Intelligent park security early warning system and method based on artificial intelligence

Abstract

The invention discloses a smart park security early warning system and method based on artificial intelligence, and relates to the technical field of early warning systems, the system comprises a security center platform, and the security center platform is in communication connection with a data acquisition module, an intelligent analysis module, an early warning decision module, a data transmission module, a data processing module and an execution feedback module. According to the invention, a multi-type video acquisition assembly, a distributed environment perception assembly and a multi-mode identity verification assembly are integrated through a data acquisition module, and park dynamic image information, environment parameter information and personnel biological characteristics and identification information are respectively acquired, so that a park security basic data set covering three dimensions of personnel, objects and environment is formed; the comprehensiveness and the dimensional integrity of park security data acquisition are improved, and the problems that blind areas exist in security monitoring and compound potential safety hazards cannot be accurately recognized due to incomplete data coverage of an existing system can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of early warning systems, in particular to an intelligent park security early warning system and method based on artificial intelligence. BACKGROUND

[0002] The park security early warning system can monitor personnel, vehicles and environment in the park in real time through the deployment of devices, combined with algorithm and big data analysis technology, automatically identify abnormal behavior and trigger early warning, and integrate access control, patrol scheduling and emergency response functions, realize early perception, accurate positioning and rapid disposal of park security risks, effectively improve the intelligence and automation level of park security management, and build a comprehensive protection barrier for personnel, property and equipment safety in the park.

[0003] At present, the park security early warning system has the problems of focusing on single dimension in data collection and being one-sided in data type, resulting in blind spots in security monitoring and inability to accurately identify complex security risks, and only single behavior recognition or simple risk judgment can be achieved, lacking predictive and systematic analysis, so that the system cannot predict potential security incidents in advance and can only respond after the event, falling into a passive security situation.

[0004] Therefore, the present application proposes an intelligent park security early warning system and method based on artificial intelligence to solve the above problems. SUMMARY

[0005] The main purpose of the present application is to provide an intelligent park security early warning system and method based on artificial intelligence to solve the problems raised in the background.

[0006] To achieve the above purpose, the technical solution adopted by the present application is: an intelligent park security early warning system and method based on artificial intelligence, comprising a security central platform, the security central platform being communicatively connected with a data collection module, an intelligent analysis module, an early warning decision module, a data transmission module, a data processing module and an execution feedback module; The data collection module is used to obtain park security data, obtain dynamic image information of personnel and objects in the park through multi-type video collection components, obtain temperature and humidity, smoke concentration and gas composition information in the park through distributed environment perception components, obtain biological feature information and identification information of personnel entering the park through multi-modal identity verification components, and integrate various collected information to form a park security basic data set; The intelligent analysis module is used to analyze the park security basic data set, analyze the motion trajectory and behavior pattern of personnel and objects through an intelligent behavior recognition component, judge potential security incidents in combination with historical security data and real-time data through a security event prediction component, determine the degree of park security risk through a dynamic risk assessment component to build an evaluation index system, and generate a park security analysis result; The early warning decision module is used for formulating a disposal scheme according to the park security analysis result, determining a risk triggering condition by combining a preset standard and real-time analysis data through a dynamic threshold judgment component, dividing an early warning category and pushing early warning information according to a risk degree through a multi-level early warning release component, formulating a response scheme by combining an early warning level and historical disposal cases through an emergency plan generation component, and outputting a park security early warning decision result; The data transmission module is used for transmitting data among modules, and realizes stable transmission of security data among modules by a combination of a double-link wired transmission component and an edge enhanced wireless transmission component; The data processing module is used for processing security data, screening and correcting abnormal items in the data through an intelligent data cleaning component, classifying and saving the processed data through a classified storage component, and searching the stored data through a search component to form a security database; The execution feedback module is used for implementing the early warning decision result and feeding back the execution situation, taking emergency measures according to the response scheme through an emergency response component, allocating park security resources through a dynamic resource scheduling component, and evaluating the event disposal situation through a disposal effect evaluation component to form a closed-loop disposal process.

[0007] Preferably, the multi-type video acquisition component includes a wide-angle panoramic camera, a low-illumination night vision camera, and a close-up behavior camera, the wide-angle panoramic camera covers the public activity area of the park, the low-illumination night vision camera is deployed in the park area without auxiliary lighting at night, and the close-up behavior camera is deployed at the entrance and equipment room of the park; The multi-type video acquisition component processes the collected continuous image frames through a space-time feature fusion algorithm, extracts the spatial position features and time motion features of the personnel and objects, establishes a dynamic tracking link of the personnel and objects, and records the moving path of the personnel and objects in the park.

[0008] Preferably, the space-time feature fusion algorithm includes a spatial feature extraction layer, a time feature processing layer, and a feature fusion layer, the spatial feature extraction layer extracts the contour and size features of the personnel and objects in the image frames by using a convolutional neural network, the time feature processing layer analyzes the motion direction and speed features of the personnel and objects in the continuous image frames by using a recurrent neural network, and the feature fusion layer performs weight distribution on the spatial features and time features by using an attention mechanism to form a dynamic feature vector of the personnel and objects.

[0009] Preferably, the distributed environment perception component includes a temperature and humidity sensor, an optical smoke sensor, and an electrochemical gas sensor, the temperature and humidity sensor is deployed at different height positions indoors and outdoors of the park, the optical smoke sensor is deployed on the top of the corridor and equipment room of the park, and the electrochemical gas sensor is deployed around the underground garage and laboratory of the park; The distributed environmental perception component integrates the data collected by each sensor through a multi-source data fusion algorithm, calculates the mean environmental parameters of different areas in the park, and establishes a regional environmental parameter matrix. When any parameter in the regional environmental parameter matrix deviates from the normal parameter matrix of the same period in history by more than a preset range, the area is marked as an environmental abnormality area.

[0010] Preferably, the multimodal identity verification component integrates a visible light face recognition device, a capacitive fingerprint collection device and an RFID reading device, wherein the visible light face recognition device collects facial feature point information of the person, the capacitive fingerprint collection device collects fingerprint texture information of the person, and the RFID reading device collects identity code information in the identification card carried by the person; The multimodal identity verification component matches the collected facial feature point information, fingerprint texture information and identity coding information through a feature fusion authentication algorithm. When the matching degree of the three types of information reaches the preset standard, the person's identity authentication is completed. At the same time, the person's activity area within the park is delineated based on the authority data associated with the identity coding information.

