An AI artificial intelligence-based security device and method

Abstract

The application belongs to the technical field of security monitoring, and discloses a security device and method based on AI artificial intelligence, which comprises: all-around monitoring by a look-around monitoring mechanism; environment data acquisition by an environment data acquisition module; image analysis by an image analysis module to identify abnormal conditions; infrared data analysis by an infrared data analysis module to identify abnormal conditions; radar data analysis by a radar data analysis module to identify abnormal conditions; comprehensive analysis by a comprehensive analysis module to predict potential risks; accurate identification of escaped criminals or dangerous elements by a police linkage module; clamping and limiting of escaped criminals or dangerous elements after the warning information of the escaped criminals or dangerous elements is monitored; and prevention of escape. The application realizes multi-dimensional and dead-angle-free monitoring and eliminates monitoring blind areas; the comprehensive analysis module fuses multi-source data to predict potential risks; escaped criminals or dangerous elements are identified; and the driving mechanism clamps and limits the target personnel quickly by using the clamping plate, effectively preventing the target personnel from escaping.

Description

Technical Field

[0001] This application relates to the field of security monitoring technology, and more specifically, to a security device and method based on AI artificial intelligence. Background Technology

[0002] For banks, financial institutions, and confidential organizations, the deployment of artificial intelligence (AI) security equipment is of irreplaceable importance. These organizations have extremely high requirements for access control, and AI technology can integrate biometric features such as facial recognition, iris scanning, and fingerprint recognition, combined with multiple authentication methods such as dynamic passwords, to avoid risks such as traditional access card duplication and password leakage.

[0003] The prior art publication CN113296160A discloses an AI-based security device, including a security gate. The security gate has a locking groove at its bottom, with first electric telescopic rods fixedly installed on both sides of the locking groove. A latch is fixedly installed at one end of each of the first electric telescopic rods, and a second electric telescopic rod is fixedly installed on one side of each first electric telescopic rod. An identification mechanism is fixedly installed on the top side of the security gate. In this security device, an AI camera performs facial recognition on people entering the premises. When the AI ​​camera identifies an unauthorized person via the internet, the first and second electric telescopic rods pop out and engage with each other to secure the latch. This utilizes artificial intelligence to achieve a more effective security effect.

[0004] While the existing technical solutions described above can achieve the relevant beneficial effects through their structure, they still have the following drawbacks: 1. Blind spots exist in monitoring, making it difficult to cover all aspects of the scene. Existing security equipment mostly relies on a single type of camera or sensor with a fixed viewing angle, failing to achieve 360-degree monitoring without blind spots. 2. When identifying abnormal behavior through video surveillance, it is easily affected by factors such as light and obstruction, and its ability to monitor non-human behaviors such as equipment overheating and left-behind items is weak. Alarm rules based on fixed thresholds also frequently result in false alarms and missed alarms, failing to meet the needs of high-precision security. 3. It is mainly based on passive response, lacking risk prediction capabilities, making it difficult to detect the abnormal movement trajectory of suspicious persons or potential equipment malfunctions in advance, and failing to proactively prevent security incidents. 4. Low linkage efficiency and weak emergency response capabilities: Conventional security systems have poor linkage with external police forces or internal security equipment. Data interaction with the public security system is delayed, making it difficult to quickly identify the identity of dangerous persons; when an anomaly is detected, there is a lack of coordinated control between various security devices, making it impossible to effectively intercept target personnel, which can easily lead to the escape of dangerous persons and make it difficult to guarantee the emergency security needs of high-security areas.

[0005] In view of this, we propose a security device and method based on AI artificial intelligence. Summary of the Invention

[0006] 1. The technical problems to be solved.

[0007] The purpose of this application is to provide an AI-based security device and method that solves the technical problems mentioned in the background section. It achieves multi-dimensional, blind-spot-free monitoring. The surround-view monitoring mechanism combines high-definition cameras, radar, and infrared thermal imagers to eliminate blind spots. It identifies anomalies from different dimensions, and the comprehensive analysis module integrates multi-source data, using multi-modal fusion decision trees and historical data for time-series prediction to detect potential risks in advance. The police linkage module quickly identifies fugitives or dangerous individuals through millisecond-level comparison with the public security system database. The control center links with the drive mechanism of the security door components, using clamps to quickly clamp and limit the target personnel, effectively preventing their escape and ensuring the safety of the area.

[0008] 2. Technical solution.

[0009] This application provides a security device based on AI (Artificial Intelligence), including...

[0010] Security door assembly: includes the door frame and two automatic control doors at the front and back.

[0011] Data collection module: Collects geological and building data of the area to be monitored; collects data on personnel allowed to enter, and labels the data as a reference sample.

[0012] Surround monitoring mechanism: includes a rotating mechanism and an infrared thermal imager, with a high-definition camera and radar fixedly mounted on the rotating mechanism; the rotating mechanism drives the high-definition camera and radar to rotate for monitoring, combined with the infrared thermal imager; reducing blind spots in monitoring.

[0013] Environmental data acquisition module: integrates light intensity sensor, rain and snow sensor, visibility meter, temperature and humidity sensor; acquires environmental data (including light intensity, rain and snow, fog, etc.) in real time.

[0014] Image analysis module: Analyzes and identifies the acquired images, promptly detects anomalies, and identifies unauthorized personnel entering the premises.

[0015] Infrared data analysis module: Analyzes and identifies infrared data to promptly detect anomalies.

[0016] Radar data analysis module: Analyzes and identifies radar data to promptly detect anomalies.

[0017] Comprehensive Analysis Module: This module combines the monitoring structures of the image analysis module, infrared data analysis module, radar data analysis module, and environmental data acquisition module to conduct a comprehensive assessment, accurately identify abnormal situations, and predict potential risks by combining historical data.

[0018] Alarm module: Includes an alarm that sounds when an abnormal situation or potential risk is detected.

[0019] Control Center: Network connected to security door components, data collection modules, surround view monitoring mechanisms, environmental data acquisition modules, image analysis modules, infrared data analysis modules, radar data analysis modules, comprehensive analysis modules, and alarm modules.

[0020] Through the above technical solution, a comprehensive monitoring system is implemented using a surround-view monitoring mechanism. An environmental data acquisition module acquires environmental data in real time. An image analysis module analyzes and identifies the acquired images, promptly detecting anomalies such as unauthorized entry. An infrared data analysis module analyzes and identifies infrared data, promptly detecting anomalies. A radar data analysis module analyzes and identifies radar data, promptly detecting anomalies. A comprehensive analysis module combines the monitoring structures of the image analysis module, infrared data analysis module, radar data analysis module, and environmental data acquisition module to conduct a comprehensive assessment, accurately judging anomalies, and predicting potential risks based on historical data. When an anomaly or potential risk is detected, an alarm module promptly issues an alert. When an entrant is detected as an authorized person, the security door assembly opens, allowing entry. When an unauthorized person is detected, both the front and rear automatic control doors remain closed, preventing entry. When an entrant is detected as a fugitive or other person requiring pursuit, the front automatic control door allows entry, then immediately closes, isolating the person between the two automatic control doors to prevent escape.

[0021] This invention provides a security method based on AI, comprising the following steps.

[0022] S1. The data collection module collects geological and building data of the area to be monitored; it also collects data on personnel allowed to enter, and labels the data as a reference sample.

