passage anomaly score of the target and the abnormal feature of the boundary intrusion angle for risk weighting. Combine the distribution of intrusion events in the park to calculate the intrusion risk index of the current target, and obtain the risk information of key intrusion behaviors in the park;

[0159] Based on the screening results of high-risk targets, call the passage anomaly score of the target and the abnormal feature of the boundary intrusion angle, set the risk weight parameters, and assign different weights to different types of abnormal behaviors. Among them, the weight of abnormal cruising behavior is set to 0.6, and the weight of abnormal boundary intrusion angle is set to 0.4. The setting of this weight is based on the impact degree of the behavior type on the park security. Abnormal cruising behavior usually involves long-term trajectory anomalies, so a higher weight (such as 0.6 to 0.8) is assigned. While the abnormal boundary intrusion angle focuses on short-term sudden behaviors, so the weight is relatively low (such as 0.3 to 0.5). Combine the distribution of historical intrusion events in the park to calculate the intrusion risk index of the current target. Set the passage anomaly score of the target as S t = 1.92, and the score of the abnormal feature of the boundary intrusion angle is B t = 3.5. The risk weighting formula is as follows:

[0160] R = 0.6×S t + 0.4×B t = 1.152 + 1.4 = 2.552;

[0161] The setting of the intrusion risk index threshold is based on the risk score distribution of past intrusion events in the park. Usually, the 95% quantile is taken as the abnormal determination threshold. In a general park environment, this threshold is set between 2.5 and 3.0. In this example, it is set to 2.5. The intrusion risk index of the target is higher than the threshold, so this target is marked as an individual at risk of key intrusion behaviors in the park, and finally the risk information of key intrusion behaviors in the park is generated.

[0162] Please refer to Figure 6 , and the specific steps for obtaining the security trigger execution information are as follows:

[0163] S511: Call the discrimination results of abnormal cruising behavior and the risk information of key intrusion behaviors in the park, extract the security weight parameters of each area, count the behavior patterns of the target in the area, combine the security event records of the area, calculate the risk deviation value of the target in the area, and obtain the regional risk deviation coefficient;

[0164] Call the discrimination results of abnormal cruising behavior and the risk information of key intrusion behaviors in the park, extract the security weight parameters of each area, count the behavior patterns of the target in this area. First, obtain the regional historical security event records, set the time window (such as 1 day or 7 days), count the number of abnormal behaviors in the area, calculate the abnormal behavior frequency, and set the security weight parameter Wr , this parameter is set according to the occurrence ratio of historical abnormal events in the area, and the specific calculation method is where N 异常 is the cumulative number of abnormal behaviors that have occurred in this area, and N 总访问 is the total number of personnel visits to this area. If there are 50 abnormal events in a certain area in the past 7 days and the total number of visits is 2000, then when this value is greater than 0.05, this area is defined as a high-risk area, and the weight parameter is set to the upper limit of 1.5. If it is less than 0.01, it is defined as a low-risk area, and the weight parameter is set to 0.8. It increases linearly with the increase of the abnormal event ratio. The behavior patterns of the target in this area are statistically analyzed, including the staying duration, the number of round trips, and the boundary approaching frequency. The discriminant result of the abnormal cruising behavior of this target is called, and the behavior pattern determination rules are set. For example, if the staying time of the target in this area is significantly higher than the average staying time of the personnel in the area, it is determined that it has an abnormal staying behavior. If the number of round trips of the target exceeds the set threshold (such as twice the area average value), it is determined that it has an abnormal round-trip behavior. Calculate the risk deviation value of the target in the area, and set the historical average intrusion risk level I of the area m , calculate the risk level I of the target within the current time window t , and finally obtain the area risk deviation coefficient.