[0011] Preferably, the intelligent behavior recognition component includes a trajectory analysis unit, a behavior classification unit and an abnormality determination unit. The trajectory analysis unit smoothes the motion trajectory of people and objects through the Kalman filtering algorithm and removes noise interference in the trajectory. The behavior classification unit uses a deep learning model to divide people's behavior into normal activities, wandering and climbing over obstacles. The abnormality determination unit compares the current behavior category with the historical normal behavior category library of the area. When the current behavior category does not appear in the historical normal behavior category library, the behavior is marked as abnormal behavior and the time and location of the abnormal behavior are recorded.

[0012] Preferably, the security incident prediction component includes a data preprocessing unit, a model training unit and a prediction execution unit. The data preprocessing unit normalizes the historical security incident data and removes redundant information in the data. The model training unit uses a spatiotemporal attention long-short-term memory network to train the preprocessed historical data and establish a security incident prediction model. The prediction execution unit inputs real-time campus status data into the security incident prediction model and outputs the types of security incidents that may occur within a preset time period in the future and the corresponding probability values.

[0013] Preferably, the dynamic threshold determination component includes a historical data statistics unit, a threshold calculation unit and a threshold update unit. The historical data statistics unit collects statistics on the park security data within the past preset period and calculates the fluctuation range of various security parameters. The threshold calculation unit uses a normal distribution algorithm to determine the initial risk threshold of various parameters based on the fluctuation range. The threshold update unit recalculates the risk threshold based on the latest statistical security data after each preset period to achieve dynamic adjustment of the risk threshold and ensure that the risk threshold matches the actual security status of the park.

[0014] Preferably, the multi-level early warning issuing component comprises an early warning level division unit, an information generation unit and a pushing unit, the early warning level division unit divides the early warning level into four categories of general early warning, relatively heavy early warning, serious early warning and extremely serious early warning, the information generation unit generates early warning text information containing early warning area, early warning type and risk degree according to the early warning level, and the pushing unit synchronously pushes the early warning text information through three channels of park broadcasting system, mobile terminal APP and indoor display screen, wherein the extremely serious early warning information additionally triggers the sound and light alarm device of the park.

[0015] Preferably, the intelligent park security early warning method based on artificial intelligence comprises the following steps: S1, through a multi-type video acquisition component, a distributed environment perception component and a multi-modal identity verification component, park dynamic image information, environment parameter information and personnel identity information are respectively acquired to form a park security basic data set; S2, through a double-link wired transmission component and an edge-enhanced wireless transmission component, the park security basic data set is transmitted to a data processing module, and the data is encrypted during transmission; S3, through an intelligent data cleaning component, abnormal data in the park security basic data set is screened and corrected, through a classified storage component, the corrected data is respectively stored in an image database, an environment database and a personnel database, and through a retrieval component, indexes are established for each database; S4, an intelligent behavior recognition component is called to analyze dynamic image data in the image database to identify abnormal behaviors, a security event prediction component is called to predict potential security events in combination with historical security data and real-time environment data, a dynamic risk assessment component is called to determine the current risk degree of the park to generate a park security analysis result; S5, through a dynamic threshold determination component, risk trigger conditions are determined according to the park security analysis result, through a multi-level early warning issuing component, early warning levels are divided and early warning information is pushed, through an emergency plan generation component, corresponding emergency disposal schemes are formulated, and an early warning decision result is output; S6, according to the early warning decision result, through an emergency response component, emergency measures are taken, through a dynamic resource scheduling component, security resources are allocated, and through a disposal effect evaluation component, the disposal process is evaluated, and the evaluation result is fed back to the security hub platform.

[0016] The present application has the following advantages: 1.The present application integrates multiple types of video acquisition components, distributed environmental perception components and multi-modal identity verification components through a data acquisition module to obtain park dynamic image information, environmental parameter information, and personnel biological characteristics and identification information, forming a park security basic data set covering personnel, objects and environment in three dimensions. Compared with the existing technology of multi-focal single dimension data acquisition and the scheme of data type being one-sided, the comprehensiveness and dimension integrity of park security data acquisition can be improved, so as to solve the problem of blind area in security monitoring caused by incomplete data coverage in the existing system, and the problem of inability to accurately identify complex security risks.

[0017] 2.The intelligent analysis module in the present application cooperates with the intelligent behavior recognition component, the security event prediction component and the dynamic risk assessment component. The intelligent behavior recognition component smoothes the trajectory based on the Kalman filter algorithm and classifies the behavior. The security event prediction component predicts potential events by combining historical and real-time data using a spatiotemporal attention long short-term memory network. The dynamic risk assessment component constructs an index system to determine the risk level, forming a multi-level security data analysis result. Compared with the existing technology of analyzing module which can only realize single behavior recognition or simple risk judgment, and lacks of predictive and systematic analysis, the depth and foresight of park security analysis can be improved, so as to solve the problem of passive security of the existing system which cannot predict potential security events in advance and can only respond afterwards.

[0018] 3.The dynamic threshold determination component, multi-level warning release component and emergency plan generation component in the early warning decision module of the present application are linked. The dynamic threshold determination component determines the initial threshold based on the normal distribution algorithm and updates it periodically. The multi-level warning release component synchronously pushes warning information through three channels and triggers sound and light alarm for particularly serious warning. The emergency plan generation component formulates a scheme combining warning level and historical cases, and outputs precise warning decision result. Compared with the existing technology of early warning decision which adopts fixed threshold, single warning channel and general emergency plan, the adaptability, information reach efficiency and scheme pertinence of early warning decision can be improved, so as to solve the problems of high false alarm rate, untimely transmission of warning information and inability of emergency plan to quickly match the actual scene in the existing system. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 It is the architecture diagram of the intelligent park security early warning system based on artificial intelligence of the present application. Figure 2 It is the process diagram of the intelligent park security early warning method based on artificial intelligence of the present application. DETAILED DESCRIPTION

[0020] In order to make the technical means, creative features, purposes and effects of the present application easy to understand, the present application is further described below in combination with specific embodiments.

[0021] Example 1, please refer to Figure 1 and Figure 2 As shown: The smart park security early warning system based on artificial intelligence includes a security central platform, which is connected to a data acquisition module, an intelligent analysis module, an early warning decision module, a data transmission module, a data processing module, and an execution feedback module; The data acquisition module is used to obtain park security data. It uses multiple video acquisition components to obtain dynamic image information of people and objects in the park. It uses distributed environmental perception components to obtain information on temperature, humidity, smoke concentration, and gas composition in the park. It uses multimodal identity verification components to obtain biometric information and identification information of people entering the park. This module integrates various types of collected information to form a basic park security data set. The multi-type video acquisition components include wide-angle panoramic cameras, low-light night vision cameras, and close-up behavior cameras. The wide-angle panoramic cameras cover the park's public activity areas, the low-light night vision cameras are deployed in areas of the park without auxiliary lighting at night, and the close-up behavior cameras are deployed at the park's entrances and exits and in the equipment room. The multi-type video acquisition component processes the collected continuous image frames through the spatiotemporal feature fusion algorithm, extracts the spatial position characteristics and temporal motion characteristics of people and objects, establishes dynamic tracking links between people and objects, and records the movement paths of people and objects within the park.