[0023] S2, the surround view monitoring mechanism drives the high-definition camera and radar to rotate and monitor through a rotating mechanism, combined with an infrared thermal imager, to conduct all-round monitoring; the environmental data acquisition module acquires environmental data in real time.

[0024] S3, the image analysis module analyzes and identifies the acquired images, promptly identifying anomalies and situations such as unauthorized personnel entering the premises.

[0025] S4, the infrared data analysis module analyzes and identifies infrared data to promptly detect abnormal situations.

[0026] S5, the radar data analysis module, analyzes and identifies radar data, promptly detecting anomalies.

[0027] S6, the comprehensive analysis module combines the monitoring structures of the image analysis module, infrared data analysis module, radar data analysis module, and environmental data acquisition module to conduct a comprehensive assessment, accurately identify abnormal situations, and predict potential risks by combining historical data.

[0028] S7, the police liaison module, compares the collected data with the public security system database to accurately identify fugitives or dangerous individuals.

[0029] S8. When an abnormal situation or potential risk is detected, the alarm module will issue an alarm in a timely manner.

[0030] S9. After receiving the warning information about the fugitive or dangerous person, the control center sends an instruction to the drive mechanism to close the two door panels; starts motor B, which drives the two clamping plates to move synchronously in opposite directions through the transmission mechanism to clamp and limit the fugitive or dangerous person; and transmits the warning information to the local public security command center and the police terminal of the local police station through the police linkage module, and automatically dials the emergency alarm number.

[0031] 3. Beneficial effects.

[0032] One or more technical solutions provided in this application have at least the following technical effects or advantages.

[0033] 1. This invention achieves multi-dimensional, blind-spot-free monitoring. The surround-view monitoring mechanism combines a high-definition camera, radar, and infrared thermal imager with a rotating mechanism to achieve 360-degree all-round monitoring and eliminate monitoring blind spots.

[0034] 2. Multimodal accurate anomaly recognition: The image, infrared, and radar data analysis modules are based on improved algorithms and models to identify anomalies from different dimensions, covering a variety of situations such as personnel intrusion, abnormal behavior, equipment overheating, and left-behind items, which greatly improves the coverage and accuracy of anomaly detection.

[0035] 3. Intelligent integrated decision-making and risk prediction: The integrated analysis module integrates multi-source data and uses multimodal fusion decision trees and historical data to make time-series predictions. It can not only accurately judge the current anomalies, but also discover potential risks in advance, realize proactive security management, and reduce the probability of security accidents.

[0036] 4. The police linkage module quickly identifies fugitives or dangerous individuals by comparing data with the public security system database in milliseconds; the control center links with the drive mechanism of the security door components to quickly clamp and limit the target person using clamps, effectively preventing their escape and ensuring area security. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating a preferred embodiment of the AI-based security method disclosed in this application.

[0038] Figure 2 This is a schematic diagram of a security door component of an AI-based security device disclosed in a preferred embodiment of this application.

[0039] Figure 3 This is a schematic diagram of a surround-view monitoring mechanism for AI-based security equipment, as disclosed in a preferred embodiment of this application.

[0040] Figure 4 This is a schematic diagram of a rotating device based on AI, as disclosed in a preferred embodiment of this application.

[0041] Reference numerals: 1. Door frame; 2. Drive mechanism; 3. Door panel; 4. Motor A; 5. Clamping plate; 6. Transmission mechanism; 7. Surround view monitoring mechanism; 21. Motor B; 22. Drive gear; 23. Driven gear; 24. Shaft; 25. Bevel gear A; 51. Foam pad; 61. Two-way lead screw; 62. Slide rod; 63. Sliding seat; 64. Connecting rod; 65. Bevel gear B; 71. Positioning seat; 72. Fixed ring seat; 721. Ring groove; 722. Tooth groove; 73. Rotary seat; 731. Arc-shaped baffle; 74. High-definition camera; 75. Motor C; 76. Drive wheel. Detailed Implementation

[0042] The present application will be further described in detail below with reference to the accompanying drawings.

[0043] Reference Figure 1 This application provides an AI-based security device, including...

[0044] Security door assembly: Includes a door frame and two automatic control doors (front and rear); the door frame is made of high-strength aviation-grade aluminum-magnesium alloy, with built-in stress sensors to monitor door frame deformation in real time, achieving an IP67 protection rating; the automatic control doors are driven by servo motors, with a single door opening time of ≤0.8 seconds; equipped with anti-pinch capacitive sensing strips, automatically rebounding when contact pressure exceeds five Newtons; supports power outage emergency mode, capable of sustaining 200 opening and closing operations via UPS power supply. It exchanges access control data with the control center in real time, supporting "dual-person, dual-factor authentication" (such as facial recognition + dynamic password); when unauthorized personnel are detected approaching, the front door will activate an audible and visual warning and delay opening, while simultaneously sending an alert to the control center.

[0045] Data collection module: Collects geological and building data of the area to be monitored; collects data on personnel allowed to enter, and annotates the data as a reference sample; establishes a dynamic personnel database to store facial features, iris texture, fingerprints, ID card information, etc.; supports API integration with enterprise OA systems, automatically synchronizes personnel permission changes, and performs data deduplication and encrypted backup every morning.

[0046] The surround-view monitoring system includes a rotating infrared thermal imager, on which a high-definition camera and radar are fixedly mounted. The rotating mechanism drives the high-definition camera and radar to rotate for monitoring, combined with the infrared thermal imager, reducing blind spots. Under normal conditions, the infrared thermal imager maintains a full-view scan. Upon detecting abnormal heat sources (such as body temperature >38℃ or abnormally high temperatures), the rotating mechanism is triggered, causing the camera and radar to turn towards the target area. The radar calculates the target distance and trajectory in real time, assisting the camera in achieving precise focusing and tracking. The high-definition camera is equipped with an automatic light-adjusting filter. When the environmental data acquisition module detects strong light, the filter automatically adjusts its transmittance to reduce overexposure. The lens uses nano-coating technology, possessing hydrophobic and oleophobic properties, preventing water droplets and snowflakes from adhering in rainy or snowy weather, maintaining a clear field of view. Multi-angle adjustable infrared and LED supplementary lights are installed around the camera. In low light or foggy weather, the intensity and angle of the supplementary light are intelligently adjusted based on environmental data. An ultrasonic snow removal device is also included, automatically activating during snowfall to clear snow from the lens surface.

[0047] Environmental data acquisition module: integrates light intensity sensor, rain and snow sensor, visibility meter, temperature and humidity sensor; acquires environmental data (including light intensity, rain and snow, fog, etc.) in real time.

[0048] Image analysis module: Analyzes and identifies acquired images, promptly detects anomalies, and identifies situations such as unauthorized personnel entering the premises; it uses an improved YOLOv8 algorithm to identify anomalies and supports real-time multi-target detection (speed ≥120fps); it has a built-in behavior recognition model that can distinguish between various abnormal behaviors such as running, falling, and fighting; and it identifies potential risks such as loitering and going against the flow through spatiotemporal trajectory analysis.

[0049] Infrared data analysis module: Analyzes and identifies infrared data to promptly detect anomalies; uses a thermal imaging feature extraction network (TIFNet) to automatically identify overheated equipment, fire sources, and abnormal human body temperatures; supports dynamic background modeling to eliminate interference from ambient temperature fluctuations.

[0050] Radar data analysis module: Analyzes and identifies radar data to promptly detect anomalies; analyzes target motion using the Doppler effect and predicts trajectories using the Kalman filter algorithm; continuously monitors stationary targets (such as abandoned packages) and triggers warnings if the time exceeds a set limit.

