Intelligent security and protection monitoring method and system based on image recognition

Abstract

The invention discloses an intelligent security and protection monitoring method and system based on image recognition. The method comprises the following steps: S1, acquiring multi-modal data; s2, performing preprocessing to generate complete target information; s3, performing real-time detection on the generated complete target information, and performing primary processing on a conventional scene; s4, based on an improved visual SLAM technology, obtaining three-dimensional information of multiple cameras; s5, when the edge device detects a potential abnormal condition, performing abnormal behavior analysis on the data by using a Swin Transform model, and optimizing a detection result in combination with the three-dimensional information of the multiple cameras; s6, performing adaptive learning on historical data by using a depth deterministic strategy gradient algorithm, and dynamically adjusting a threshold value of abnormal behavior detection; and S7, after an abnormal condition is detected, processing data through differential privacy protection and an end-to-end encryption technology, automatically controlling security and protection equipment, and giving a real-time alarm. According to the invention, image recognition and deep learning technologies are fully utilized, and accurate real-time monitoring in a complex scene is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent security monitoring, and in particular to an intelligent security monitoring method and system based on image recognition. Background Art

[0002] With the acceleration of urbanization and the increase in social security needs, intelligent security monitoring systems have been widely used in the fields of public safety, traffic management, and site monitoring. Traditional security monitoring systems mainly rely on a single optical camera for image acquisition and identify and detect target objects through fixed rule algorithms or simple image processing technology. This method has a certain effect in static and low-complexity environments, but when faced with complex and dynamic monitoring scenes, such as crowded areas, places with poor lighting or frequent occlusion, traditional monitoring systems often show problems such as inaccurate detection and insufficient real-time response, making it difficult to meet the growing demand for intelligent monitoring.

[0003] The existing monitoring systems have the following technical defects. First, a single optical camera is difficult to provide sufficiently clear and stable monitoring images in weak light, bad weather, occlusion, and complex environments. This camera method that relies on visible light not only performs poorly at night or in low light conditions, but is also easily limited by obstructions or monitoring angles, resulting in a significant reduction in monitoring effectiveness. At the same time, it is difficult for a single image data to capture the complete information of the target object, especially when monitoring dynamic scenes, it is difficult to accurately perceive the target's spatial position and behavioral changes.

[0004] Secondly, traditional monitoring systems rely on central servers for data processing and analysis. All data collected by front-end devices needs to be transmitted to the central server for processing. This centralized computing architecture brings significant bandwidth and computing resource burdens. On the one hand, the transmission of a large amount of high-resolution video data will occupy huge bandwidth resources, resulting in transmission delays, which in turn affects the response speed of real-time monitoring; on the other hand, the central server has a large computing pressure and it is difficult to complete complex image analysis and behavior recognition tasks in a short period of time, which can easily lead to response delays and miss the best time window for handling abnormal situations.

[0005] In addition, most of the existing abnormal behavior detection technologies rely on preset rules or thresholds and cannot be adaptively adjusted according to the complex situations in actual scenarios. This detection method based on fixed rules lacks flexibility and exhibits a high false alarm rate or missed alarm rate when facing a dynamically changing monitoring environment. For example, the system may not be able to distinguish between normal fast-moving behaviors (such as pedestrians rushing to catch a bus) and real abnormal behaviors (such as running and fleeing), resulting in false alarms; or in some complex scenarios, the system fails to identify potential threats in time, resulting in missed alarms. These problems are particularly evident in places with large traffic and frequent scene changes, such as shopping malls, stations, airports and other public places.

[0006] Existing anomaly detection methods also have shortcomings in the coordination of multiple cameras. Although multiple cameras have been used to expand coverage in some monitoring scenarios, the coordination between these cameras is mostly simple image stitching or region segmentation, lacking effective spatial perception and three-dimensional modeling capabilities. The lack of deep information fusion between cameras makes it impossible to provide accurate target positioning and tracking, especially when the monitoring target is partially obscured or appears at the junction of different cameras. It is difficult for the system to continuously and effectively track the target, resulting in poor monitoring results.