[0022] The spatiotemporal feature fusion algorithm includes a spatial feature extraction layer, a temporal feature processing layer, and a feature fusion layer. The spatial feature extraction layer uses a convolutional neural network to extract the contours and size features of people and objects in image frames. The temporal feature processing layer uses a recurrent neural network to analyze the movement direction and speed features of people and objects in continuous image frames. The feature fusion layer uses an attention mechanism to assign weights to spatial features and temporal features to form dynamic feature vectors of people and objects.

[0023] Distributed environmental sensing components include temperature and humidity sensors, optical smoke sensors, and electrochemical gas sensors. Temperature and humidity sensors are deployed at different heights indoors and outdoors in the park. Optical smoke sensors are deployed in park corridors and on top of equipment rooms. Electrochemical gas sensors are deployed in the park's underground garages and around laboratories. The distributed environmental perception component integrates the data collected by each sensor through a multi-source data fusion algorithm, calculates the mean environmental parameters of different areas in the park, and establishes a regional environmental parameter matrix. When any parameter in the regional environmental parameter matrix deviates from the normal parameter matrix of the same period in history by more than a preset range, the area is marked as an environmental abnormality area.

[0024] The multi-modal identity verification assembly integrates a visible light face recognition device, a capacitive fingerprint collection device, and an RFID reading device. The visible light face recognition device collects personnel facial feature point information, the capacitive fingerprint collection device collects personnel fingerprint texture information, and the RFID reading device collects identity code information in a personnel carried identification card. The multi-modal identity verification assembly matches the collected facial feature point information, fingerprint texture information, and identity code information through a feature fusion authentication algorithm. When the matching degrees of the three types of information all reach a preset standard, personnel identity authentication is completed. Meanwhile, according to permission data associated with the identity code information, the activity area range of the personnel in the park is delimited.

[0025] Further, the data collection module is used to acquire park security data. Different dimensions of security information are collected by a multi-type video collection assembly, a distributed environment sensing assembly, and a multi-modal identity verification assembly, and then various collected information is integrated to form a park security basic data set. The data set is transmitted to a security central platform in real time. The security central platform classifies and stores the data and pre-processes the data to remove redundant information and abnormal noise, thereby ensuring data quality and providing a reliable basis for subsequent feature extraction and risk judgment.

[0026] The multi-type video collection assembly includes a wide-angle panoramic camera, a low-illumination night vision camera, and a close-up behavior camera. The wide-angle panoramic camera has a 180-degree to 360-degree field of view and is deployed in public activity areas such as park squares and playgrounds, which can achieve large-range scene coverage and record the dynamic distribution of multiple targets. The low-illumination night vision camera uses a starlight-level photosensitive chip and can still clearly image when the ambient light intensity is less than 0.01 lux. It is specially deployed in areas without auxiliary lighting at night, such as park fence perimeters and green belts, to solve the problem of night monitoring blind spots. The close-up behavior camera is equipped with a 4K high-definition lens and an AI behavior analysis chip and is mainly deployed at entrances and exits such as park gates and side gates, as well as equipment rooms such as power distribution rooms and server rooms, to accurately capture the subtle actions and facial features of targets.

[0027] When the multi-type video acquisition assembly is running, the continuous image frames collected are processed by a space-time feature fusion algorithm. In the algorithm running process, first, the space feature extraction layer plays a role. A 16-layer convolutional neural network is adopted. Through the alternate operation of 3x3 and 5x5 convolution kernels, the space features such as the contour edges and size ratios of the persons and objects in the image frames are gradually extracted, and a space feature matrix is generated. Then, the time feature processing layer is started. A bidirectional recurrent neural network is adopted to perform time series analysis on the space feature matrices of the continuous 20 frames of images, calculate the pixel displacement amount of the targets between adjacent frames, and thus infer the time features such as the moving direction and moving speed of the persons and objects, form a time feature vector, and finally, the feature fusion layer allocates different weights to the space features and the time features according to the activity level of the target movement through the attention mechanism. The higher the activity of the moving target, the higher the weight proportion of the time features, which can be up to 0.6, and the weight proportion of the space features is correspondingly reduced. Conversely, the weight proportion of the space features is increased. The dynamic feature vector of the persons and objects is formed through weighted calculation.

[0028] Based on the dynamic feature vector, the multi-type video acquisition assembly can establish a dynamic tracking link of the persons and objects, assign a unique identification code to each target, and record the moving path of the target in the park in real time. When the target enters a preset forbidden area or the moving track is abnormal, a track abnormality signal is generated and transmitted to the security hub platform. The priority of the signal is set to level two.

[0029] The distributed environment perception assembly includes a temperature and humidity sensor, an optical smoke sensor, and an electrochemical gas sensor. The temperature and humidity sensor adopts a digital measurement method, and the measurement accuracy can reach ±0.5℃ and ±3%RH. It is respectively arranged at different height positions such as the corridor ceiling of the 1st to 5th floors of the office building in the park, the outdoor green belt, and the underground garage upright column, to realize all-around monitoring of the indoor and outdoor environment temperature and humidity. The optical smoke sensor is based on the light scattering principle, and the detection sensitivity of the smoke particles is 0.01dB / m. It is mainly installed at the top of the corridor of each building in the park and the top of the equipment room such as the power distribution room and the fire control room, to monitor the smoke concentration change in real time. The electrochemical gas sensor is for toxic and harmful gases such as carbon monoxide and hydrogen sulfide, and the measurement range is 0-1000ppm, and the detection resolution is 1ppm. It is specially arranged near the ventilation opening of the underground garage and the outer wall of the chemical laboratory in the park, to ensure that the gas leakage risk is found in time.

[0030] The distributed environment perception component runs, through a multi-source data fusion algorithm, the data collected by each sensor is integrated, first, the collected data of three or more than three sensors of the same type in the same area is calculated by mean value, the abnormal data beyond the normal fluctuation range is eliminated, the mean value of the environment parameters of the area is obtained, then according to the geographical zoning of the park, the mean value of the environment parameters of each area is summarized, the regional environment parameter matrix containing temperature, humidity, smoke concentration and gas composition is established, at the same time, the security central platform will call the environment parameter data of the same period in the past 12 months, construct the historical same period normal parameter matrix, compare the regional environment parameter matrix with the historical same period normal parameter matrix one by one.

