Smart city key area security and protection and large-screen AI linkage control method

Abstract

The invention relates to a smart city key area security and protection and large-screen AI linkage control method, and relates to the technical field of smart city security and protection monitoring, and the method comprises the steps: carrying out the frequency conversion collection of video stream data based on target density, and normalizing environment sensing data; visual and numerical features are extracted, and comprehensive confidence is calculated through a weighted fusion algorithm; determining an event level according to the confidence coefficient, and constructing a security equipment action vector and a large-screen display control vector; through a heterogeneous communication protocol, double-track distribution of instructions is carried out, a large-screen picture refreshing time sequence is delayed based on mechanical action prediction time consumption of security equipment, and synchronous response of equipment action and picture content is realized; and iteratively updating the fusion weight by using a gradient descent algorithm. Through multi-source data fusion and time sequence cooperative control, the false alarm rate is effectively reduced, and the problem of picture asynchronization in the linkage process is solved.

Description

Technical Field

[0006] , , , ,

[0005]

[0001] This application relates to the technical field of smart city security monitoring, in particular to an AI linkage control method for security and large screens in key areas of smart cities. Background Art

[0002] With the advancement of smart city construction, the security monitoring systems in key areas are becoming increasingly complex, usually integrating high-definition video monitoring, various environmental sensors, and the multimedia large-screen display system in the command center. The existing security systems mainly rely on computer vision technology to analyze video streams or collect environmental data through sensors to judge abnormal situations, aiming to achieve real-time early warning of emergencies such as fires and intrusions.

[0003] However, in practical applications, there are still some limitations in the existing security linkage control technologies. First, video analysis and environmental sensing often operate as independent subsystems, lacking an effective data fusion mechanism. Single visual detection is prone to false alarms due to changes in light, background interference, or dynamic targets in non-jurisdictional areas, while single environmental sensors are also prone to numerical drift due to local airflows or equipment aging. This monitoring mode lacking multi-dimensional verification is difficult to balance the sensitivity and accuracy of alarms. At the same time, most monitoring devices record all day long at a fixed sampling frequency and do not have the ability to dynamically adjust the acquisition strategy according to the density of scene targets, resulting in a waste of a large amount of network bandwidth and storage resources under normal conditions with no targets or few targets.

[0004] More critically, when an abnormal event occurs and security devices and large-screen displays need to be linked, the existing control logic is usually executed concurrently. When the command system issues an instruction, the electronic display terminal can switch the signal source in milliseconds, while the physical security devices need significant mechanical movement time to adjust the angle and focus. This mismatch in the timing between physical actions and electronic displays results in the command large screen often showing blurred images or invalid intermediate scenes during the process of camera rotation at the initial stage of switching the screen, rather than clear images对准事件现场的清晰画面. This not only affects the visual experience of monitoring personnel but also is not conducive to the command center quickly obtaining effective on-site information for decision-making.

[0005] In view of the above related technologies, an AI linkage control method for security and large screens in key areas of smart cities is provided. Summary of the Invention

[0006] The purpose of this application is to provide an AI linkage control method for security and large screens in key areas of smart cities, aiming to improve the problems in the existing smart city security systems, such as high false alarm rates due to the lack of fusion between visual and environmental data, resource waste caused by fixed acquisition strategies, and the out-of-sync between large-screen displays and physical actions due to ignoring the mechanical movement time of devices.

[0007] By adopting the above technical solution, a method for AI-linked control of security and large screens in key areas of smart cities is provided, comprising the following steps:

[0008] S1. Perform adaptive data acquisition and preprocessing: Perform frequency conversion acquisition based on target density on video stream data and perform normalization processing on environmental sensor data;

[0009] S2. Perform parallel feature extraction and multi-source data fusion: extract the visual features of the video stream data and the numerical features of the environmental sensor data respectively, and calculate the comprehensive confidence level at the current moment through a weighted fusion algorithm;

[0010] S3. Execution Event Classification and Instruction Construction: Determine the event level based on the comprehensive confidence level, and construct the security device action vector for security equipment and the large screen display control vector for large screen display terminal accordingly.