[0165] S512: Based on the area risk deviation coefficient, combined with the identity level of the target, use the formula:

[0166]

[0167] Calculate the risk level value of the target in the area and obtain the area risk level evaluation value;

[0168] Among them, R s represents the risk level value of the target in the area, W r represents the security weight parameter of this area, S t represents the passage abnormal score of the target at time t, I t represents the intrusion behavior risk value of the target at time t, I m represents the average intrusion risk level of the area, L R represents the identity level of the target;

[0169] Based on the area risk deviation coefficient, combined with the identity level of the target, calculate the risk level value of the target in the area. First, obtain the passage abnormal score S of the target t , call the area security weight parameter W r , set the identity level value L R , the identity level L RBased on the park management authority settings, the value usually ranges from 1 to 5. Among them, ordinary visitors are set to 1, ordinary employees are set to 2, personnel in core departments are set to 3, senior management is set to 4, and specially approved personnel with the highest security permissions are set to 5. This value is directly related to the access permissions and passage ranges of the target within the park. The higher the identity level value, the wider the passage permissions of the target within the park and the fewer behavioral restrictions it is subject to. The calculation formula is as follows:

[0170]

[0171] Set the parameters of a certain target as follows: W r = 1.5, and the security weight parameter of this area is calculated based on the proportion of historical abnormal events; S t = 2.0, the passage anomaly score of the target, which is calculated by comparing the abnormal cruising behavior of the target within the park with the historical passage pattern; I t = 3.2, the current intrusion behavior risk value of the target, which is calculated based on the abnormal cruising behavior weight of the target and the boundary abnormal behavior characteristics; I m = 2.5, the average intrusion risk level of this area, which is derived from the historical occurrence frequency of intrusion events in this area; L R = 2, the identity level of this target, which is set according to the management authority of the target in the park;

[0172] Calculate the risk level value of the target:

[0173]

[0174] Finally, the obtained regional risk level assessment value of this target is 1.85.

[0175] S513: Based on the regional risk level assessment value, compare the set trigger threshold, determine whether the notification condition is met, screen the abnormal area information to be pushed to the security management personnel, calculate the urgency of the target security event, and obtain the security trigger execution information;

[0176] Based on the regional risk level assessment value, compare the set trigger threshold, determine whether the security management personnel notification condition is met, and set the trigger threshold R threshold (This value is usually set according to the park security management level. For ordinary office areas, it is set to 2.5, for important facility areas, it is set to 2.0, and for key security control areas, it is set to 1.8). If the regional risk level assessment value of the target is higher than this threshold, it is determined that the target belongs to a high-risk individual, screen the abnormal area information to be pushed to the security management personnel, calculate the urgency of the target security event, and set the urgency classification rule. For example, if R s ≤ 1.5, it is a low risk, 1.5 < R s

[0177] If it is ≤ 2.0, it is a medium risk, R s If it is > 2.0, it is a high risk. The threshold is set based on the statistics of the urgency of security incidents in the park. Usually, similar types of security incidents in the past year are counted, the risk level distribution is calculated, and a reasonable alarm trigger point is set. For example, if the risk level value of a target is 1.85, it is determined as a medium risk and does not meet the trigger condition, so no abnormal warning is pushed. If the risk level value of the target is 2.2, it meets the trigger condition, the security trigger execution information of the target is calculated, and finally the security management push data of the target is generated.

[0178] The intelligent park security management system based on AI visual recognition is used to execute the above-mentioned intelligent park security management method based on AI visual recognition. The system includes:

[0179] The trajectory retention analysis module obtains the video stream captured by the monitoring cameras in the intelligent park, uses AI to extract the personnel movement trajectory data, counts the stay duration and access times of the target object in each grid, constructs the retention heat distribution of the grid area, and obtains the personnel trajectory retention heat value;

[0180] The abnormal cruise evaluation module, based on the personnel trajectory retention heat value, counts the round-trip times of the target object between multiple grid units, screens the targets with abnormal retention time and abnormal increase in access times, and judges whether the trajectory pattern of the target conforms to the characteristics of abnormal cruise behavior, and obtains the discriminant result of abnormal cruise behavior;