[0007] In addition, data privacy protection is also a major challenge for traditional monitoring systems. Existing monitoring systems usually lack effective privacy protection mechanisms when processing large amounts of video data, and there is a risk of data leakage and abuse. Especially in monitoring scenarios involving public places or private areas, data privacy issues are particularly important. As the global attention to personal privacy protection continues to increase, how to effectively protect data privacy while ensuring monitoring functions has become a key technical issue that smart security systems must face.

[0008] Therefore, how to provide an intelligent security monitoring method and system based on image recognition is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0009] One purpose of the present invention is to propose an intelligent security monitoring method and system based on image recognition. The present invention makes full use of multimodal sensor fusion, edge computing, deep learning and adaptive learning technology, and combines multiple sensors such as optical cameras, infrared sensors, depth cameras and radars to achieve accurate real-time monitoring in complex scenarios. By deploying lightweight neural network models on edge devices and applying Swin Transformer models for abnormal behavior analysis, combined with three-dimensional modeling and multi-camera collaboration, the accuracy of target tracking and the real-time response capability of the system are improved. At the same time, the system also has a differential privacy protection function to ensure data security, and has the advantages of high efficiency, flexibility, accuracy and privacy protection.

[0010] The intelligent security monitoring method based on image recognition according to an embodiment of the present invention comprises the following steps:

[0011] S1, collects visible light images through optical cameras, and obtains multimodal data through infrared sensors, depth cameras and radars;

[0012] S2, pre-process the acquired multi-modal data, and integrate the sensor data in real time through the Kalman filter algorithm to generate complete target information;

[0013] S3, deploy lightweight neural network models on edge devices to detect the generated complete target information in real time and perform preliminary processing on conventional scenarios;

[0014] S4, based on the improved visual SLAM technology, multiple cameras work together to build a three-dimensional scene model of the monitoring area and obtain three-dimensional information of multiple cameras;

[0015] S5. When the edge device detects a potential abnormal situation, it transmits the data to the central server, uses the SwinTransformer model to analyze the data for abnormal behavior, and combines the three-dimensional information of multiple cameras to optimize the detection results;

[0016] S6. Use the deep deterministic policy gradient algorithm to adaptively learn historical data, dynamically adjust the threshold of abnormal behavior detection, and optimize behavior recognition and response strategies;

[0017] S7. After detecting an abnormal situation, the data is processed through differential privacy protection and end-to-end encryption technology, while automatically controlling security equipment and issuing real-time alerts to managers.

[0018] Optionally, the S2 specifically includes:

[0019] S21, performing data synchronization and time alignment on the visible light image collected by the optical camera, the thermal imaging data of the infrared sensor, the depth image data of the depth camera, and the distance and speed information of the radar;

[0020] S22, performing denoising processing on the synchronized data, including applying a Gaussian filter algorithm to the visible light image to filter out high-frequency noise, applying a median filter algorithm to the depth image to remove isolated noise points, and applying a mean smoothing process to the infrared data to reduce random noise in the thermal imaging;

[0021] S23, performing standardization processing on the denoised data, mapping the data of different sensors to the same value range;

[0022] S24, applying the Kalman filter algorithm to the standardized data for fusion processing, and predicting the target state and covariance matrix at each time step t:

[0023]

[0024] in, represents the predicted state at time step t, A represents the state transfer matrix, x t-1 represents the state at time step t-1, B represents the control input matrix, u t-1 represents the control input at time step t-1;

[0025]

[0026] in, represents the prediction covariance matrix at time step t, P t-1 represents the covariance matrix at time step t-1, A T represents the transposed matrix of the state transfer matrix, Q represents the process noise covariance matrix;

[0027] S25, through the Kalman gain K t Update the state estimate:

[0028]

[0029] Among them, H represents the observation matrix, H T represents the transposed matrix of the observation matrix, and R represents the observation noise covariance matrix;

[0030] S26. Finally, the fused target state is calculated through the updated estimate value and the updated covariance of the Kalman filter:

[0031]

[0032] Among them, x t new represents the updated state estimate at time step t, z t Represents the observation value of the current time step;

[0033]

[0034] Among them, P t new represents the updated covariance matrix at time step t, and I represents the identity matrix.