[0031] In order to accurately determine the degree of environmental abnormality, an environmental abnormality determination coefficient calculation formula is introduced, the formula is: In the formula, EAC is the environmental abnormality determination coefficient, which comprehensively reflects the degree of environmental abnormality, the value range is 0~1, EAC≥0.3 is determined as abnormal, the larger the value, the more serious the abnormality; W t is the temperature sub-item weight; W h is the humidity sub-item weight; W s is the smoke concentration sub-item weight; W9 is the gas concentration sub-item weight; k1 is the temperature parameter weight coefficient, which is set according to the area function, 0.25 for office area, 0.35 for equipment room, and 0.2 for outdoor area; T 实时 is the real-time temperature mean value, which is calculated by the mean value of the data collected by three or more than three temperature and humidity sensors in the same area after eliminating abnormal values; T 历史 is the historical same period normal temperature mean value, which is the temperature mean value of the same period in the past 12 months of the area, the reference value after excluding the influence of extreme weather; k2 is the humidity parameter weight coefficient, which is uniformly taken as 0.2, because the influence of humidity on the park security is relatively stable, and there is no need to be differentiated according to the area; H 实时 is the real-time humidity mean value, which is calculated by the mean value of the data collected by three or more than three temperature and humidity sensors in the same area after eliminating abnormal values; H 历史 is the historical same period normal humidity mean value, which is the humidity mean value of the same period in the past 12 months of the area, the reference value after excluding the influence of extreme weather; k3 is the smoke concentration parameter weight coefficient, which is set according to the risk level of the area: 0.4 for equipment room, 0.2 for corridor and outdoor area; S 实时 is the real-time smoke concentration mean value, which is calculated by the mean value of the data collected by three or more than three optical smoke sensors in the same area after eliminating abnormal values; S 历史 ​​​​is the historical same period normal smoke concentration average, which is the average of the smoke concentration in the same period of the past 12 months in the region, and the baseline value close to 0 in the normal state; k4 is a gas concentration parameter weight coefficient, which is set according to the danger degree of the region, 0.5 for the laboratory, 0.3 for the underground garage, and 0.1 for other regions; 实时 is the real-time average toxic gas concentration, which is the average of the data collected by three or more electrochemical gas sensors in the same region after removing outliers; 历史 is the historical same period normal gas concentration average, which is the average of the toxic gas concentration in the same period of the past 12 months in the region, and the baseline value close to 0 in the normal state; 安全阈值 is the upper limit of the safe concentration of toxic gas, which is set according to the national "Indoor Air Quality Standard"; Compare EAC with the preset threshold value 0.3, EAC < 0.3 is normal, 0.3 ≤ EAC < 0.5 is mild abnormal, 0.5 ≤ EAC < 0.8 is moderate abnormal, and EAC ≥ 0.8 is severe abnormal, different levels correspond to different early warning response strategies.

[0032] The multi-modal identity verification assembly integrates visible light face recognition equipment, capacitive fingerprint collection equipment and RFID reading equipment. The visible light face recognition equipment uses a binocular camera and can collect 68 feature point information of the face with an accuracy of 99.8%. It is installed on the verification terminal at each entrance of the park to quickly collect the facial features of the personnel. The capacitive fingerprint collection equipment records the fingerprint texture information through the change of the capacitance value, with a collection resolution of 500 dpi and supporting dry and wet finger recognition. It is used together with the face recognition equipment to further improve the identity verification accuracy. The RFID reading equipment uses ultra-high frequency reading and writing technology with a reading distance of 0-50 cm, which can quickly collect the identity code information stored in the identification card carried by the personnel. It is also compatible with the existing access control card system in the park to ensure the compatibility and convenience of identity information collection.

[0033] When the multi-modal identity verification assembly is running, the collected facial feature point information, fingerprint texture information and identity code information are matched by a feature fusion authentication algorithm. First, the facial feature point information and the fingerprint texture information are converted into digital feature codes, and then compared with the pre-stored feature codes corresponding to the identity code information obtained by the RFID reading device. When the matching degrees of the three types of information all reach the preset standard, i.e., the facial feature matching degree is greater than or equal to 95%, the fingerprint texture matching degree is greater than or equal to 98%, and the identity codes are completely consistent, the personnel identity authentication is completed. At the same time, the multi-modal identity verification assembly will determine the activity area range of the personnel in the park according to the permission data associated with the identity code information. Ordinary employees can only enter the office area and the public activity area, and equipment maintenance personnel can additionally enter the equipment room area. When the personnel attempt to enter an area beyond the permission range, the assembly will generate a permission boundary crossing signal and transmit it to the security hub platform. The priority of the signal is set to level one. After receiving the signal, the security hub platform starts the early warning response mechanism according to the priority. The instant response is triggered by the level one signal, and the regular response is triggered by the level two signal, so that high-risk events are disposed of in priority.

[0034] After receiving the trajectory anomaly signal, the environmental anomaly signal or the permission boundary crossing signal, the security hub platform will immediately start the early warning response mechanism, issue a warning prompt through the park broadcast system, send a warning message to the mobile terminal of the security personnel, and call the real-time video picture of the abnormal area to assist the security personnel in quickly disposing, so as to ensure that the park security risks are timely controlled. The above algorithms and formulas are obtained by collecting a large amount of park security data for software simulation and selecting a model close to the true value. The parameter threshold in the algorithm and the weight coefficient in the formula are set by the personnel in the art according to the actual security needs of the park, the functional partition of the building and the environmental characteristics.

[0035] In use, the data acquisition module collects dynamic image information, environmental parameter information and personnel identity information of personnel and objects in the park through the multi-type video acquisition assembly, the distributed environment sensing assembly and the multi-modal identity verification assembly, integrates to form a park security basic data set, processes image data through the spatio-temporal feature fusion algorithm by the multi-type video acquisition assembly, establishes a target dynamic tracking link, identifies trajectory anomalies, quantifies the environmental anomaly degree through the multi-source data fusion algorithm and the environmental anomaly determination coefficient formula by the distributed environment sensing assembly, marks the environmental anomaly area, matches the identity information through the feature fusion authentication algorithm by the multi-modal identity verification assembly, controls the personnel activity permission, and transmits the abnormal signals generated by each component to the security hub platform. The platform starts a hierarchical early warning response, realizes intelligent and accurate early warning and control of park security.

[0036] Embodiment two, please refer to Figure 1 and Figure 2As shown: the intelligent park security early warning system based on artificial intelligence, the data transmission module is used for transmitting data between modules, through the combination of double-link wired transmission components and edge-enhanced wireless transmission components, the stable transmission of security data between modules is realized.