[0011] S4. Dual-track distribution and collaborative control of execution instructions: The action vector of the security device and the control vector of the large screen display are distributed separately through a heterogeneous communication protocol. Based on the estimated time consumption of the mechanical action of the security device, the screen refresh sequence of the large screen display terminal is delayed to achieve synchronous response of device action and screen content.

[0012] S5. Perform self-evolution of model parameters: Construct a loss function based on historical event logs and use the gradient descent algorithm to iteratively update the weight parameters in the weighted fusion algorithm.

[0013] Preferably, the step of performing frequency conversion acquisition based on target density on the video stream data specifically includes:

[0014] Count the total number of target objects identified in the video frames at the current moment;

[0015] The total number of the target objects is compared with a preset density threshold;

[0016] If the total number of the target objects is greater than or equal to the density threshold, the sampling frame rate of the video acquisition device is set to the frame rate setting value in the high-frequency acquisition mode;

[0017] If the total number of the target objects is less than the density threshold, the sampling frame rate of the video acquisition device is set to the frame rate setting value in the low-frequency acquisition mode.

[0018] Preferably, the extraction of visual features from the video stream data specifically includes:

[0019] Use deep learning models to identify target objects in video frames and obtain the classification confidence score for each target object;

[0020] Calculate the overlap ratio between the predicted bounding box of each target object and the preset effective monitoring area;

[0021] Calculate the product of the classification confidence score and the overlap ratio to obtain the region overlap confidence score for each target object;

[0022] The maximum confidence level of regional overlap for all target objects is determined and used as the visual anomaly probability score at the current moment.

[0023] Preferably, the step of calculating the comprehensive confidence level at the current moment using a weighted fusion algorithm specifically includes:

[0024] Obtain the environmental anomaly probability score from the environmental sensor data at the current moment;

[0025] Obtain preset visual weights, environmental weights, and relational weights;

[0026] Calculate the first product of the visual weight and the visual anomaly probability score, and the second product of the environmental weight and the environmental anomaly probability score;

[0027] Calculate the product of the visual anomaly probability score and the environmental anomaly probability score, and calculate the result of the operation between the correlation weight and the product to obtain the third product;

[0028] The sum of the first product, the second product, and the third product is calculated to obtain the overall confidence level.

[0029] Preferably, the security device motion vector includes a PTZ attitude control component, an access control switch status component, and an alarm trigger mode component; the large screen display control vector includes a screen display mode component, a signal source index component, and a UI layout template component.

[0030] Preferably, the distribution via heterogeneous communication protocols specifically includes:

[0031] Establish a message queue telemetry transmission protocol channel based on the publish-subscribe model to distribute the action vectors of the security devices to the security devices;

[0032] A Hypertext Transfer Protocol version channel based on a multiplexing mechanism is established to send the large screen display control vector to the large screen control module.

[0033] Preferably, the estimated time required for the mechanical actions of the security equipment, which delays the screen refresh timing of the large-screen display terminal, specifically includes:

[0034] Obtain the current attitude angle of the security device and the attitude angle of the target, and calculate the absolute value of the difference between the horizontal and vertical angles respectively.

[0035] Divide the absolute value of the horizontal angle difference by the horizontal rotation angular velocity to obtain the horizontal rotation time, and divide the absolute value of the vertical angle difference by the vertical rotation angular velocity to obtain the vertical rotation time.

[0036] The maximum value between the horizontal rotation time and the vertical rotation time is determined, and this maximum value is added to the preset focusing buffer time to obtain the estimated time of the mechanical action;

[0037] The moment when the large screen control module receives the large screen display control vector is taken as the reference moment. The sum of the reference moment and the product of the estimated time of the mechanical action and the synchronization coefficient is calculated to obtain the effective display time of the screen.

[0038] The control module of the large screen keeps the currently displayed content unchanged until the effective display time is reached.