[0051] The comprehensive analysis module combines the monitoring structures of the image analysis module, infrared data analysis module, radar data analysis module, and environmental data acquisition module to conduct a comprehensive assessment, accurately identify anomalies, and predict potential risks by incorporating historical data. It constructs a multimodal fusion decision tree, assigning data weights (e.g., increasing the weight of infrared data to 60% at night), performs time-series predictions based on historical data, and outputs the risk level.

[0052] Alarm module: Includes an alarm that promptly issues an alert when abnormal situations or potential risks are detected.

[0053] Control Center: Network connected to security door components, data collection modules, surround view monitoring mechanisms, environmental data acquisition modules, image analysis modules, infrared data analysis modules, radar data analysis modules, comprehensive analysis modules, and alarm modules.

[0054] In this technical solution, a comprehensive monitoring system is implemented using a surround-view monitoring mechanism. An environmental data acquisition module acquires real-time environmental data (including illumination, rain, snow, fog, etc.). An image analysis module analyzes and identifies the acquired images, promptly detecting anomalies such as unauthorized personnel intrusion. An infrared data analysis module analyzes and identifies infrared data, promptly detecting anomalies. A radar data analysis module analyzes and identifies radar data, promptly detecting anomalies. A comprehensive analysis module combines the monitoring structures of the image analysis, infrared data analysis, radar data analysis, and environmental data acquisition modules to conduct a comprehensive assessment, accurately determining anomalies and predicting potential risks based on historical data. When anomalies or potential risks are detected, an alarm module promptly issues an alert. When the system detects that the person entering is an authorized person, the security door assembly opens to allow entry; when the system detects that the person entering is not an authorized person, neither the front nor rear automatic control door opens to prevent entry; when the system detects that the person entering is a fugitive or other person who needs to be apprehended, the front automatic control door allows entry, and then immediately closes the front automatic control door, isolating the person between the two automatic control doors to prevent escape.

[0055] Furthermore, the image analysis module analyzes and identifies the acquired images, promptly detecting any anomalies. This includes the following steps.

[0056] 1. Equipment Parameter Pre-adjustment: The environmental data acquisition module collects environmental parameters such as light intensity, rain / snow level, and fog concentration in real time. Based on the environmental data, the control center pre-adjusts the camera hardware parameters: automatically adjusts the light transmittance using a light-changing filter in strong light; activates the lens's nano-coating hydrophobic function in rainy or snowy weather; and activates multi-angle supplementary lighting in heavy fog or low light conditions, while also adjusting the operating modes of the infrared thermal imager and radar.

[0057] 2. Image Acquisition: High-definition cameras acquire image data. In adverse environmental conditions (such as heavy rain or dense fog), infrared thermal imagers and radar data are used to assist in target location.

[0058] 3. Image preprocessing.

[0059] 3.1 Weather interference elimination.

[0060] Foggy Weather: When fog concentration exceeds the standard, the improved DCPDN dehazing algorithm is invoked to enhance image contrast and clarity. The improved DCPDN dehazing algorithm is as follows.

[0061] J(x)=[I(x)-A] / {max[t(x),t0]}+A; t(x)=1-w×min y∈Ω(x) [I c (y)] / A c In the formula, t(x) represents the transmittance at pixel x in the image, which describes the proportion of light that can pass through a medium such as fog or haze and reach the camera. The lower the transmittance, the greater the obstruction of light by the medium, and the more blurred the image. I(x) represents the pixel value at pixel x in the foggy image, which is a vector containing information from the RGB channels. J(x) represents the pixel value at pixel x in the dehazed image, calculated as the pixel result of the image without fog or with minimal fog impact. A represents the atmospheric light value, that is, the intensity of light scattered by the atmosphere in the scene. In foggy images, the pixel values ​​in distant areas tend to be close to the atmospheric light value. c This is the c-channel value of atmospheric light value A, where c takes values ​​of R (red), G (green), and B (blue), representing the intensity components of atmospheric light in the three color channels. c `(y)` is the c-channel value at pixel `y` in the hazy image `I`, used to calculate the relationship with atmospheric light within a local region. `c` represents R, G, and B. `w` is the dehazing intensity coefficient, typically ranging from [0,1]. `Ω(x)` is the local region centered at pixel `x`, a square window such as 3×3 or 5×5. `t0` is a fixed parameter, a threshold to prevent excessively low transmittance. Because `t(x)` may have excessively small values ​​when calculating transmittance, it can lead to over-enhancement and noise amplification in the dehazed image. `min` is the minimum value operator. `max` is the maximum value operator.

[0062] For both bright and dark scenes: An improved Retinex algorithm is used to compress the dynamic range of bright areas, while EnlightenGAN is combined to enhance the brightness and reduce noise in dark images. The improved Retinex algorithm model is as follows.

[0063] L(x) = I(x) * [1 / W(x)]; R(x) = I(x) / L(x); where I(x) represents the pixel value at pixel x in the original input image, a vector containing information from the RGB channels, representing image data affected by lighting interference. In security scenarios, it corresponds to images captured by high-definition cameras that are affected by strong or dark light. L(x) represents the illumination component at pixel x in the image. It reflects the distribution of light in the image. The calculated L(x) can be used to separate the lighting information in the image for subsequent processing of the reflection component. R(x) represents the reflection component at pixel x in the image. This component removes the influence of lighting factors and mainly reflects the inherent characteristics of the object itself, such as color and texture. W(x) is a Gaussian weighting function, a function based on Gaussian distribution, used to weight the original image I(x) when calculating the illumination component L(x). * indicates a convolution operation.

[0064] Rain and snow weather: The motion trajectories of raindrops and snowflakes are analyzed using a spatiotemporal filtering algorithm to remove interference; a clear image without rain or snow is generated using a GAN algorithm. The rain and snow interference elimination model is as follows.

[0065] I clean (x,y)=I N (x,y)-Σ N-1 i=1 [a i (x,y)M(I N-i V scale )).

[0066] a i (x,y)={e -λi S[I N (x,y)]} / {[Σ N-1 k-1 (e -λk )]S[I N (x,y)]}.

[0067] S[I N [x,y)]=SQT{(1 / |W|)Σ (m,n)∈W [I N (m,n)-I - N] 2}

[0068] M(I N-i ,v scale (x,y)=I N-i [xu scale (x,y),yv scale [x,y]; where I clean(x,y) represents the pixel value at pixel (x,y) in the image after rain and snow interference has been removed; it is the final clear image output by the model. N (x,y) represents the pixel value at pixel (x,y) in the current frame (the Nth frame), which is the original image pixel data before rain and snow interference removal processing. N-i (x,y) represents the pixel value at pixel (x,y) in the i-th frame image preceding the current frame, used in conjunction with the current frame for rain and snow interference cancellation calculations. The value of i ranges from 1 to N-1. i (x,y) is an adaptive weighting coefficient used to measure the contribution of the Ni-th frame image at pixel (x,y) to the elimination of rain and snow interference in the current frame. λ is an adjustable attenuation coefficient, set by the user or system according to the actual scene requirements. The larger the value of λ, the faster the weight of historical frames decreases over time; i is the sequence number of the historical frame. S(I N (x,y) is the spatial importance function, used to evaluate the importance of the current frame image within the local region containing pixel (x,y). W is a local window centered on pixel (x,y). The pixels within the window collectively determine the spatial importance weight of that point. - N It is the average value of pixels within a local window W, serving as an intermediate variable in calculating local variance and used to measure the average pixel value within the window. scale It is a multi-scale motion vector, a combination of coarse-grained and fine-grained motion vectors, used to describe the motion of pixels in the image from historical frames to the current frame, in order to achieve motion compensation. M is the motion compensation function. scale (x,y) and v scale (x, y) is a multi-scale motion vector v scale The horizontal and vertical coordinate components are used to map historical frame pixels to the current frame position. SQT stands for square root operation.