[0037] Further, the data transmission module realizes the stable transmission of security data between modules through the combination of double-link wired transmission components and edge-enhanced wireless transmission components. The double-link wired transmission component adopts an industrial-grade gigabit Ethernet switch, supports double-link redundancy design, and two links are deployed in different weak electric shafts in the park. The transmission bandwidth of each link can reach 1000Mbps, and the transmission delay is controlled within 5ms. When one of the links fails, the component will automatically switch to the other link, and the switching time is less than 100ms, ensuring the continuity of wired transmission. The double-link wired transmission component is mainly used to transmit dynamic feature vector data generated by multiple types of video acquisition components. The data uses H.265 encoding format, which can reduce bandwidth occupation by 50% while ensuring image quality. Through double-link wired transmission, the probability of video freezing caused by single-link congestion can be significantly reduced. When the single-link bandwidth utilization rate exceeds 80%, the system automatically splits part of the data to the standby link, further improving transmission stability.

[0038] The edge-enhanced wireless transmission component adopts 5G+LoRa dual-mode transmission technology. The 5G transmission part supports Sub-6GHz frequency band, with downlink rate up to 300Mbps and uplink rate up to 100Mbps. It is mainly used to transmit small-capacity and high-priority data such as emergency warning signals and device status feedback. When the data transmission module is running, it will perform real-time integrity check on the transmitted data. CRC32 check algorithm is used to generate a check code for each frame of data. The receiving end judges whether the data is complete by comparing the check code.

[0039] If the check is passed, the data will normally enter the next module. If the check fails, the data transmission module will generate a data retransmission signal to request retransmission from the sending end. The number of retransmissions does not exceed 3 times. If 3 times of retransmission still fails, a transmission fault signal is generated and sent to the security hub platform. After receiving the signal, the platform will start the standby transmission channel and send a maintenance prompt to the maintenance personnel.

[0040] Example three, please refer to Figure 1 and Figure 2 As shown: the intelligent park security early warning system based on artificial intelligence, the data processing module is used for processing security data. The intelligent data cleaning component is used to filter and correct abnormal items in the data. The classified storage component is used to save the processed data. The retrieval component is used to search the stored data to form a security database.

[0041] Further, the intelligent data cleaning component adopts an abnormal data identification model based on machine learning, which automatically identifies abnormal items in the data by training the historical security data of the park in the past 24 months. First, the received raw data is preliminarily screened to eliminate data that obviously does not conform to physical laws, such as records with a temperature lower than -40℃ or higher than 80℃. Then, trend analysis is performed on the remaining data, and the fluctuation amplitude of the data is calculated. When the average deviation of a certain data point from the adjacent 10 data points exceeds the preset 20%, it is determined to be an abnormal item. Finally, the abnormal items are corrected. For environmental parameter data, the mean value of normal data in adjacent time periods is used for filling. For video image data, the previous clear image is used for replacement, ensuring the accuracy of the cleaned data.

[0042] The classified storage component adopts a distributed storage architecture, which divides data into three categories: video image data, environmental parameter data, and personnel identity data. Video image data is stored hierarchically by collection time and area, retaining high-definition video data for nearly 30 days, standard-definition compressed data for 30 days to 1 year, and key event video clips for more than 1 year. Environmental parameter data is summarized and stored by hour, retaining historical data for the past 5 years. Personnel identity data is stored in association with permission information using an encryption storage method, with the key managed separately by the security hub platform to prevent data leakage.

[0043] The retrieval component adopts a hybrid retrieval technology based on keywords and feature values, supporting multi-dimensional retrieval requirements. Users can quickly retrieve corresponding security data by inputting keywords such as time range, area name, and event type. For video image data, target features such as personnel clothing color and object shape are inputted to automatically match video segments that meet the features. The retrieval response time is controlled within 3 seconds, and the retrieval result export function is supported. When the retrieval component does not query corresponding data, a no result prompt signal is generated and fed back to the user.

[0044] Embodiment four, please refer to Figure 1 and Figure 2 The intelligent park security early warning system based on artificial intelligence is shown: the intelligent analysis module is used to analyze the park security basic data set. The intelligent behavior recognition component analyzes the motion trajectory and behavior pattern of personnel and objects. The security event prediction component combines historical security data and real-time data to judge potential security events. The dynamic risk assessment component constructs an evaluation index system to determine the security risk level of the park, generating a park security analysis result. The intelligent behavior recognition component includes a trajectory analysis unit, a behavior classification unit, and an abnormality determination unit. The trajectory analysis unit smoothes the motion trajectories of personnel and objects through a Kalman filtering algorithm, removing noise interference in the trajectories. The behavior classification unit divides personnel behavior into normal activity, loitering and staying, and crossing obstacles categories using a deep learning model. The abnormality determination unit compares the current behavior category with the historical normal behavior category library for the area. When the current behavior category does not appear in the historical normal behavior category library, the behavior is marked as abnormal behavior, and the time and location of the abnormal behavior are recorded.

[0045] The security event prediction component includes a data preprocessing unit, a model training unit, and a prediction execution unit. The data preprocessing unit normalizes historical security event data, removing redundant information in the data. The model training unit trains the preprocessed historical data using a spatiotemporal attention long short-term memory network to establish a security event prediction model. The prediction execution unit inputs real-time park state data into the security event prediction model, outputting the types and corresponding probability values of possible security events in a future preset time period.

[0046] Further, the intelligent analysis module works collaboratively with the intelligent behavior recognition component, the security event prediction component, and the dynamic risk assessment component to complete security data analysis. The intelligent behavior recognition component includes a trajectory analysis unit, a behavior classification unit, and an abnormality determination unit. The trajectory analysis unit smoothes the motion trajectories of personnel and objects through a Kalman filtering algorithm. First, a motion model is established based on the trajectory position, speed, and acceleration at the previous moment to predict the trajectory coordinates at the current moment. Then, the trajectory noise interference caused by lens shaking and occlusion is removed by combining the position data collected in real time by the multi-type video acquisition component, so that the trajectory error is controlled within 0.5 meters, ensuring the smoothness and accuracy of the trajectory data.

[0047] The behavior classification unit uses a deep learning model to recognize the behavior of the smoothed trajectory data. This model is based on the ResNet50 network architecture and can divide personnel behavior into normal activity, loitering and staying, and crossing obstacles categories by training 100,000 personnel behavior samples in different scenarios in the park. Normal activity includes continuous movement behaviors such as walking and running. Loitering and staying refers to behavior where personnel stay in an area for more than 10 minutes and move less than 5 meters. Crossing obstacles refers to behavior where personnel cross pre-set prohibited crossing facilities such as fences and guardrails. The recognition accuracy of the model for each type of behavior can reach more than 98%.