[0039] Preferably, the step of iteratively updating the weight parameters in the weighted fusion algorithm using the gradient descent algorithm specifically includes:

[0040] A binary cross-entropy loss function is constructed by combining the real labels in historical event logs with the combined confidence of model predictions;

[0041] Calculate the partial derivatives of the loss function with respect to the visual weights, the environmental weights, and the association weights, respectively.

[0042] The visual weights, environmental weights, and association weights are updated using the partial derivatives and a preset learning rate.

[0043] Normalize the updated weight parameters so that the sum of the weight parameters equals one.

[0044] Preferably, determining the event level based on the comprehensive confidence level specifically includes:

[0045] Obtain the preset first-level emergency threshold and second-level early warning threshold, as well as the operating status of the equipment;

[0046] If the overall confidence level is greater than or equal to the Level 1 emergency threshold, the event level is determined to be Level 1 emergency.

[0047] If the overall confidence level is less than the first-level emergency threshold and greater than or equal to the second-level warning threshold, the event level is determined to be a second-level warning.

[0048] If the overall confidence level is less than the secondary warning threshold and the equipment is in an abnormal operating state, the event level is determined to be Level 3 maintenance.

[0049] A smart city key area security and large-screen AI linkage control system, the system comprising:

[0050] The front-end perception and execution module is used to perform frequency conversion acquisition of video stream data, acquire environmental sensing data, and receive motion vectors from security devices to drive the security devices to perform physical actions.

[0051] The edge computing and processing module is used to perform normalization processing on the collected data, extract visual and numerical features, and perform multi-source data fusion to calculate the comprehensive confidence score.

[0052] The decision-making and scheduling module is used to determine the event level based on the comprehensive confidence level, construct the action vector of the security equipment and the control vector of the large screen display, and perform dual-track distribution through a heterogeneous communication protocol;

[0053] The visualization module is used to receive the control vector of the large screen display and, based on the estimated time delay of the screen refresh sequence according to the mechanical movements of the security equipment, realize the synchronous response of the screen content and the movements of the security equipment.

[0054] The model optimization module is used to construct a loss function based on historical event logs and iteratively update the weight parameters in the edge computing and processing module using the gradient descent algorithm.

[0055] In summary, this application includes at least one of the following beneficial technical effects:

[0056] 1. This application effectively improves the accuracy of security event determination by introducing ROI region overlap confidence and multi-source heterogeneous data weighted fusion algorithm; the system not only uses visual algorithms to identify abnormal targets, but also combines environmental sensor data for nonlinear cross-validation. In particular, by calculating the correlation product between visual and environmental features, it can automatically filter out visual false alarms caused by light and shadow reflection or non-jurisdictional areas, as well as numerical fluctuation interference from single sensors, thus reducing the false alarm rate.

[0057] 2. This application establishes a timing synchronization mechanism based on the estimated time consumption of mechanical actions of security equipment, which solves the problem of asynchronous screen display and physical action in traditional linkage control; by accurately calculating the physical time required for pan-tilt rotation and focusing, the system can automatically delay the screen refresh time of the large screen display terminal, ensuring that the content displayed on the command screen is always a clear and effective image after the security equipment has completed its actions, and avoiding invalid image display during the equipment rotation process;

[0058] 3. This application realizes adaptive frequency conversion acquisition and vectorized distribution of commands based on target density, which optimizes the utilization efficiency of system resources; the system dynamically adjusts the video sampling frame rate according to the number of targets in the monitoring area, which reduces data transmission bandwidth and storage pressure while ensuring complete recording of key events; at the same time, it adopts MQTT and HTTP / 2 dual-track protocols to transmit security action vectors and large screen control vectors respectively, adapting to the communication characteristics of different terminals and improving the overall response speed of the system. Attached Figure Description

[0059] Figure 1 This is a flowchart of the overall process of the security and large-screen AI linkage control method provided in the embodiments of the present invention;

[0060] Figure 2 This is an overall architecture diagram of the smart city key area security and large-screen AI linkage control system provided in the embodiments of the present invention. Detailed Implementation

[0061] The following is in conjunction with the appendix Figure 1 -Appendix Figure 2 This application will be described in further detail below.