[0069] 3.2 Format Conversion: The processed image is standardized in terms of size, color space, and other formats, and the pixel value range is standardized to provide a unified input for subsequent algorithms.

[0070] 4. Target Detection and Behavior Recognition: Based on the YOLOv8 algorithm, multi-target real-time detection is performed on preprocessed images, marking the location and category of people and objects (authenticated personnel, unauthenticated personnel, etc.), with a detection speed of ≥120fps. The image sequence of the detected target areas is input into a behavior recognition model (such as 3D-CNN) to distinguish abnormal behaviors such as running, falling, and fighting.

[0071] 5. Spatiotemporal Trajectory Analysis: Employing multi-target tracking algorithms such as DeepSORT, each detected target is assigned an ID, generating motion trajectories in consecutive frames. The trajectory data is analyzed, and potential risk behaviors such as loitering and reverse movement are identified through features such as dwell time and direction changes.

[0072] 6. Dynamic weight adjustment: The weight of each data point is dynamically adjusted according to weather conditions: during heavy fog, the focus is on images and infrared data after defogging; during heavy rain, the focus is on images and radar data after rain and snow processing, thereby improving the accuracy of anomaly detection.

[0073] 7. Abnormal Situation Determination: Based on preset rules (such as unauthorized personnel entering, detection of abnormal behavior, etc.), determine whether an abnormality has been triggered. After confirming the abnormality, the image analysis module sends an alarm message to the alarm module, along with detailed data such as the image, target category, and behavior type, triggering subsequent handling procedures.

[0074] 8. Performance Evaluation: The system regularly calculates false alarm rates, missed alarm rates, and other indicators under different weather conditions to evaluate the image analysis performance. If the accuracy of recognizing a certain weather scene decreases, incremental learning or model fine-tuning is automatically triggered to update the image analysis algorithm and optimize its ability to withstand weather interference.

[0075] Furthermore, the infrared data analysis module analyzes and identifies infrared data to promptly detect anomalies; this includes the following steps.

[0076] 1. Infrared Data Preprocessing: Thermal imaging data of the monitored area is continuously acquired using an infrared thermal imager. Combined with the rotating mechanism of the surround-view monitoring system, key areas are dynamically scanned, covering the entire monitoring range. Data preprocessing includes noise reduction and normalization; median filtering or Gaussian filtering is used to eliminate random noise and reduce the impact of environmental electromagnetic interference. Temperature data is normalized, mapping pixel values ​​to the [0,1] interval to unify temperature quantization standards for different scenarios.

[0077] 2. Thermal Imaging Feature Extraction: The preprocessed infrared image is converted into a tensor input to the TIFNet model. Feature extraction network operations are then performed, including low-level feature extraction, mid-level semantic feature extraction, and high-level semantic abstraction.

[0078] Low-level feature extraction: Extract basic features such as temperature gradient and edges through convolutional layers (such as 3×3 convolutional kernels).

[0079] Mid-level semantic features: Residual blocks or attention mechanisms are used to enhance the feature representation of temperature anomaly regions.

[0080] High-level semantic abstraction: Extract global features of thermal imaging through a global pooling layer to identify thermal patterns of targets such as overheated equipment, fire sources, and human bodies.

[0081] Feature vector generation outputs a 128-dimensional feature vector containing key features such as temperature distribution, hot spot shape, and dynamic temperature change rate.

[0082] 3. Abnormal target identification: including identification of overheated equipment and fire sources and identification of abnormal human body temperature.

[0083] Overheating equipment and fire source identification: Set temperature thresholds (e.g., equipment surface temperature > 60℃, fire source temperature > 400℃), and combine the hotspot area and temperature gradient in the feature vector for judgment. Non-maximum suppression (NMS) is used to filter overlapping hotspots to avoid repeated alarms.

[0084] Abnormal body temperature detection: Locates human body regions (based on human contour features in infrared thermography) and extracts the temperature of exposed areas such as the forehead. Compares the temperature with normal body temperature thresholds (e.g., >38℃) and triggers an alert based on dynamic time series analysis.

[0085] 4. Dynamic Background Modeling and Interference Removal: Initial background model construction: Upon system startup, 100 frames of infrared images of anomaly-free scenes are acquired, and an initial background model is established using a Gaussian Mixture Model (GMM) or the median method. The background model is dynamically updated using an online learning mechanism, updating its parameters every 100 frames. For stable background areas (such as walls and fixed equipment), the mean and variance are updated; for temporary interference (such as brief passage of people), Kalman filtering is used to predict the background state and avoid misjudgments. Environmental interference filtering is performed by calculating the difference image between the current frame and the background model, and foreground targets are extracted through threshold segmentation. Morphological operations (erosion, dilation) are combined to remove minor noise and retain true anomaly targets.

[0086] 5. Results Output: Push abnormal data (coordinates, temperature, risk level) to the comprehensive analysis module and store it in the historical database. Generate infrared data statistical reports every hour for trend analysis and model optimization.

[0087] Furthermore, the radar data analysis module analyzes and identifies radar data to promptly detect anomalies; this includes the following steps.

[0088] 1. Radar Data Preprocessing: Millimeter-wave radar (such as 77GHz) or lidar (LiDAR) is used to collect point cloud data or echo signals. Preprocessing of the radar data includes: point cloud denoising, data dimensionality reduction, and time synchronization.

[0089] Point cloud denoising: Outliers are removed through statistical filtering.

[0090] Data dimensionality reduction: Voxel grid downsampling is performed on high-density point clouds to reduce computational load.

[0091] Time synchronization: Align radar frames with timestamps from other sensors with an error of ≤10ms.

[0092] 2. Target Feature Extraction: Point cloud clustering and segmentation are performed using the Euclidean clustering algorithm to classify adjacent points in space into independent targets (such as people, vehicles, and packages). Clustering parameters are set: neighborhood radius (eps) is 0.5–1.5 meters, and minimum number of points (minPts) is five to ten. Target features are then extracted, including geometric and motion features.

[0093] Geometric features: Calculate the target volume, surface area, aspect ratio, etc.

[0094] Motion characteristics: Radial velocity was calculated based on Doppler frequency shift (through correlation analysis between radar wavelength and frequency shift value).

[0095] 3. Motion state analysis and trajectory prediction.

[0096] 3.1 Kalman filter initialization: A Kalman filter is established for each target, and the state vector contains three-dimensional position and velocity components.

[0097] 3.2 State prediction update.

[0098] Prediction phase: Estimate the current position based on the state of the previous moment.

[0099] Update phase: Integrate current observations to revise the prediction results.

[0100] 3.3 Trajectory Association Management: The Hungarian algorithm combined with Mahalanobis distance is used for target association, with a threshold set to 3σ. The trajectory lifecycle is maintained: the presence of the target is confirmed after ≥3 consecutive detections, and the trajectory is deleted after ≥5 consecutive undetected frames.