[0035] Optionally, the S3 specifically includes:

[0036] S31, inputting the complete target information generated by step S2 into the lightweight neural network model in the edge device, wherein the lightweight neural network model is a pre-trained improved adaptive frequency domain decomposition model;

[0037] S32, perform spectrum transformation on the input image data, convert the spatiotemporal domain image into frequency domain representation, and use discrete Fourier transform to convert the image I(x,y) into frequency domain representation I(u,v):

[0038]

[0039] Among them, I(u,v) represents the frequency domain representation, M represents the width of the image, N represents the height of the image, u and v represent the frequency components, and j represents the imaginary unit;

[0040] S33, performing adaptive frequency selection on the converted frequency domain data, dynamically decomposing the frequency components and applying different weights to each frequency band, setting a trainable adaptive threshold λ f :

[0041]

[0042] Among them, W f (u,v) represents the adaptive frequency weight matrix, and Z represents the normalization factor;

[0043] S34, applying an adaptive scale interpolation algorithm to the filtered frequency domain feature data to dynamically adjust the feature scale of each frequency band:

[0044] F rescaled (u,v)=F(u,v)·γ(u,v);

[0045] Among them, F rescaled (u,v) represents the adjusted frequency domain features, F(u,v) represents the input frequency domain features, and γ(u,v) represents the frequency adaptive scaling factor;

[0046] S35, after the frequency domain features are adaptively screened and scaled, the features are returned to the spatiotemporal domain and processed by a sparse spatiotemporal convolutional network. The sparse spatiotemporal convolutional network dynamically adjusts the sparsity of the convolution kernel S k , controls the computational complexity of feature extraction:

[0047]

[0048] Among them, F i represents the input features, W k represents the convolution weight of the kth layer, θ represents the adaptive threshold, δ(·) represents the sparsity function, K represents the total number of convolutional layers, and ∥·∥1 represents the L1 norm;

[0049] When the feature sparsity meets the set threshold θ, some convolution calculations are skipped to reduce the computational burden of edge devices;

[0050] S36, the final output includes the category probability p of each targeti and the spatiotemporal bounding box (x i ,y i ,t i ,w i ,h i ), where t i Represents the boundary information in the time dimension, (x i ,y i ) represents the spatial coordinate, w i Indicates the width of the bounding box, h i Indicates the height of the bounding box.

[0051] Optionally, the S4 specifically includes:

[0052] S41, collecting two-dimensional image data in the monitoring area from different angles through multiple cameras i (x, y, t), where i represents the i-th camera, x and y represent pixel coordinates, and t represents the time frame;

[0053] S42, using the improved visual SLAM algorithm, extract feature points f in the view of each camera i (x,y,t), and based on the dynamic sparsity adaptive mechanism, the feature points are screened:

[0054]

[0055] Among them, S(f i ) represents the feature point f i The sparsity of n represents the total number of feature points in the current feature point set, ∥·∥1 represents the L1 norm, and W k Represents the weight matrix of the k-th layer feature, θ k represents the adaptive threshold;

[0056] S43, using the matching of feature points between multiple cameras, and using an adaptive multi-view geometric transformation model based on geometric optimization, the spatial correlation between feature points is calculated:

[0057]

[0058] Among them, T ij Represents the feature point f i and feature point f j The three-dimensional geometric transformation matrix between i ,f j ) represents the distance difference between two points, R ij represents the rotation matrix, t ij represents the translation vector, ∥·∥2 represents the L2 norm;

[0059] S44, based on the improved deep learning pose estimation network, using the multi-view feature point matching results, the pose T of all cameras is estimated through a global optimization algorithm i (t) Perform adaptive estimation:

[0060]

[0061] Among them, X v Represents the feature points in three-dimensional space, P i and P j represents the projection matrix of the camera, τ represents the regularization coefficient, m represents the total number of feature points, ∥P i X v -P j X v ∥ 2 represents the error after projection between camera i and camera j, A regularization constraint that represents the change in camera pose in adjacent time frames;

[0062] S45, by using the acquired camera position and feature point depth information, using the adaptive 3D reconstruction algorithm based on graph optimization, 3D modeling of the monitoring area is performed:

[0063]

[0064] Among them, Z(f i ) represents the feature point f i The three-dimensional depth information, B represents the camera baseline, f represents the focal length, d(f i ) represents parallax;

[0065] S46. Finally, the three-dimensional information of all cameras is combined and a complete three-dimensional scene model is output through global optimization. The three-dimensional scene model includes the spatial position, contour and structure information of the objects in the monitoring area.