[0062] Reference Figure 1 The method for AI-driven linkage control of security and large screens in key areas of smart cities includes the following steps:

[0063] S1. Perform adaptive data acquisition and preprocessing. The security sensing device cluster dynamically adjusts the video acquisition frame rate according to the personnel density in the monitoring area, and encodes, compresses and normalizes the acquired video image data and environmental sensing data.

[0064] S2. Perform multimodal data fusion analysis based on edge computing. The AI ​​core processing module extracts visual features from video image data and numerical features from environmental sensing data. Through weighted calculation, it obtains the comprehensive confidence level of the event and generates an analysis report containing event type, location coordinates and confidence level parameters.

[0065] S3. Execute hierarchical linkage rule matching and instruction generation. The linkage scheduling module determines the urgency level of the event based on the comprehensive confidence level and generates corresponding security equipment control instructions and large screen display instructions based on the preset rule library.

[0066] S4. Perform collaborative control between security equipment and the large screen, drive the security sensing device cluster to perform physical actions through the MQTT protocol, and drive the large screen control module to switch the displayed content through the HTTP / 2 protocol;

[0067] S5 executes feedback recording and model parameter optimization. The data storage module records the entire process log, and the AI ​​core processing module calculates the loss function based on the log data and updates the weight parameters of the multi-source data fusion algorithm.

[0068] In step S1, adaptive data acquisition and preprocessing are performed. This process specifically includes a frequency conversion acquisition sub-step for video stream data and an acquisition and standardization sub-step for environmental sensor data.

[0069] The system performs real-time analysis of the raw video stream acquired by the video capture device. The AI ​​core processing module or the computing unit integrated in the front end of the device runs a lightweight object detection algorithm to analyze the current moment. video frames A scan is performed. This lightweight object detection algorithm is used to identify target objects of preset categories in the image, including people, vehicles, or specific objects to be detected. The system uses the algorithm to calculate the current time. The total number of target objects identified in the image is denoted as . .

[0070] The system calculates the target quantity based on statistics. With the preset density threshold The comparison results determine the current frame rate of the video capture device. To optimize bandwidth and storage resources while ensuring the capture capability of critical events, the system establishes a sampling frame rate... With target quantity The mapping relationship model.

[0071] This mapping model is represented by a piecewise function, and the specific calculation formula is as follows:

[0072] ;

[0073] In the above formula: Indicates the current sampling time; Indicates at time The sampling frame rate set by the video capture device; Indicates at time Total number of detected target objects; This indicates a preset density threshold. This indicates the frame rate setting value in high-frequency acquisition mode; This represents the frame rate setting in low-frequency acquisition mode. When the number of detected targets is greater than or equal to the threshold, the system determines the area of ​​high interest and switches to a high frame rate; otherwise, it maintains a low frame rate. After acquisition is complete, the system uses the current frame rate as a time base parameter to encode and compress the video data.

[0074] Various environmental sensors in the security sensing device cluster acquire physical environmental parameters within the monitored area. The system receives the raw output signals from these sensors and defines them as raw analog quantities.

[0075] To eliminate the influence of different physical dimensions on subsequent calculations, the system performs a normalization mapping on the original analog quantities. The system pre-stores the physical range parameters corresponding to various sensors and uses linear transformation to map data with different dimensions to a unified dimensionless standard range. .

[0076] The specific mathematical model for normalization calculation is as follows:

[0077] ;

[0078] In the above formula: Indicates the current sampling time; Indicates the category index of the sensor; This represents the standardized value after normalization. Represents the original analog data; and They represent the first The maximum and minimum effective monitoring range of the sensor.

[0079] The system encrypts and encapsulates encoded and compressed video data and normalized environmental data, preferably using the SM4 block cipher algorithm to ensure data security during transmission to the edge computing layer.

[0080] In step S2, the AI ​​core processing module performs parallel feature extraction on the received standardized data stream, including visual feature extraction, environmental feature transformation, and multi-source data fusion.