[0101] 4. Monitoring and early warning of stationary targets.

[0102] Stationary target determination: Targets with a speed of less than 0.1 m / s and a duration of ≥ 10 seconds are marked as stationary, and the stationary state is verified by position variance.

[0103] Abandoned object detection: A timer is started to record the duration of stationary targets. If the duration exceeds a set threshold (e.g., 30 minutes), an alert is triggered. Multi-frame point cloud comparison is used to confirm that the target has not changed. The background model is dynamically updated to eliminate interference from fixed facilities; the detection threshold for stationary targets is increased in densely populated areas.

[0104] 5. Abnormal behavior identification and risk assessment.

[0105] Abnormal movement pattern detection: including loitering behavior and retrograde behavior detection.

[0106] Loitering behavior: Calculate the entropy value of the target trajectory. If the entropy value is less than 0.5 and the dwell time is greater than five minutes, it is judged as loitering.

[0107] Reverse movement behavior: The angle between the target's movement direction and the preset travel direction exceeds 60°, triggering a warning.

[0108] Risk Level Assessment: The risk level is output based on a combination of target speed, distance to sensitive areas, and degree of abnormality in movement patterns. If the risk level is ≥ Medium, an early warning message is sent to the comprehensive analysis module, triggering confirmation via cameras and infrared thermal imagers.

[0109] Furthermore, the comprehensive analysis module combines the monitoring structures of the image analysis module, infrared data analysis module, radar data analysis module, and environmental data acquisition module to conduct a comprehensive assessment, accurately identify abnormal situations, and predict potential risks by combining historical data; including the following steps.

[0110] 1. Multimodal Data Preprocessing: Real-time reception of output data from image analysis module (video stream, target recognition results), infrared data analysis module (thermal imaging features), radar data analysis module (point cloud trajectory, velocity), and environmental data module (temperature, humidity, illumination, etc.). A timestamp alignment mechanism is employed to ensure that the time error of each modality's data is ≤50ms, and data is cached using a sliding window. Data preprocessing includes normalization and missing value handling; missing value handling employs forward imputation or mean interpolation based on historical data.

[0111] 2. Multi-source data feature fusion: The coordinates of targets from various modalities are uniformly transformed to the world coordinate system, and the same physical target is associated through a spatial distance threshold. A target association table is established to record the features of each modality (such as visual appearance, thermal imaging temperature, and radar velocity). The feature fusion strategy is as follows.

[0112] Early fusion: Fusion at the raw data layer (such as overlaying infrared images with visible light images), suitable for environmental visualization.

[0113] Late-stage fusion: The abnormal probabilities output by each module are fused at the decision level, and a voting mechanism is adopted (if more than two modules trigger alarms, an early warning is activated). The feature fusion model is as follows.

[0114] P 融合 ={Σ D i=1 [w i Pi×e -δ|Pi-PA| ]} / [Σ D i=1 (w i ×e -θ|Pi-PA| )];Σ D i=1 w i =1.

[0115] w i =[(A i +E i )γ t (1+δ×EnvF)] / {ΣD n i=1 [(A i +E i )γ t (1+δ×EnvF)]};where, P 融合 This is the final anomaly probability value after fusion, ranging from [0,1]. D is the number of modalities involved in the fusion (e.g., vision, infrared, radar, etc.). Pi is the anomaly probability of the i-th modality output (e.g., the probability of the vision module judging something as "abnormal" is 0.8). w i is the dynamic weight of the i-th mode, determined by environmental and historical data. PA is the mean of the anomaly probabilities of all modes; θ is the conflict suppression coefficient (default value is 2), controlling the sensitivity to modal divergence: the larger θ is, the faster the weight decays when there are large differences between modes. e is the base of a natural number. A i E is the historical accuracy of the i-th modality (based on sliding window statistics over the past 24 hours; for example, the accuracy of the infrared module at night is 0.9). i γ is the environmental matching degree (0-1) interval for the i-th modality, calculated from the current environmental data (e.g., 0.8 for nighttime infrared). γ is the time decay factor (default 0.95), which gives higher weight to recent data. t is the time step (in hours / frames, depending on the data update frequency). δ is the environmental sensitivity coefficient (default 0.8), which controls the influence of the environmental correction term. EnvF is the environmental correction term (dimensionless): EnvF = 0.5 for nighttime / low-light scenes; EnvF = 0.3 for rainy / snowy weather.

[0116] 3. Construction of multimodal fusion decision model.

[0117] 3.1 Decision Tree Initialization. The basic model is constructed using the CART (Classification and Regression Tree) algorithm, with input features including...

[0118] Visual characteristics (target category, behavioral pattern).

[0119] Infrared characteristics (temperature anomalies, hot spot area).

[0120] Radar characteristics (speed, trajectory entropy value).

[0121] Environmental characteristics (light intensity, rain and snow conditions).

[0122] 3.2 Dynamic weight setting: The modal weights are automatically adjusted according to environmental conditions.

[0123] Nighttime / low-light scenarios: Infrared data weighting is increased to 60%, while visual data weighting is reduced to 30%.

[0124] Rainy or snowy weather: Radar data weighting increased to 50% to reduce visual misjudgment.

[0125] The weight calculation formula is determined through training with historical data.

[0126] After every 100 decisions, the model is incrementally trained using newly labeled data, and the split nodes and weight parameters are updated. Pruning strategies (such as cost complexity pruning) are employed to avoid overfitting and preserve key decision paths.

[0127] 4. Historical data time series prediction and risk assessment.

[0128] 4.1 Time Series Data Modeling: Establish a sliding window to extract the time distribution characteristics of historical anomaly data (e.g., a high incidence of anomalies between 2 AM and 4 AM). Use an ARIMA or LSTM model to predict the probability of risk in future periods. Inputs include...

[0129] Time series of historical anomalous events.

[0130] Periodic characteristics (weekdays / weekends, seasonal variations).

[0131] 4.2 Potential Risk Prediction: Combining real-time data with prediction results, identify trend risks: If a target loiters in a certain area for three consecutive hours, and the theft incidents in the same area have been frequent in the same period in history, then it is predicted to be high risk.

[0132] 4.3 Quantitative Output of Risk Level: Define risk levels (low / medium / high / urgent), and include assessment dimensions.

[0133] The severity of the situation (e.g., fire source vs. people loitering).

[0134] Scope of impact (single target vs. regional anomaly).

[0135] Prediction confidence (model output probability).

[0136] 5. Early warning linkage: Multi-module linkage response, triggering different measures according to risk level.

[0137] Low risk: Record to logs and push to security personnel's mobile devices.

[0138] High risk: The linkage alarm module provides audible and visual warnings, automatically locks onto the camera to track the target, and sends a text message to the emergency response manager.

[0139] Reference Figure 2 The security door assembly includes a door frame 1, a drive mechanism 2, a door panel 3, a motor A4, a clamping plate 5, and a transmission mechanism 6.

[0140] Door panels 3 are rotatably mounted at both ends of the door frame 1; a rectangular channel is provided inside the door frame 1; motors A4 are fixedly mounted at both ends of the top of the door frame 1, and the output end of motor A4 is fixedly connected to the corresponding door panel 3, and motor A4 can drive the door panel 3 to rotate.