[0048] The anomaly determination unit pre-constructs a historical normal behavior category library for each area, which is dynamically updated according to the functional characteristics of different areas in the park. For example, the historical normal behavior category library for the office area includes normal activities and short-time lingering stays, while the historical normal behavior category library for the equipment room area only includes normal activities. The anomaly determination unit compares the current behavior category with the historical normal behavior category library of the corresponding area. When the current behavior category does not appear in the historical normal behavior category library, such as lingering stay behavior in the equipment room area, the behavior is marked as abnormal behavior, and the time and location of the abnormal behavior are recorded. An abnormal behavior signal is generated and transmitted to the security hub platform.

[0049] The security event prediction component includes a data preprocessing unit, a model training unit, and a prediction execution unit. The data preprocessing unit processes historical security event data from the past 5 years. First, it eliminates duplicate records and invalid data. Then, it uses the min-max normalization method to map event occurrence time, area, and environmental parameter feature data to the 0-1 interval, removing the interference caused by data magnitude differences.

[0050] The model training unit uses a spatio-temporal attention long short-term memory network to train the preprocessed historical data. This network adds a spatio-temporal attention mechanism to the traditional LSTM network, focusing on significant spatio-temporal features that affect security events, such as personnel flow characteristics in entrance and exit areas during holidays. During training, the historical data is divided into a training set and a validation set in a 7:3 ratio. The network parameters are adjusted through iterative training. When the prediction accuracy of the model on the validation set exceeds 92%, training is stopped, and the security event prediction model is saved.

[0051] The prediction execution unit inputs real-time park state data into the security event prediction model. These real-time data include current personnel distribution, environmental parameters, and equipment operating status. The model outputs the types of security events that may occur within the next 24 hours and the corresponding probability values. When the probability value of a certain type of security event exceeds 60%, the prediction execution unit generates an event warning signal and sends it to the security hub platform. The platform will allocate security resources in advance based on the warning signal.

[0052] The dynamic risk assessment component constructs an evaluation index system containing four dimensions: personnel density, equipment status, environmental parameters, and historical events. Each dimension has specific evaluation factors and weights. The personnel density dimension has a weight of 0.3, including regional personnel quantity and personnel aggregation degree. The equipment status dimension has a weight of 0.25, including equipment operating temperature and failure frequency. The environmental parameter dimension has a weight of 0.25, including temperature and humidity deviation value, smoke concentration, and gas composition. The historical event dimension has a weight of 0.2, including the number of events in the past three months and the severity of the events in the region.

[0053] The overall security risk value of the park is calculated through weighted summation. The risk value range is 0 to 100, with 0 to 30 being low risk, 31 to 60 being medium risk, and 61 to 100 being high risk. The dynamic risk assessment component updates the risk value every hour.

[0054] Example 5, please refer to Figure 1 and Figure 2 As shown: The AI-based smart campus security early warning system has an early warning decision module that formulates disposal plans based on the results of campus security analysis. The dynamic threshold judgment component combines preset standards with real-time analysis data to determine risk trigger conditions. The multi-level early warning release component divides early warning categories according to risk level and pushes early warning information. The emergency plan generation component combines early warning levels with historical disposal cases to formulate response plans and output the results of the campus security early warning decision. The dynamic threshold determination component includes a historical data statistics unit, a threshold calculation unit and a threshold update unit. The historical data statistics unit collects statistics on the park security data within the past preset period and calculates the fluctuation range of various security parameters. The threshold calculation unit uses a normal distribution algorithm to determine the initial risk threshold of various parameters based on the fluctuation range. The threshold update unit recalculates the risk threshold based on the latest statistical security data after each preset period to achieve dynamic adjustment of the risk threshold and ensure that the risk threshold matches the actual security situation of the park.

[0055] The multi-level warning release component includes a warning level division unit, an information generation unit and a push unit. The warning level division unit divides the warning levels into four categories: general warning, heavy warning, serious warning and particularly serious warning. The information generation unit generates warning text information including the warning area, warning type and risk level according to the warning level. The push unit pushes the warning text information simultaneously through three channels: the park broadcasting system, mobile terminal APP and indoor display screen. The particularly serious warning information additionally triggers the park's sound and light alarm device.

[0056] Furthermore, the dynamic threshold determination component includes a historical data statistics unit, a threshold calculation unit and a threshold update unit. The historical data statistics unit collects statistics on the park security data in the past 6 months. Selecting 6 months as the preset period can not only cover the changes in the park's security conditions in different seasons, but also ensure the timeliness of the data.

[0057] The threshold calculation unit uses the normal distribution algorithm to determine the initial risk thresholds of various parameters based on the fluctuation range. The mean and standard deviation of the parameter data are first calculated, and then the mean plus 2 times the standard deviation is used as the upper limit of the initial risk threshold, and the mean minus 2 times the standard deviation is used as the lower limit of the initial risk threshold.

[0058] The threshold updating unit recalculates the risk threshold every 6 months in combination with the latest statistical security data. If the change rate of the fluctuation range of a certain parameter exceeds 30%, such as the expansion of the temperature fluctuation range caused by high temperature in summer, the standard deviation coefficient in the normal distribution algorithm is adjusted from 2 to 2.2-2.5. For every 10% increase in the expansion rate of the fluctuation range, the standard deviation coefficient increases by 0.1. This ensures that the updated risk threshold can cover more than 95% of normal data and match the actual security situation of the park.

[0059] The multi-level early warning release component includes an early warning level division unit, an information generation unit, and a pushing unit. The early warning level division unit divides the early warning level into four categories: general early warning, relatively severe early warning, severe early warning, and extremely severe early warning. The division is based on the risk impact range, the degree of emergency disposal, and the potential loss. The general early warning only affects a single area and poses no risk to personnel and property. The extremely severe early warning affects the entire park and poses a major safety hazard.

[0060] The information generation unit generates early warning text information containing the early warning area, early warning type, and risk level based on the early warning level. The pushing unit synchronously pushes the early warning text information through three channels: the park broadcast system, mobile terminal APP, and indoor display screen. The pushing delay is controlled within 10 seconds to ensure that relevant personnel quickly obtain early warning information. In addition, the extremely severe early warning information triggers the sound and light alarm device in the park. The sound and light alarm device is installed at major intersections and building entrances in the park. The alarm sound intensity reaches 85 decibels, and the light flickering frequency is 1 time per second. The triggering continues until the early warning is lifted to maximize the attention of personnel.