[0081] The system uses a pre-built deep learning model to analyze video frames. Feature map analysis is performed to identify anomalous target categories such as flames, smoke, fallen objects, and running objects. To quantify the degree of visual anomalies, a visual anomaly discrimination function is defined. For the detected [object / type]... For each target object, the model outputs a classification confidence score. Meanwhile, to eliminate interference from targets located in non-priority areas, the system calculates the intersection-union ratio (IUU) between the predicted target bounding box and the preset effective monitoring area, defined as the regional overlap confidence level. The system iterates through all target object collections. Calculate the probability score of visual anomalies .

[0082] The formula for calculating the probability score of visual anomalies is as follows:

[0083] ;

[0084] In the above formula: This represents the anomaly probability score for the visual dimension, with a value range of... ; The set of detected targets; For the first The classification probability of each target; For the first The overlap ratio between the predicted bounding box of a target and the pre-calibrated effective monitoring area mask of the system. If the target is completely outside the monitoring area, this value is 0, thus effectively filtering out visual false alarms from invalid areas.

[0085] The system processes the normalized values ​​from environmental sensors using a nonlinear activation function to calculate the probability score of environmental anomalies. :

[0086] ;

[0087] In the above formula: Use the Sigmoid activation function; This is the sensitivity adjustment coefficient; For the first The abnormal activation threshold for this type of sensor. This formula highlights abnormal fluctuations exceeding the threshold through a nonlinear mapping.

[0088] The system performs weighted fusion of multi-source heterogeneous data. The system reads the preset visual weights. Environmental weight and related weights Calculate the overall confidence level :

[0089] ;

[0090] in, This is a non-linear correlation term used to increase confidence through cumulative effects when both visual and environmental features are abnormal, thus achieving cross-validation. The system then encapsulates the combined confidence level, event category, and location information into an event analysis report.

[0091] In step S3, the linkage scheduling module performs hierarchical judgment and instruction construction based on the event analysis report.

[0092] The system is based on the overall confidence level Compared with the preset Level 1 emergency threshold Level II warning threshold Combined with the equipment operating status The event level is determined by a piecewise mapping function. :

[0093] ;

[0094] This logic categorizes events into Level 1 Emergency, Level 2 Warning, Level 3 Maintenance, and Normal Status.

[0095] The system constructs vectorized instructions based on the determined event level. First, it constructs the action vectors of the security devices. This controls the pan-tilt attitude, access control switch status, and alarm mode, respectively. Simultaneously, a large-screen display control vector is constructed. Define the screen display mode, video signal source index, and UI layout template ID respectively.

[0096] In step S4, the system performs dual-track distribution and coordinated control of instructions.

[0097] The system uses the MQTT protocol to publish action vectors of security devices. The system defines topics based on device location and type, with security devices acting as subscribers to receive instructions. This mechanism leverages MQTT's low bandwidth consumption to adapt to edge network environments.

[0098] The system uses the HTTP / 2 protocol to send control vectors for the large screen display. Leveraging the multiplexing and header compression features of HTTP / 2, the large-screen control module quickly loads and renders the static resources and video streams required for multimedia data processing, meeting the high throughput demands of multimedia data.

[0099] The system synchronizes device actions with the displayed content. Upon receiving a command, the security device drives its mechanical structure. The system calculates the estimated time required for the security device to complete its physical actions. :

[0100] ;

[0101] The large screen control module sets the effective display time of the screen based on this time consumption. (in (As a synchronization coefficient), the screen refresh is delayed until the physical equipment reaches a stable state. This timing control logic ensures that the image displayed on the command center's large screen is always clear and accurately aligned with the event scene, avoiding invalid image displays during equipment rotation.

[0102] In step S5, the system performs self-evolution of model parameters.

[0103] The system constructs a training set based on historical event logs, and the samples include input features and manually confirmed real labels. The system constructs a binary cross-entropy loss function. Predicting deviations using quantitative models.