[0141] Two symmetrically sliding clamping plates 5 are installed inside the door frame 1; a drive mechanism 2 is fixedly installed above the door frame 1; a transmission mechanism 6 is installed inside the door frame 1, and the transmission mechanism 6 is connected to the drive mechanism 2; the transmission mechanism 6 is also connected to the corresponding clamping plate 5. The drive mechanism 2 drives the two clamping plates 5 to move synchronously in opposite directions through the transmission mechanism 6. A speaker and a camera are fixedly installed at the top inside the door frame 1. The speaker can be used to issue audible and visual warnings, while the camera can monitor the situation inside the passage in real time, enhancing security.

[0142] In this technical solution, the drive mechanism 2 drives the two clamping plates 5 to move synchronously in opposite directions through the transmission mechanism 6, so that the two clamping plates 5 clamp and limit the fugitive or dangerous person to prevent him from escaping.

[0143] Furthermore, the drive mechanism 2 includes a motor B21, a drive gear 22, a driven gear 23, a shaft 24, and a bevel gear A25.

[0144] Motor B21 is fixedly mounted above door frame 1; a drive gear 22 is coaxially fixedly mounted on the output end of motor B21; a shaft 24 is rotatably mounted above door frame 1, and a driven gear 23 is coaxially fixedly mounted on shaft 24; bevel gears A25 are coaxially fixedly mounted on both ends of shaft 24; bevel gears A25 are meshed and connected to transmission mechanism 6.

[0145] In this technical solution, the motor B21 is started, which drives the driven gear 23, shaft 24 and bevel gear A25 to rotate through the driving gear 22. The bevel gear A25 drives the transmission mechanism 6 to move, so that the clamping plate 5 moves.

[0146] Furthermore, the transmission mechanism 6 includes a two-way lead screw 61, a slide bar 62, a sliding seat 63, a connecting rod 64, and a bevel gear B65.

[0147] A double-acting screw 61 is provided on both sides of the door frame 1, and a slide rod 62 is provided on both sides of the double-acting screw 61 on the door frame 1; two sliding seats 63 are symmetrically threaded on the double-acting screw 61, and the sliding seats 63 are slidably connected to the slide rod 62.

[0148] Two connecting rods 64 are rotatably mounted on the sliding seat 63; the connecting rods 64 are rotatably mounted on the corresponding clamping plate 5.

[0149] A bevel gear B65 is coaxially fixed at the upper end of the double-acting lead screw 61, and the bevel gear B65 meshes with the bevel gear A25 for transmission.

[0150] In this technical solution, when the drive mechanism 2 drives the bevel gear A25 to rotate, the bevel gear A25 drives the double-acting lead screw 61 to rotate, and the double-acting lead screw 61 drives the two sliding seats 63 mounted on it to move synchronously in opposite directions. The sliding seats 63 drive the clamping plate 5 to move through the connecting rod 64.

[0151] Furthermore, a foam pad 51 is fixedly installed on the inner side of the clamping plate 5, and several pressure sensors are evenly distributed and fixed on the foam pad 51. The foam pad 51 is soft in texture, which can avoid serious injury to personnel during clamping, while providing a certain degree of cushioning. Several pressure sensors are evenly distributed and fixed on the foam pad 51, and these pressure sensors can monitor the clamping pressure of the clamping plate 5 on the target personnel in real time. When the pressure sensor detects that the pressure exceeds the safety threshold, it will feed the signal back to the control system. The control system can adjust the clamping force of the clamping plate 5 according to the preset program to prevent injury to personnel due to excessive clamping force, thus ensuring safety protection while taking into account humanitarian principles.

[0152] Reference Figure 3 and Figure 4 The surround view monitoring mechanism includes a positioning seat 71, a fixed ring seat 72, a rotating seat 73, an arc baffle 731, a high-definition camera 74, a motor C75, and a drive wheel 76.

[0153] The positioning seat 71 is fixedly installed in the required position by bolts or expansion bolts.

[0154] A fixing ring seat 72 is fixedly provided on the positioning seat 71; the fixing ring seat 72 has an annular groove 721, and a toothed groove 722 is provided on the outer side of the annular groove 721.

[0155] A rotating seat 73 is slidably mounted on the fixed ring seat 72; two high-definition cameras 74 are fixedly mounted on the rotating seat 73, one high-definition camera 74 facing downwards and the other high-definition camera 74 facing forwards (at a certain tilt angle).

[0156] A motor C75 is fixedly mounted on top of the rotary seat 73; a drive wheel 76 is fixedly mounted coaxially on the output end of the motor C75; the drive wheel 76 is slidably mounted in the annular groove 721, and the drive wheel 76 is meshed with the tooth groove 722 for transmission.

[0157] An arc-shaped baffle 731 is fixedly installed on the inner side of the rotary base 73; the arc-shaped baffle 731 slides with the rotary base 73.

[0158] In this technical solution, the starter motor C75 drives the drive wheel 76 to rotate. Since the drive wheel 76 is meshed with the fixed ring seat 72, and the fixed ring seat 72 remains stationary, the drive wheel 76 drives the rotating base 73 and the high-definition camera 74 to rotate, achieving all-around monitoring and avoiding blind spots. The different orientations and angles of the two high-definition cameras 74 form a multi-view monitoring system covering vertical, horizontal, front, and rear angles. During the rotation of the rotating base 73, the image data collected by the two cameras complement each other. Through fusion processing by the image analysis module, a more comprehensive and three-dimensional monitoring scene model can be constructed, greatly improving the accuracy of identifying abnormal situations in complex environments and effectively avoiding safety hazards caused by blind spots.

[0159] Furthermore, a police linkage module is fixedly installed on door frame 1. The hardware of this module includes a dedicated encrypted communication module, a high-performance data processing unit, and a storage module. The encrypted communication module uses national cryptographic algorithms to ensure the security of data interaction with the public security system; the data processing unit supports multi-threaded parallel processing to quickly handle massive identity comparison requests; the storage module is used to temporarily cache monitoring data and interaction information. The police linkage module integrates a real-time identity recognition algorithm, comparing facial and iris biometric features captured by high-definition cameras with the public security system's fugitive database and key monitoring personnel database at millisecond levels; it also features an intelligent early warning rule engine that can set response strategies according to different risk levels. This includes the following steps.

[0160] 1. Data Preprocessing: High-definition cameras continuously collect biometric image data such as facial features and iris scans, and transmit the raw data as a video stream to the police linkage module. The data processing unit receives and parses the video stream, extracting valid image data from each frame. Data standardization is performed, with the storage module temporarily caching the raw image data. The data processing unit then performs standardization operations, including image grayscale conversion and size normalization, converting images of different resolutions and formats into a standard format recognizable by the algorithm. Simultaneously, the data is anonymized, retaining only key biometric information.

[0161] 2. Real-time Identity Verification: The data processing unit initiates a real-time identity recognition algorithm, segmenting the preprocessed biometric image data into feature vectors, such as extracting key facial coordinates and iris texture features. The encrypted communication module retrieves relevant data from the public security system's fugitive database and key monitoring personnel database, transmitting it to the data processing unit. The data processing unit employs multi-threaded parallel processing, performing millisecond-level comparisons between the feature vector to be compared and the template feature vectors in the database, calculating a similarity score. The comparison result is then judged based on preset similarity thresholds (e.g., 90% for facial recognition and 95% for iris recognition) to determine if a match is successful. If the similarity score exceeds the threshold, it is determined that a fugitive or dangerous individual has been identified, generating preliminary warning information; if it does not exceed the threshold, the data is stored as an anonymous sample in the temporary cache of the storage module, retained for a certain period, and then automatically deleted.