[0066] Optionally, the S5 specifically includes:

[0067] S51, when the edge device detects potential abnormal behavior, transmits preliminary detection information of the target, including the two-dimensional coordinates, time frame, bounding box size, preliminary behavior classification label and motion trajectory of the target, to the central server;

[0068] S52. After receiving the data, the central server uses the Swin Transformer model to perform multi-scale feature extraction on the two-dimensional image and motion features. The Swin Transformer model applies an adaptive attention mechanism in each window, calculates local features, and merges the features of each window through weights;

[0069] S53, based on the features extracted by Swin Transformer, combined with time series information, conduct spatiotemporal analysis of abnormal behaviors, use the motion trajectory estimation model to analyze the motion changes of the target, and calculate the change in the three-dimensional spatial position of the target at time t and the previous time t-1:

[0070] ΔP t =∥P t -P t-1 ∥;

[0071] Among them, ΔP t represents the three-dimensional position change of the target between time t and time t-1, P t represents the three-dimensional position coordinates of the target at time t, P t-1 Represents the three-dimensional position coordinates of the target at time t-1;

[0072] S54. Optimize the target's motion trajectory using a three-dimensional behavior optimization model based on Bayesian inference, and estimate the target's position in three-dimensional space using the Bayesian inference model:

[0073]

[0074] Among them, P(P t |Z t ) represents the observed data Z at time t t Given the probability of the target's three-dimensional position, P(Z t |P t represents the observation probability of a given target position, P(P t |P t-1 ) represents the prior probability of the target in the time series, and C represents the normalization constant;

[0075] S55, combining the three-dimensional position information of multiple cameras, and optimizing the three-dimensional target trajectory by fusing the geometric information of multiple cameras:

[0076]

[0077] in, represents the 3D position of the fused target in the mth frame, N p Indicates the number of cameras involved in the fusion, R j represents the pose matrix of the jth camera, P m Represents the three-dimensional position of the feature point in the mth time frame;

[0078] S56. Based on the fused features, make a final judgment on the abnormal behavior:

[0079]

[0080] Among them, Fanomaly Represents the binary judgment of abnormal behavior, γ a represents the probability of abnormal behavior output by the Swin Transformer model, represents the 3D position of the fused target in the m-1th frame, and μ represents the preset abnormal behavior detection threshold.

[0081] Optionally, the S6 specifically includes:

[0082] S61, historical abnormal behavior data D h The current real-time monitoring data is input into the deep deterministic policy gradient algorithm. The deep deterministic policy gradient algorithm optimizes the abnormal behavior detection strategy through the actor-critic architecture. The actor network generates the optimal detection threshold adjustment strategy according to the current state.

[0083] S62, the system is in the current state S t Perform behavior evaluation and generate the optimal strategy A through the actor network t , used to adjust the abnormal behavior detection threshold η in real time. At the same time, the critic network evaluates the effect of the current strategy through historical data and gives feedback;

[0084] S63, through abnormal behavior feedback, the system continuously updates the policy network parameters and uses the current state S t and the current action A t The feedback information is used to optimize the adjustment strategy of abnormal behavior detection threshold;

[0085] S64. At each time step, the system dynamically adjusts the abnormal behavior detection threshold η based on the target's real-time motion trajectory and historical behavior pattern:

[0086]

[0087] Among them, η t+1 represents the anomaly detection threshold at the next time t+1, η t represents the detection threshold at the current moment, ρ represents the adaptive adjustment coefficient, ΔO t Indicates the target's motion change at the current moment, ΔO h represents the motion change of historical normal behavior, R t Represents the reward value of the current detection strategy, R max Indicates the maximum value of the reward;

[0088] S65. When abnormal behavior is detected, the system adjusts the threshold η according to the adaptive adjustment. t+1 Determine whether to trigger an alarm. If the target's motion change exceeds the threshold and the system's adaptive strategy assesses the probability of abnormal behavior to be high, an abnormal alarm is triggered and the preset countermeasures are executed.