[0104] The system uses the gradient descent algorithm to update the weight parameters:

[0105] ;

[0106] After the update is complete, the system performs a normalized projection on the weight vector to ensure that the sum of all weight components is 1. Through this feedback mechanism, the system can automatically adjust the fusion weights of visual and environmental data based on historical false positive and false negative data for specific scenarios, thereby improving the algorithm's scene adaptability.

[0107] refer to Figure 2 The AI-powered integrated control system for security and large-screen displays in key smart city areas includes:

[0108] The front-end perception and execution module is used to perform frequency conversion acquisition of video stream data, acquire environmental sensing data, and receive motion vectors from security devices to drive the security devices to perform physical actions.

[0109] The edge computing and processing module is used to perform normalization processing on the collected data, extract visual and numerical features in parallel, and perform multi-source data fusion to calculate the comprehensive confidence score.

[0110] The decision-making and scheduling module is used to determine the event level based on the comprehensive confidence level, construct the action vector of the security equipment and the control vector of the large screen display, and perform dual-track distribution through a heterogeneous communication protocol;

[0111] The visualization module is used to receive the control vector of the large screen display and, based on the estimated time delay of the screen refresh sequence according to the mechanical movements of the security equipment, realize the synchronous response of the screen content and the movements of the security equipment.

[0112] The model optimization module is used to construct a loss function based on historical event logs and iteratively update the weight parameters in the edge computing and processing module using the gradient descent algorithm.