[0181] The joint angle monitoring module, based on the video stream captured by the monitoring cameras in the intelligent park, detects the target personnel entering the key control area of the park, calculates the angular change rate at adjacent moments, analyzes the gait, the swing pattern of the upper limbs and the change trend of the trunk tilt angle, and obtains the dynamic change information of the personnel joint angle;

[0182] The behavior risk assessment module calls the dynamic change information of the personnel joint angle and the discriminant result of abnormal cruise behavior, judges whether the target has features such as jumping in, climbing and abnormal staying, compares the access records in the park security management database, screens high-risk targets, calculates the invasion behavior risk assessment value, and obtains the key invasion behavior risk information in the park;

[0183] The security management execution module calls the discriminant result of abnormal cruise behavior and the key invasion behavior risk information in the park, extracts the required security weights for each area, combines the security event records and the target identity level of each area, calculates the risk level value of each area, compares it with the set trigger threshold, and notifies the security management personnel to obtain the security trigger execution information.

[0184] The above are only the preferred embodiments of the present invention, and do not impose other forms of limitations on the present invention. Any person skilled in the relevant art may use the technical content disclosed above to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A smart park security management method based on AI visual recognition, characterized in that, It includes the following steps: S1: Obtain the video stream captured by the surveillance cameras in the smart park, use AI to extract the personnel movement trajectory data, count the stay duration and access times of the target object in each grid, construct the retention heat distribution of the grid area, and obtain the personnel trajectory retention heat value; S2: Based on the personnel trajectory retention heat value, count the round-trip times of the target object between multiple grid cells, screen the targets with abnormal retention time and abnormal increase in access times, and judge whether the trajectory pattern of the target conforms to the characteristics of abnormal cruising behavior to obtain the abnormal cruising behavior discrimination result; S3: Based on the video stream captured by the surveillance cameras in the smart park, detect the target personnel entering the key control area of the park, calculate the angular change rate at adjacent moments, and analyze the gait, upper limb swing pattern and trunk tilt angle change trend to obtain the dynamic change information of the personnel joint angles; S4: Call the dynamic change information of the personnel joint angles and the abnormal cruising behavior discrimination result, judge whether the target has features of jumping in, climbing and abnormal staying, compare the access records in the park security management database, screen the high-risk targets, calculate the risk assessment value of the intrusion behavior, and obtain the key intrusion behavior risk information of the park.

2. The intelligent park security management method based on AI visual recognition according to claim 1, wherein, The personnel trajectory retention heat value includes the grid stay duration distribution, the access frequency per unit time and the spatial heat probability parameter. The abnormal cruising behavior discrimination result is specifically the abnormally high-frequency moving target, the target with the trajectory round-trip frequency threshold exceeded, and the target with abnormal short-term retention. The dynamic change information of the personnel joint angles includes the gait angle change rate, the deviation of the upper limb swing pattern and the trunk tilt angle change trend. The key intrusion behavior risk information of the park is specifically the target with sudden change in intrusion angle, the target crossing the warning boundary and the target with risk of illegal retention.

3. The intelligent park security management method based on AI visual recognition according to claim 1, characterized in that, The specific steps for obtaining the personnel trajectory retention heat value are as follows: S111: Obtain the video stream captured by the surveillance cameras in the smart park, use AI to identify the personnel movement trajectory data, and establish a personnel trajectory data set; S112: Based on the personnel trajectory data set, divide the park entrances and exits, passages, and public areas into grid cells of a fixed size, call the personnel trajectory data, count the cumulative stay duration and access times of the target object in each grid cell, and obtain the grid cell stay cumulative value; S113: Call the grid cell stay cumulative value, combine the number of stays and the stay duration per unit time, and use the formula: Perform the operation to obtain the stay probability value of each grid area, construct the retention heat distribution of the grid area, and generate the personnel trajectory retention heat value; Among them, P represents the heat value of the personnel trajectory stay, T represents the cumulative stay duration of the grid cell, F represents the cumulative access times of the grid cell, represents the average cumulative stay duration of the grid cell, represents the average cumulative access times of the grid cell, D i represents the continuous stay duration of the target object's i-th stay in the grid cell, represents the average value of the continuous stay duration in the grid cell, T total represents the total sum of the cumulative stay durations of all grid cells, F total represents the total sum of the cumulative access times of all grid cells, and n represents the number of stays of the target object in the grid cell.