[0089] The intelligent security monitoring system based on image recognition according to an embodiment of the present invention includes the following modules:

[0090] A data acquisition module, which is used to collect visible light images through an optical camera and obtain multimodal data through an infrared sensor, a depth camera, and a radar;

[0091] The data preprocessing module is used to preprocess the collected multimodal data and integrate the sensor data through the Kalman filter algorithm to generate complete target information in real time;

[0092] Edge computing module, used to deploy lightweight neural network models on edge devices to perform real-time detection and preliminary processing of generated target information;

[0093] The 3D modeling module is used to collect data from multiple cameras based on the improved visual SLAM technology and build a 3D scene model of the monitoring area;

[0094] The anomaly detection module is used to transmit relevant data to the central server when the edge device detects potential anomalies, and perform multi-scale feature extraction and abnormal behavior analysis on the data through the Swin Transformer model;

[0095] Adaptive learning module, which is used to adaptively learn historical data through a deep deterministic policy gradient algorithm and dynamically adjust the threshold for abnormal behavior detection;

[0096] The alarm and control module is used to trigger differential privacy protection and end-to-end encryption technology after detecting abnormal behavior, automatically control security equipment, and issue real-time alerts to managers.

[0097] The beneficial effects of the present invention are:

[0098] First of all, through the fusion of multiple sensors such as optical cameras, infrared sensors, depth cameras and radars, the system can not only continue to work under adverse conditions such as low light, bad weather or partial occlusion, but also can accurately identify and track the target objects in the monitoring area in real time through the integration of multimodal data. The real-time fusion of multimodal sensor data by the Kalman filter algorithm enables the system to generate more accurate target information and avoid the limitations of single sensor data in different environments. This greatly improves the adaptability and monitoring effect of the system, especially at night, under complex lighting conditions, or in scenes with limited field of view, and can provide more reliable monitoring data.

[0099] Secondly, the introduction of edge computing technology has significantly improved the real-time response capability of the system. Deploying lightweight neural network models on edge devices allows some data processing tasks to be completed in local devices, greatly reducing the computing burden of the central server and reducing the bandwidth occupied by data transmission. Edge devices can directly complete data processing when no abnormalities are detected in normal scenarios, and only transmit data to the central server for in-depth processing when further analysis is required. This not only improves the processing speed of the system and ensures real-time performance, but also optimizes the resource allocation of the entire system, allowing each monitoring node in a large-scale monitoring system to operate efficiently and collaboratively.

[0100] Through improved visual SLAM technology and 3D modeling, the system realizes the collaborative work of multiple cameras and generates high-precision 3D scene models in wide-area complex environments. The system can obtain 3D information from multiple cameras in real time and reconstruct the monitored area in 3D through visual SLAM technology. The fusion of multi-view information enables the system to accurately track the position and movement trajectory of the target. Even when the target is partially obscured or appears in the boundary area of ​​multiple cameras, it can maintain a stable tracking effect through data complementarity between cameras. This multi-camera collaboration and 3D modeling greatly improves the accuracy of target positioning, especially for large-scale, high-complexity scenes such as airports and shopping malls, further improving the practicality of the system.

[0101] In terms of abnormal behavior detection, the Swin Transformer model combines multi-scale feature extraction and time series analysis, enabling the system to analyze and judge complex behavior patterns more accurately. At the same time, the application of adaptive learning mechanisms enables the system to self-optimize based on real-time monitoring data and historical data, dynamically adjust the threshold for abnormal behavior detection, and greatly reduce the occurrence of false positives and false negatives. The deep deterministic policy gradient algorithm automatically adjusts the response strategy by evaluating changes in the monitoring scene in real time, ensuring that the system can respond flexibly to diverse behavior patterns. This feature is particularly suitable for scenes with dense crowds or high dynamic changes, such as stations and subways, and can effectively respond to various complex behavior patterns in the environment.