Claims

1. A smart city key area security and large screen AI linkage control method, characterized in that, The method comprises the following steps: S1, performing adaptive data acquisition and preprocessing: performing variable frequency acquisition based on target density on video stream data, and performing normalization processing on environmental sensing data; S2, performing parallel feature extraction and multi-source data fusion: respectively extracting visual features of video stream data and numerical features of environmental sensing data, and calculating a comprehensive confidence at the current time through a weighted fusion algorithm; S3, performing event classification determination and instruction construction: determining the event level according to the comprehensive confidence, and respectively constructing a security device action vector for a security device and a large-screen display control vector for a large-screen display terminal according to the event level; S4, performing double-track distribution and cooperative control of instructions: respectively distributing the security device action vector and the large-screen display control vector through heterogeneous communication protocols, and delaying the picture refresh timing of the large-screen display terminal based on the mechanical action prediction time consumption of the security device, to realize the synchronous response of device action and picture content; S5, performing model parameter self-evolution: constructing a loss function based on historical event logs, and iteratively updating the weight parameters in the weighted fusion algorithm using a gradient descent algorithm. 2.The AI linkage control method of smart city key area security and large screen according to claim 1, characterized in that, The variable frequency acquisition based on target density on video stream data specifically comprises: counting the total number of target objects recognized in the current time video frame; comparing the total number of target objects with a preset density threshold; if the total number of target objects is greater than or equal to the density threshold, setting the sampling frame rate of the video acquisition device to the frame rate setting value in the high frequency acquisition mode; if the total number of target objects is less than the density threshold, setting the sampling frame rate of the video acquisition device to the frame rate setting value in the low frequency acquisition mode. 3.The AI linkage control method of smart city key area security and large screen according to claim 1, characterized in that, The extraction of visual features of video stream data specifically comprises: identifying target objects in the video frame using a deep learning model, and obtaining the classification confidence of each target object; calculating the overlap ratio between the predicted bounding box of each target object and the preset effective monitoring area; calculating the product of the classification confidence and the overlap ratio to obtain the region overlap confidence of each target object; determining the maximum value of the region overlap confidence of all target objects as the visual anomaly probability score at the current time. 4.The AI linkage control method of smart city key area security and large screen according to claim 3, characterized in that, The calculation of the comprehensive confidence at the current time through the weighted fusion algorithm specifically comprises: obtaining the environmental anomaly probability score of the environmental sensing data at the current time; obtaining preset visual weight, environmental weight and correlation weight; calculating the first product of the visual weight and the visual anomaly probability score, and the second product of the environmental weight and the environmental anomaly probability score; calculating the product of the visual anomaly probability score and the environmental anomaly probability score, and calculating the operation result of the correlation weight and the product to obtain a third product; calculating the sum of the first product, the second product and the third product to obtain the comprehensive confidence. 5.The AI linkage control method of smart city key area security and large screen according to claim 1, characterized in that, The security device action vector includes pan-tilt posture control components, access control switch state components and alarm trigger mode components; the large-screen display control vector includes screen display mode components, signal source index components and UI layout template components. 6.The AI linkage control method of smart city key area security and large screen according to claim 1, characterized in that, The distributing respectively through the heterogeneous communication protocol specifically includes: A message queue telemetry transport protocol channel based on a publish-subscribe mode is established, and the security device action vector is distributed to a security device; A hypertext transfer protocol version channel based on a multiplexing mechanism is established, and the large screen display control vector is sent to a large screen control module. 7.The AI linkage control method of smart city key area security and large screen according to claim 1, characterized in that, The mechanical action prediction time consumption of the security device is used to delay the picture refresh timing of the large screen display terminal, specifically including: The absolute values of the horizontal direction angle difference and the vertical direction angle difference are calculated respectively by obtaining the posture angle and the target posture angle of the security device at the current time; The horizontal rotation time consumption is obtained by dividing the absolute value of the horizontal direction angle difference by the horizontal rotation angular velocity, and the vertical rotation time consumption is obtained by dividing the absolute value of the vertical direction angle difference by the vertical rotation angular velocity; The maximum value of the horizontal rotation time consumption and the vertical rotation time consumption is determined, and the maximum value is added to a preset focus buffer time to obtain the mechanical action prediction time consumption; The time when the large screen control module receives the large screen display control vector is taken as a reference time, and the sum of the reference time, the mechanical action prediction time consumption, and the product of a synchronization coefficient is calculated to obtain the effective display time of the picture; The large screen control module is controlled to keep the current display content unchanged until the effective display time is reached. 8.The AI linkage control method of smart city key area security and large screen according to claim 4, characterized in that, The weight parameters in the weighted fusion algorithm are updated iteratively using the gradient descent algorithm, specifically including: A binary cross-entropy loss function is constructed using the real label in the historical event log and the comprehensive confidence of the model prediction; The partial derivatives of the loss function with respect to the visual weight, the environment weight, and the correlation weight are calculated respectively; The visual weight, the environment weight, and the correlation weight are updated using the partial derivatives and a preset learning rate; Normalization processing is performed on the updated weight parameters, so that the sum of the weight parameters is equal to one. 9.The AI linkage control method of smart city key area security and large screen according to claim 1, characterized in that, The event level is determined according to the comprehensive confidence, specifically including: A first-level emergency threshold and a second-level warning threshold are obtained, as well as the running state of the device; If the comprehensive confidence is greater than or equal to the first-level emergency threshold, the event level is determined to be first-level emergency; If the comprehensive confidence is less than the first-level emergency threshold and greater than or equal to the second-level warning threshold, the event level is determined to be second-level warning; If the comprehensive confidence is less than the second-level warning threshold and the running state of the device is abnormal, the event level is determined to be third-level maintenance.

10. The AI linkage control system for security and large screen in key areas of smart city according to any one of claims 1-9, characterized in that, It includes: A front-end perception and execution module is used to perform variable frequency acquisition of video stream data, obtain environmental sensing data, and receive a security device action vector to drive a security device to perform physical actions; An edge computing and processing module is used to perform normalization processing on the collected data, extract visual features and numerical features in parallel, and perform multi-source data fusion to calculate a comprehensive confidence; A decision and scheduling module is used to determine the event level according to the comprehensive confidence, construct the security device action vector and the large screen display control vector, and perform double-track distribution through a heterogeneous communication protocol. A visual display module is configured to receive the large-screen display control vector, and to synchronize the picture content with the mechanical action of the security device according to the expected time delay of the refresh timing of the picture. A model optimization module is configured to construct a loss function based on the historical event log, and to iteratively update the weight parameters in the edge computing and processing module using a gradient descent algorithm.

Images