4. The intelligent park security management method based on AI visual recognition according to claim 1, characterized in that The specific steps for obtaining the abnormal cruising behavior discrimination result are as follows: S211: Based on the personnel trajectory retention heat value, calculate the ratio of the difference in access times between adjacent grid cells to the cumulative access times to obtain the grid access times change rate; S212: Based on the grid access times change rate, call the trajectory records of the target object in multiple grid cells, count the cumulative number of times the target travels back and forth between multiple grid cells, and obtain the grid round-trip access cumulative amount; S213: Based on the cumulative amount of round-trip access in the grid, call the data of the park entrance and exit management system, compare the park access permission of the target object with the actual trajectory behavior, and according to the change range of the staying duration in the grid cell and the change rate of the grid access times, use the formula: Calculate the deviation degree of the cruising behavior of the target object, compare it with the park cruising behavior detection threshold, judge whether the trajectory pattern of the target conforms to the characteristics of abnormal cruising behavior, and generate the discrimination result of abnormal cruising behavior; Among them, S represents the deviation degree of the target object's cruising behavior, V S represents the change rate of the number of grid visits, W S represents the change range of the staying duration, represents the change rate of the average number of grid visits of all targets, represents the change range of the average staying duration of all targets, R S represents the total cumulative number of cruises of the target object, A S represents the number of authorized access areas in the park for the target object, Q S represents the actual number of grid cells cruised by the target object.

5. The intelligent park security management method based on AI visual recognition according to claim 1, wherein, The specific steps for obtaining the dynamic change information of the joint angles of the personnel are as follows: S311: Based on the video stream captured by the intelligent park monitoring camera, detect the target personnel entering the key control area of the park, extract the contour information of the target object, calibrate the center point coordinates of the target, judge whether the target is within the monitoring range of the control area, screen out the invalid targets beyond the monitoring range, and obtain the information of the target personnel in the key control area; S312: Based on the information of the target personnel in the key control area, extract the key point data of the target personnel, including the position coordinates of the shoulder, elbow, knee and ankle joints, calculate the Euclidean distance between adjacent key points, calibrate the key point coordinates, and normalize them to the standard human body proportion model to obtain the standardized key point coordinate data; S313: Based on the standardized key point coordinate data, calculate the angular change rate at adjacent times, count the swing amplitude of the upper limb within the gait cycle, extract the gait parameters, including the swing time, support time and the change range of the trunk tilt angle, and use the formula: Calculate the dynamic change rate of the gait angle to obtain the dynamic change information of the joint angles of the personnel; Among them, B t represents the dynamic change rate of the joint angle at time t, K represents the total number of selected joints, and θ t,k represents the angle value of joint k at time t, and v t,k represents the angular velocity of joint k at time t.