[0102] Finally, the system also combines differential privacy protection and end-to-end encryption technology to ensure that the privacy of users is effectively protected during the collection, processing, transmission and storage of monitoring data. Through privacy protection technology, the system can prevent the leakage or abuse of sensitive data while ensuring the normal operation of monitoring functions. In the current social context of increasingly stringent requirements for data privacy, this provides an intelligent security monitoring solution that meets security and compliance standards, greatly improving the security and social acceptance of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0104] Figure 1 This is a flow chart of the intelligent security monitoring method based on image recognition proposed by the present invention;

[0105] Figure 2 This is a structural schematic diagram of the intelligent security monitoring system based on image recognition proposed by the present invention. DETAILED DESCRIPTION

[0106] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0107] refer to Figure 1 , an intelligent security monitoring method based on image recognition, comprising the following steps:

[0108] S1, collects visible light images through optical cameras, and obtains multimodal data through infrared sensors, depth cameras and radars;

[0109] S2, pre-process the acquired multi-modal data, and integrate the sensor data in real time through the Kalman filter algorithm to generate complete target information;

[0110] S3, deploy lightweight neural network models on edge devices to detect the generated complete target information in real time and perform preliminary processing on conventional scenarios;

[0111] S4, based on the improved visual SLAM technology, multiple cameras work together to build a three-dimensional scene model of the monitoring area and obtain three-dimensional information of multiple cameras;

[0112] S5. When the edge device detects a potential abnormal situation, it transmits the data to the central server, uses the SwinTransformer model to analyze the data for abnormal behavior, and combines the three-dimensional information of multiple cameras to optimize the detection results;

[0113] S6. Use the deep deterministic policy gradient algorithm to adaptively learn historical data, dynamically adjust the threshold of abnormal behavior detection, and optimize behavior recognition and response strategies;

[0114] S7. After detecting an abnormal situation, the data is processed through differential privacy protection and end-to-end encryption technology, while automatically controlling security equipment and issuing real-time alerts to managers.

[0115] In this implementation, S2 specifically includes:

[0116] S21, performing data synchronization and time alignment on the visible light image collected by the optical camera, the thermal imaging data of the infrared sensor, the depth image data of the depth camera, and the distance and speed information of the radar;

[0117] S22, performing denoising processing on the synchronized data, including applying a Gaussian filter algorithm to the visible light image to filter out high-frequency noise, applying a median filter algorithm to the depth image to remove isolated noise points, and applying a mean smoothing process to the infrared data to reduce random noise in the thermal imaging;

[0118] S23, performing standardization processing on the denoised data, mapping the data of different sensors to the same value range;

[0119] S24, applying the Kalman filter algorithm to the standardized data for fusion processing, and predicting the target state and covariance matrix at each time step t:

[0120]

[0121] in, represents the predicted state at time step t, A represents the state transfer matrix, x t-1 represents the state at time step t-1, B represents the control input matrix, u t-1 represents the control input at time step t-1;

[0122]

[0123] in, represents the prediction covariance matrix at time step t, P t-1 represents the covariance matrix at time step t-1, A T represents the transposed matrix of the state transfer matrix, Q represents the process noise covariance matrix;

[0124] S25, through the Kalman gain K t Update the state estimate:

[0125]

[0126] Among them, H represents the observation matrix, H T represents the transposed matrix of the observation matrix, and R represents the observation noise covariance matrix;

[0127] S26. Finally, the fused target state is calculated through the updated estimate value and the updated covariance of the Kalman filter:

[0128]

[0129] Among them, x tnew represents the updated state estimate at time step t, z t Represents the observation value of the current time step;

[0130]

[0131] Among them, P t new represents the updated covariance matrix at time step t, and I represents the identity matrix.