[0061] Embodiment six, please refer to Figure 1 and Figure 2 The intelligent park security early warning system based on artificial intelligence. The execution feedback module is used to implement the early warning decision results and feedback the execution situation. The emergency response component takes emergency measures according to the response plan. The dynamic resource scheduling component allocates park security resources. The disposal effect evaluation component evaluates the event disposal situation to form a closed-loop disposal process.

[0062] Further, the execution feedback module works collaboratively with the emergency response component, dynamic resource scheduling component, and disposal effect evaluation component. The emergency response component takes emergency measures according to the response plan. The measure type is determined according to the early warning level and event type. General early warning corresponds to on-site verification and simple disposal, such as general early warning device failure events. The component will instruct maintenance personnel to carry tools to the site for repair. Extremely severe early warning corresponds to emergency lockdown and professional disposal.

[0063] The dynamic resource scheduling component allocates campus security resources, including security personnel, firefighting equipment, monitoring equipment, and emergency vehicles. It monitors the distribution and usage of these resources in real time. Upon receiving a resource allocation instruction, it uses a greedy algorithm to calculate the optimal resource allocation path to ensure that resources reach the scene as quickly as possible. For example, when deploying security personnel, priority is given to patrol teams closest to the incident area and currently idle. When deploying firefighting equipment, priority is given to fire hydrants and extinguishers within 300 meters of the incident area. Once resource allocation is complete, the component sends a confirmation signal to the central security platform confirming the availability of these resources.

[0064] The disposal effect evaluation component evaluates the incident disposal situation. The evaluation indicators include the timeliness of disposal, the effectiveness of measures and the degree of risk elimination. The timeliness of disposal is measured by the time from the start of execution to the initial control of the risk. General warnings require risk control within 30 minutes, and particularly serious warnings require risk control within 10 minutes. The effectiveness of measures is determined by comparing the changes in risk parameters before and after disposal, such as whether the smoke concentration has dropped to a safe range after the fire hazard is disposed of. The degree of risk elimination determines whether the incident has been completely resolved and there are no secondary risks.

[0065] The disposal effect evaluation component evaluates the incident disposal situation. The evaluation indicators include the timeliness of disposal, effectiveness of measures and degree of risk elimination, with weights of 0.3, 0.4 and 0.3 respectively.

[0066] The timeliness of disposal is measured by the time from the start of execution to the initial control of the risk. The general warning requires that the risk be controlled within 30 minutes. If the standard is met, 100 points will be awarded, and 10 points will be deducted for every 5 minutes of delay. The particularly serious warning requires that the risk be controlled within 10 minutes. If the standard is met, 100 points will be awarded, and 10 points will be deducted for every 1 minute of delay. The effectiveness of the measures is measured by comparing the changes in risk parameters before and after the disposal. If the risk parameters are reduced to a safe range, 100 points will be awarded, and 10 points will be deducted for every 10% of residual risk. The degree of risk elimination is judged by whether the incident is completely resolved and there is no secondary risk. A complete resolution will be awarded 100 points, and 30 to 80 points will be deducted for any residual risk.

[0067] The weighted scoring method is used to calculate the comprehensive score of the disposal effect, with the score range from 0 to 100 points. 80 points and above are considered excellent disposal, 60 to 80 points are considered qualified disposal, and below 60 points are considered unqualified disposal.

[0068] Example 7, please refer to Figure 1 and Figure 2 As shown: The smart park security early warning system based on artificial intelligence and the smart park security early warning method based on artificial intelligence include the following steps: S1, through the multi-type video acquisition component, the distributed environment perception component and the multi-modal identity verification component, respectively collect park dynamic image information, environment parameter information and personnel identity information, and integrate to form a park security basic data set; S2, through the double-link wired transmission component and the edge enhancement wireless transmission component, the park security basic data set is transmitted to the data processing module, and the data is encrypted during transmission; S3, through the intelligent data cleaning component, the abnormal data in the park security basic data set is screened and corrected, the corrected data is stored in the image database, the environment database and the personnel database through the classified storage component, and the index for each database is established through the retrieval component; S4, calling the intelligent behavior recognition component to analyze the dynamic image data in the image database, identifying abnormal behavior, calling the security event prediction component to combine historical security data and real-time environment data to predict potential security events, calling the dynamic risk assessment component to determine the current risk degree of the park, and generating the park security analysis result; S5, through the dynamic threshold determination component, the risk trigger condition is determined according to the park security analysis result, the warning level is divided and the warning information is pushed through the multi-level warning release component, the corresponding emergency disposal scheme is formulated through the emergency plan generation component, and the warning decision result is output; S6, according to the warning decision result, through the emergency response component, the emergency measures are taken, through the dynamic resource scheduling component, the security resources are allocated, through the disposal effect evaluation component, the disposal process is evaluated, and the evaluation result is fed back to the security central platform.