[0162] 3. Intelligent Early Warning: The intelligent early warning rule engine classifies individuals into risk levels based on identified information (such as crime type, criminal record, etc.) and pre-defined rules. For example, fugitives involved in violent crimes are classified as Level 1 risk, fugitives involved in ordinary economic crimes as Level 2 risk, and suspicious individuals under close monitoring as Level 3 risk. The data processing unit generates a complete early warning data package, including text descriptions, images, and video clips, based on the risk level, target location, and real-time imagery. 4. Tiered Response Execution.

[0163] Level 1 Danger: The encrypted communication module immediately transmits the warning data packet with the highest priority to the local public security command center and the police terminal of the local police station. At the same time, it automatically dials the emergency alarm number to explain the situation to the police in detail. The control center sends a command to the drive mechanism 2 to start the motor B21, which drives the two clamps 5 to move synchronously in the opposite direction through the transmission mechanism 6 to clamp and limit the target person. The system activates the panoramic recording function and uploads the on-site video in encrypted form to the public security system in real time.

[0164] Level 2 Hazard: The encrypted communication module transmits the warning data packet to the relevant police terminal. The control center activates the warning mode of the security door component. For example, the front door will activate an audible and visual warning and delay opening. At the same time, the warning information will be pushed to security personnel. The storage module will perform long-term backup of the relevant data of the event for subsequent investigation.

[0165] Level 3 Danger: The encrypted communication module only sends warning alerts to the local security personnel's terminals, reminding them to pay closer attention; the system records the personnel's activity trajectory and continuously monitors their behavior.

[0166] In this technical solution, once the high-definition camera captures a person's biometric features, the police linkage module immediately activates a real-time identity recognition algorithm, comparing the collected data with the public security system database. If a fugitive or dangerous individual is identified, the system automatically generates a warning message containing the target person's location, characteristics, and real-time image, and classifies the danger level according to preset rules (e.g., Level 1 danger is for violent crime fugitives, Level 2 danger is for ordinary fugitives). Upon receiving the warning message, the control center sends a command to the drive mechanism 2 to start motor B21, which drives the two clamps 5 to move synchronously in opposite directions through the transmission mechanism 6, clamping and limiting the fugitive or dangerous individual. Simultaneously, the police linkage module encrypts and transmits the warning message to the local public security command center and the police terminal of the jurisdictional police station, and automatically dials the emergency alarm number. At the same time, the system activates the panoramic recording function, uploading the on-site video to the public security system in real time, providing decision-making basis for police response.

[0167] This invention provides a security method based on AI, comprising the following steps.

[0168] S1. The data collection module collects geological and building data of the area to be monitored; it also collects data on personnel allowed to enter, and labels the data as a reference sample.

[0169] S2, the surround view monitoring mechanism drives the high-definition camera and radar to rotate and monitor through a rotating mechanism, combined with an infrared thermal imager, to conduct all-round monitoring; the environmental data acquisition module acquires environmental data (including light, rain, snow, fog, etc.) in real time.

[0170] S3, the image analysis module analyzes and identifies the acquired images, promptly identifying anomalies and situations such as unauthorized personnel entering the premises.

[0171] S4, the infrared data analysis module analyzes and identifies infrared data to promptly detect abnormal situations.

[0172] S5, the radar data analysis module, analyzes and identifies radar data, promptly detecting anomalies.

[0173] S6, the comprehensive analysis module combines the monitoring structures of the image analysis module, infrared data analysis module, radar data analysis module, and environmental data acquisition module to conduct a comprehensive assessment, accurately identify abnormal situations, and predict potential risks by combining historical data.

[0174] S7, the police liaison module, compares the collected data with the public security system database to accurately identify fugitives or dangerous individuals.

[0175] S8. When an abnormal situation or potential risk is detected, the alarm module will issue an alarm in a timely manner.

[0176] S9. Upon receiving a warning message about a fugitive or dangerous individual, the control center sends an instruction to the drive mechanism 2 to open the front door panel 3, allowing the fugitive to enter the door frame 1, and then closes both door panels 3. The motor B21 is activated, driving the two clamping plates 5 to move synchronously in opposite directions via the transmission mechanism 6, thus clamping and limiting the fugitive or dangerous individual. The warning message is then encrypted and transmitted to the local public security command center and the police terminal of the jurisdictional police station through the police linkage module, and an emergency alarm is automatically dialed. Simultaneously, the system activates the panoramic recording function, uploading the on-site video to the public security system in real time to provide decision-making support for police response.

[0177] This invention achieves multi-dimensional, blind-spot-free monitoring. The surround-view monitoring mechanism combines high-definition cameras, radar, and infrared thermal imagers, along with a rotating mechanism, to achieve 360-degree all-around monitoring, eliminating blind spots. Multi-modal, precise anomaly identification: image, infrared, and radar data analysis modules, based on improved algorithms and models, identify anomalies from different dimensions, covering various situations such as intrusion, abnormal behavior, equipment overheating, and left-behind items, significantly improving the coverage and accuracy of anomaly detection. Intelligent comprehensive decision-making and risk prediction: the comprehensive analysis module integrates multi-source data, using multi-modal fusion decision trees and historical data for time-series prediction. This not only accurately identifies current anomalies but also proactively detects potential risks, enabling proactive security management and reducing the probability of security incidents. Police linkage module: through millisecond-level comparison with the public security system database, it quickly identifies fugitives or dangerous individuals; the control center, in conjunction with the drive mechanism of the security door components, uses clamps to quickly clamp and limit the target personnel, effectively preventing escape and ensuring area security.

[0178] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A security method based on AI (Artificial Intelligence), characterized in that, Includes the following steps: S1, the data collection module collects geological data, building data, and data on permitted personnel. S2: The surrounding monitoring agency conducts comprehensive monitoring; the environmental data acquisition module acquires environmental data. S3, the image analysis module analyzes and identifies the acquired images to detect abnormal situations; S4, the infrared data analysis module analyzes and identifies infrared data to detect anomalies; S5, the radar data analysis module analyzes and identifies radar data, and identifies abnormal situations; S6, the comprehensive analysis module combines multi-source data to conduct a comprehensive assessment and predict potential risks; S7, the police liaison module accurately identifies fugitives or dangerous individuals; S8. When an abnormal situation or potential risk is detected, the alarm module will issue an alarm in a timely manner. S9. Upon detecting a warning message about a fugitive or dangerous individual, open the front door panel of the security door assembly to allow them to enter the door frame, and close both door panels; start motor B of the drive mechanism, which drives the two clamping plates to move synchronously in opposite directions through the transmission mechanism to clamp and limit the fugitive or dangerous individual; prevent escape, and automatically dial the emergency alarm number.

2. The AI-based security method according to claim 1, characterized in that: Step S3 includes the following steps: S31. Pre-adjustment of equipment parameters: Based on environmental data, pre-adjust the camera hardware parameters; S32. Image Acquisition: High-definition cameras acquire image data. In harsh environments, infrared thermal imagers and radar data assist in locating targets. S33. Image preprocessing: including weather interference removal and format conversion; S34. Target Detection and Behavior Recognition: Based on the YOLOv8 algorithm, perform real-time multi-target detection on preprocessed images, mark the location and category of people and objects, and distinguish abnormal behaviors; S35. Spatiotemporal trajectory analysis: The DeepSORT multi-target tracking algorithm is used to assign an ID to each detected target, generate motion trajectories in consecutive frames, analyze trajectory data, and identify potential risk behaviors. S36. Dynamic weight adjustment: The weights of each data point are dynamically adjusted according to weather conditions. S37. Abnormal Situation Judgment: Determine whether an abnormality is triggered based on preset rules; S38. Effect Evaluation: Evaluate the image analysis effect and update the image analysis algorithm.