6. The intelligent park security management method based on AI visual recognition according to claim 1, characterized in that, The specific steps for obtaining the risk information of key intrusion behaviors in the park are as follows: S411: Call the dynamic change information of the joint angles of the personnel and the discrimination result of abnormal cruising behavior, compare with the set boundary intrusion angle change threshold, extract the sequence of pose angle changes of the target in the boundary area, and combine with the angle change rate at adjacent time points to screen the candidate targets with features of jumping in, climbing and abnormal staying, and obtain the abnormal features of the boundary intrusion angle; S412: Based on the abnormal features of the boundary intrusion angle, combine with the discrimination result of abnormal cruising behavior, compare with the access records in the park security management database, and use the formula: Calculate the access abnormality score of the target, compare with the abnormal screening threshold, and obtain the screening result of high-risk targets; Among them, S t represents the passing anomaly score of the target at time t, A t represents the cumulative passing times of the target at time t, A m represents the average passing times within the time window, W t represents the anomaly weight at time t, A s represents the standard deviation of the passing times within the time window, λ represents the decay coefficient of the passing anomaly change, and e is the base of the natural logarithm; S413: Based on the screening result of high-risk targets, use the access abnormality score of the target and the abnormal features of the boundary intrusion angle for risk weighting, and combine with the distribution of intrusion events in the park to calculate the intrusion risk index of the current target to obtain the risk information of key intrusion behaviors in the park.

7. The intelligent park security management method based on AI visual recognition according to claim 1, characterized in that, The method further includes: S5: Call the discrimination result of abnormal cruising behavior and the risk information of key intrusion behaviors in the park, extract the required security weights for each area, combine with the security event records and the target identity level of each area, calculate the risk level value of each area, compare with the set trigger threshold, and notify the security management personnel to obtain the security trigger execution information; The security trigger execution information includes the enhanced area of park monitoring, the adjustment parameter of illegal intrusion sensitivity and the optimization factor of the security patrol path.

8. The intelligent park security management method based on AI visual recognition according to claim 7, characterized in that The specific steps for obtaining the security trigger execution information are as follows: S511: Invoke the abnormal patrol behavior discrimination result and the risk information of key intrusion behaviors in the park, extract the security weight parameters of each area, count the behavior patterns of the target in the area, combine the security event records of the area, calculate the risk deviation value of the target in the area, and obtain the area risk deviation coefficient; S512: Based on the area risk deviation coefficient, combined with the identity level of the target, use the formula: Calculate the risk level value of the target in the area and obtain the area risk level evaluation value; Among them, R s represents the risk level value of the target in the area, W r represents the safety weight parameter of the area, S t represents the passage anomaly score of the target at time t, I t represents the risk value of the target's intrusion behavior at time t, I m represents the average intrusion risk level of the area, L R represents the identity level of the target; S513: Based on the area risk level evaluation value, compare the set trigger threshold, determine whether the notification condition is met, filter the abnormal area information to be pushed to the security management personnel, calculate the urgency of the target security event, and obtain the security trigger execution information.

9. The intelligent park security management system based on AI visual recognition is characterized in that, According to the intelligent park security management method based on AI visual recognition according to any one of claims 1-8, the system includes: The trajectory stay analysis module obtains the video stream captured by the monitoring cameras in the intelligent park, uses AI to extract the personnel movement trajectory data, constructs the stay heat distribution of the grid area, and obtains the personnel trajectory stay heat value; The abnormal patrol evaluation module determines whether the trajectory pattern of the target conforms to the characteristics of abnormal patrol behavior based on the personnel trajectory stay heat value, and obtains the abnormal patrol behavior discrimination result; The joint angle monitoring module detects the target personnel entering the key control area of the park based on the video stream captured by the monitoring cameras in the intelligent park, analyzes the gait, upper limb swing pattern and the change trend of the trunk tilt angle, and obtains the dynamic change information of the personnel joint angle; The behavior risk assessment module invokes the dynamic change information of the personnel joint angle and the abnormal patrol behavior discrimination result, determines whether the target has features such as jumping in, climbing, and abnormal staying, calculates the risk assessment value of the intrusion behavior, and obtains the risk information of key intrusion behaviors in the park; The security management execution module invokes the abnormal patrol behavior discrimination result and the risk information of key intrusion behaviors in the park, calculates the risk level value of each area, notifies the security management personnel, and obtains the security trigger execution information.

Images