[0069] The above shows and describes the basic principles and main features of the present application and the advantages of the present application. It should be understood by those skilled in the art that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. The intelligent park security early warning system based on artificial intelligence, characterized in that, The security central platform is communicatively connected with a data acquisition module, an intelligent analysis module, an early warning decision module, a data transmission module, a data processing module, and an execution feedback module; The data acquisition module is configured to acquire park security data, acquire dynamic image information of personnel and objects in the park through multiple types of video acquisition components, acquire temperature and humidity, smoke concentration, and gas composition information in the park through distributed environment sensing components, acquire biological feature information and identification information of personnel entering the park through multi-modal identity verification components, and integrate various types of collected information to form a park security basic data set; The intelligent analysis module is configured to analyze the park security basic data set, analyze the motion trajectory and behavior pattern of personnel and objects through an intelligent behavior recognition component, determine potential security events in combination with historical security data and real-time data through a security event prediction component, determine the degree of park security risk through an evaluation index system constructed by a dynamic risk assessment component, and generate a park security analysis result; The early warning decision module is configured to develop a disposal scheme based on the park security analysis result, determine a risk triggering condition in combination with preset standards and real-time analysis data through a dynamic threshold determination component, divide early warning categories and push early warning information according to the degree of risk through a multi-level early warning publishing component, develop a response scheme in combination with the early warning level and historical disposal cases through an emergency plan generation component, and output a park security early warning decision result; The data transmission module is configured to transmit data between modules, and realizes stable transmission of security data between modules through a combination of a double-link wired transmission component and an edge-enhanced wireless transmission component; The data processing module is configured to process security data, filter and correct abnormal items in the data through an intelligent data cleaning component, save the processed data in a classified manner through a classified storage component, and search for stored data through a search component to form a security database; The execution feedback module is configured to implement the early warning decision result and feed back the execution situation, takes emergency measures according to the response scheme through an emergency response component, allocates park security resources through a dynamic resource scheduling component, and evaluates the event disposal situation through a disposal effect evaluation component to form a closed-loop disposal process. 2.The AI-based intelligent park security early warning system according to claim 1, characterized in that, The multi-type video acquisition component includes a wide-angle panoramic camera, a low-illumination night-vision camera, and a close-up behavior camera, the wide-angle panoramic camera covers the public activity area of the park, the low-illumination night-vision camera is deployed in the area without auxiliary lighting at night in the park, and the close-up behavior camera is deployed at the entrance and equipment room of the park; The multi-type video acquisition component processes the collected continuous image frames through a spatio-temporal feature fusion algorithm, extracts the spatial position features and temporal motion features of personnel and objects, establishes a dynamic tracking link of personnel and objects, and records the moving path of personnel and objects in the park. 3.The AI-based intelligent park security early warning system according to claim 2, characterized in that, The spatio-temporal feature fusion algorithm comprises a spatial feature extraction layer, a temporal feature processing layer and a feature fusion layer. The spatial feature extraction layer extracts the contour and size features of the personnel and objects in the image frames by using a convolutional neural network. The temporal feature processing layer analyzes the motion direction and speed features of the personnel and objects in the continuous image frames by using a recurrent neural network. The feature fusion layer performs weight distribution on the spatial features and the temporal features by using an attention mechanism to form a dynamic feature vector of the personnel and objects. 4.The AI-based intelligent park security early warning system according to claim 1, characterized in that, The distributed environment perception component comprises a temperature and humidity sensor, an optical smoke sensor and an electrochemical gas sensor. The temperature and humidity sensor is deployed at different height positions indoors and outdoors of the park. The optical smoke sensor is deployed on the top of the corridors and equipment rooms of the park. The electrochemical gas sensor is deployed in the underground garage and around the laboratories of the park. The distributed environment perception component integrates the data collected by each sensor through a multi-source data fusion algorithm, calculates the average of the environmental parameters of different areas of the park, and establishes a regional environmental parameter matrix. When the deviation of any parameter in the regional environmental parameter matrix from the historical same-period normal parameter matrix exceeds the preset range, the area is marked as an environmental abnormal area. 5.The AI-based security early warning system for smart park according to claim 1, wherein, The multi-modal identity verification component integrates a visible light face recognition device, a capacitive fingerprint collection device and an RFID reading device. The visible light face recognition device collects facial feature point information of personnel. The capacitive fingerprint collection device collects fingerprint texture information of personnel. The RFID reading device collects identity code information in the identification card carried by personnel. The multi-modal identity verification component matches the collected facial feature point information, fingerprint texture information and identity code information through a feature fusion authentication algorithm. When the matching degrees of the three types of information all reach the preset standard, the personnel identity authentication is completed. Meanwhile, according to the permission data associated with the identity code information, the activity area range of the personnel in the park is delimited. 6.The AI-based security early warning system for smart park according to claim 1, wherein, The intelligent behavior recognition component comprises a trajectory analysis unit, a behavior classification unit and an abnormality determination unit. The trajectory analysis unit smoothes the motion trajectories of personnel and objects by using a Kalman filtering algorithm to remove noise interference in the trajectories. The behavior classification unit divides the personnel behaviors into normal activities, wandering and staying and barrier crossing categories by using a deep learning model. The abnormality determination unit compares the current behavior category with the historical normal behavior category library of the area. When the current behavior category does not appear in the historical normal behavior category library, the behavior is marked as an abnormal behavior, and the time and position of the occurrence of the abnormal behavior are recorded. 7.The intelligent park security early warning system based on artificial intelligence according to claim 1, characterized in that, The security event prediction component comprises a data preprocessing unit, a model training unit and a prediction execution unit. The data preprocessing unit performs normalization processing on the historical security event data to remove redundant information in the data. The model training unit trains the preprocessed historical data by using a spatio-temporal attention long short-term memory network to establish a security event prediction model. The prediction execution unit inputs the real-time park state data into the security event prediction model to output the types and corresponding probability values of the possible security events in a future preset time period. 8.The intelligent AI-based security early warning system for smart park according to claim 1, characterized in that, The dynamic threshold determination component includes a historical data statistics unit, a threshold calculation unit and a threshold updating unit. The historical data statistics unit performs statistics on park security data in a past preset period, calculates the fluctuation range of each type of security parameter, the threshold calculation unit determines the initial risk threshold of each type of parameter according to the fluctuation range using a normal distribution algorithm, and the threshold updating unit recalculates the risk threshold every preset period combined with the latest statistical security data, realizes the dynamic adjustment of the risk threshold, and ensures that the risk threshold matches the actual security situation of the park. 9.The intelligent AI-based security early warning system for smart park according to claim 1, characterized in that, The multi-level early warning release component includes a warning level division unit, an information generation unit and a pushing unit. The warning level division unit divides the warning level into four categories: general warning, heavier warning, serious warning and extremely serious warning. The information generation unit generates warning text information containing warning area, warning type and risk degree according to the warning level. The pushing unit synchronously pushes the warning text information through three channels: park broadcast system, mobile terminal APP and indoor display screen. The extremely serious warning information additionally triggers the park sound and light alarm device.

10. The intelligent park security early warning method based on artificial intelligence, applicable to the intelligent park security early warning system based on artificial intelligence of any one of claims 1-9, characterized in that, The method comprises the following steps: S1, through the multi-type video acquisition component, the distributed environment perception component and the multi-modal identity verification component, respectively acquire park dynamic image information, environment parameter information and personnel identity information, and integrate to form a park security basic data set; S2, through the double-link wired transmission component and the edge-enhanced wireless transmission component, transmit the park security basic data set to the data processing module, and encrypt the data during transmission; S3, through the intelligent data cleaning component, screen and correct abnormal data in the park security basic data set, store the corrected data into the image database, environment database and personnel database through the classified storage component, and establish indexes for each database through the retrieval component; S4, call the intelligent behavior recognition component to analyze the dynamic image data in the image database and identify abnormal behaviors, call the security event prediction component to predict potential security events combined with historical security data and real-time environment data, call the dynamic risk assessment component to determine the current risk degree of the park, and generate a park security analysis result; S5, through the dynamic threshold determination component, determine the risk trigger condition according to the park security analysis result, through the multi-level early warning release component, divide the warning level and push the warning information, through the emergency plan generation component, formulate the corresponding emergency disposal scheme, and output the warning decision result; S6, according to the warning decision result, take emergency measures through the emergency response component, allocate security resources through the dynamic resource scheduling component, and evaluate the disposal process through the disposal effect evaluation component, and feed back the evaluation result to the security hub platform.

Images