3. The AI-based security method according to claim 1, characterized in that: Step S4 includes the following steps: S41. Infrared data preprocessing: Preprocess the data, including noise reduction and normalization; S42. Thermal imaging feature extraction: Perform feature extraction network operations, including low-level feature extraction, mid-level semantic features, and high-level semantic abstraction; S43. Abnormal Target Identification: including identification of overheated equipment and fire sources and identification of abnormal human body temperature; S44. Dynamic Background Modeling and Interference Removal: Initial background model construction, environmental interference filtering, morphological operations to eliminate minor noise, and preservation of real abnormal targets; S45. Output Results: Push abnormal data to the comprehensive analysis module and store it in the historical database.

4. The AI-based security method according to claim 1, characterized in that: Step S5 includes the following steps: S51. Radar data preprocessing: Preprocessing radar data, including point cloud denoising, data dimensionality reduction, and time synchronization; S52. Target Feature Extraction: Perform point cloud clustering and segmentation, and extract target features, including geometric features and motion features; S53, Motion State Analysis and Trajectory Prediction; S54. Monitoring and early warning of stationary targets; S55. Abnormal Behavior Identification and Risk Assessment: Detect abnormal movement patterns, including loitering and reverse movement; output risk level based on target speed, distance to sensitive areas, and degree of abnormality in movement pattern.

5. The AI-based security method according to claim 1, characterized in that: Step S6 includes the following steps: S61, Multimodal data preprocessing: Real-time reception of data from image analysis module, infrared data analysis module, radar data analysis module and environmental data module, and preprocessing of the data; S62. Multi-source data feature fusion: The coordinates of targets from various modalities are uniformly transformed to the world coordinate system, and the same physical target is associated through a spatial distance threshold. S63, Construction of a multimodal fusion decision model; S64. Historical Data Time Series Prediction and Risk Assessment; S65. Early warning linkage: Multi-module linkage response, triggering different measures according to risk level.

6. The AI-based security method according to claim 5, characterized in that: Step S64 includes the following steps: S64.1 Time Series Data Modeling: Establish a sliding window to extract the time distribution characteristics of historical abnormal data; use the ARIMA model to predict the risk probability in future periods; S64.2 Potential Risk Prediction: Identify trend risks by combining real-time data with prediction results. S64.3, Risk Level Quantification Output: Define and output the risk level.

7. The AI-based security method according to claim 1, characterized in that: Step S7 includes the following steps: S71. Data Preprocessing: Parse the video stream and extract valid image data from each frame; perform data standardization processing on the data. S72. Real-time identity comparison: The preprocessed biometric image data is segmented into feature vectors, and the feature vectors to be compared are compared with the template feature vectors in the database in milliseconds to calculate the similarity score; to identify fugitives or dangerous individuals. S73, Intelligent Early Warning: Based on the identified personnel information and pre-set rules, the danger level is classified; S74, Hierarchical response execution.

8. The AI-based security method according to claim 1, characterized in that: The security door assembly includes a door frame, a drive mechanism, a door panel, a motor A, a clamping plate, and a transmission mechanism; Door panels are rotatably mounted at both ends of the door frame; a rectangular channel is provided inside the door frame; motor A is fixedly mounted at both ends of the top of the door frame, and the output end of motor A is fixedly connected to the corresponding door panel, which can drive the door panel to rotate; two clamping plates are symmetrically and slidably mounted inside the door frame; a drive mechanism is fixedly mounted above the door frame; a transmission mechanism is provided inside the door frame, which is connected to the drive mechanism; the transmission mechanism is connected to the corresponding clamping plate, and the drive mechanism drives the two clamping plates to move synchronously in opposite directions through the transmission mechanism.

9. The AI-based security method according to claim 8, characterized in that: The drive mechanism includes motor B, driving gear, driven gear, shaft and bevel gear A; Motor B is fixedly mounted above the door frame; a drive gear is coaxially fixedly mounted on the output end of motor B; a shaft is rotatably mounted above the door frame, and a driven gear is coaxially fixedly mounted on the shaft; bevel gears A are coaxially fixedly mounted on both ends of the shaft; bevel gears A are meshed and connected to the transmission mechanism.

10. The AI-based security method according to claim 9, characterized in that: The transmission mechanism includes a two-way lead screw, a slide bar, a sliding seat, a connecting rod, and a bevel gear B; A double-acting screw is rotatably mounted on both sides of the door frame, and slide rods are mounted on both sides of the double-acting screw on the door frame. Two sliding seats are symmetrically threaded on the double-acting screw, and the sliding seats are slidably connected to the slide rods. Two connecting rods are rotatably mounted on the sliding seats. The connecting rods are rotatably mounted on the corresponding clamps. A bevel gear B is coaxially fixed at the upper end of the double-acting screw, and bevel gear B is meshed with bevel gear A for transmission.

11. The AI-based security method according to claim 10, characterized in that: The surround-view monitoring mechanism includes a positioning seat, a fixed ring seat, a rotating seat, an arc-shaped baffle, a high-definition camera, a motor C, and a drive wheel. The positioning seat is fixedly mounted on the fixed ring seat. The fixed ring seat has an annular groove with toothed grooves on the outer side. The rotating seat is slidably mounted on the fixed ring seat. Two high-definition cameras are fixedly mounted on the rotating seat, and the motor C is fixedly mounted on the rotating seat. The output end of the motor C is coaxially fixedly mounted on the drive wheel. The drive wheel is slidably mounted in the annular groove and is connected to the toothed groove for transmission.

12. The AI-based security method according to claim 1, characterized in that: A foam pad is fixedly installed on the inner side of the clamp, and several pressure sensors are evenly distributed and fixed on the foam pad.

13. A security device based on AI (Artificial Intelligence), comprising: The system comprises a control center, security door components, a data collection module, a surround-view monitoring mechanism, an environmental data acquisition module, an image analysis module, an infrared data analysis module, a radar data analysis module, a comprehensive analysis module, and an alarm module; its features are: Security door components: including the door frame and two automatically controlled doors at the front and back; Data collection module: collects geological data, building data, and data on permitted personnel access; Surround monitoring mechanism: includes a rotating mechanism and an infrared thermal imager, with a high-definition camera and radar fixedly mounted on the rotating mechanism; the rotating mechanism drives the high-definition camera and radar to rotate for monitoring. Environmental data acquisition module: Acquires environmental data in real time; Image analysis module: Analyzes and identifies the acquired images to detect anomalies; Infrared data analysis module: Analyzes and identifies infrared data to detect anomalies; Radar data analysis module: Analyzes and identifies radar data to detect anomalies; Comprehensive Analysis Module: Combines data from multiple sources to conduct a comprehensive assessment and predict potential risks and problems; Alarm module: includes an alarm that sounds when an abnormal situation or potential risk is detected; Control Center: Network connected to security door components, data collection modules, surround view monitoring mechanisms, environmental data acquisition modules, image analysis modules, infrared data analysis modules, radar data analysis modules, comprehensive analysis modules, and alarm modules.

Images