[0132] In this implementation, S3 specifically includes:

[0133] S31, inputting the complete target information generated by step S2 into the lightweight neural network model in the edge device, wherein the lightweight neural network model is a pre-trained improved adaptive frequency domain decomposition model;

[0134] S32, perform spectrum transformation on the input image data, convert the spatiotemporal domain image into frequency domain representation, and use discrete Fourier transform to convert the image I(x,y) into frequency domain representation I(u,v):

[0135]

[0136] Among them, I(u,v) represents the frequency domain representation, M represents the width of the image, N represents the height of the image, u and v represent the frequency components, and j represents the imaginary unit;

[0137] S33, performing adaptive frequency selection on the converted frequency domain data, dynamically decomposing the frequency components and applying different weights to each frequency band, setting a trainable adaptive threshold λ f :

[0138]

[0139] Among them, W f (u,v) represents the adaptive frequency weight matrix, and Z represents the normalization factor;

[0140] S34, applying an adaptive scale interpolation algorithm to the filtered frequency domain feature data to dynamically adjust the feature scale of each frequency band:

[0141] F rescaled (u,v)=F(u,v)·γ(u,v);

[0142] Among them, F rescaled (u,v) represents the adjusted frequency domain features, F(u,v) represents the input frequency domain features, and γ(u,v) represents the frequency adaptive scaling factor;

[0143] S35, after the frequency domain features are adaptively screened and scaled, the features are returned to the spatiotemporal domain and processed by a sparse spatiotemporal convolutional network. The sparse spatiotemporal convolutional network dynamically adjusts the sparsity of the convolution kernel S k , controls the computational complexity of feature extraction:

[0144]

[0145] Among them, F i represents the input features, W k represents the convolution weight of the kth layer, θ represents the adaptive threshold, δ(·) represents the sparsity function, K represents the total number of convolutional layers, and ∥·∥1 represents the L1 norm;

[0146] When the feature sparsity meets the set threshold θ, some convolution calculations are skipped to reduce the computational burden of edge devices;

[0147] S36, the final output includes the category probability p of each target i and the spatiotemporal bounding box (x i ,y i ,t i ,w i ,h i ), where t i Represents the boundary information in the time dimension, (x i ,y i ) represents the spatial coordinate, w i Indicates the width of the bounding box, h i Indicates the height of the bounding box.

[0148] In this implementation manner, the S4 specifically includes:

[0149] S41, collecting two-dimensional image data in the monitoring area from different angles through multiple cameras i (x, y, t), where i represents the i-th camera, x and y represent pixel coordinates, and t represents the time frame;

[0150] S42, using the improved visual SLAM algorithm, extract feature points f in the view of each camera i (x,y,t), and based on the dynamic sparsity adaptive mechanism, the feature points are screened:

[0151]

[0152] Among them, S(f i ) represents the feature point f i The sparsity of n represents the total number of feature points in the current feature point set, ∥·∥1 represents the L1 norm, and W k Represents the weight matrix of the k-th layer feature, θ krepresents the adaptive threshold;

[0153] S43, using the matching of feature points between multiple cameras, and using an adaptive multi-view geometric transformation model based on geometric optimization, the spatial correlation between feature points is calculated:

[0154]

[0155] Among them, T ij Represents the feature point f i and feature point f j The three-dimensional geometric transformation matrix between i ,f j ) represents the distance difference between two points, R ij represents the rotation matrix, t ij represents the translation vector, ∥·∥2 represents the L2 norm;

[0156] S44, based on the improved deep learning pose estimation network, using the multi-view feature point matching results, the pose T of all cameras is estimated through a global optimization algorithm i (t) Perform adaptive estimation:

[0157]

[0158] Among them, X v Represents the feature points in three-dimensional space, P i and P j represents the projection matrix of the camera, τ represents the regularization coefficient, m represents the total number of feature points, ∥P i X v -P j X v ∥ 2 represents the error after projection between camera i and camera j, A regularization constraint that represents the change in camera pose in adjacent time frames;

[0159] S45, by using the acquired camera position and feature point depth information, using the adaptive 3D reconstruction algorithm based on graph optimization, 3D modeling of the monitoring area is performed:

[0160]

[0161] Among them, Z(f i ) represents the feature point f i The three-dimensional depth information, B represents the camera baseline, f represents the focal length, d(f i ) represents parallax;

[0162] S46. Finally, the three-dimensional information of all cameras is combined and a complete three-dimensional scene model is output through global optimization. The three-dimensional scene model includes the spatial position, contour and structure information of the objects in the monitoring area.

[0163] In this implementation manner, S5 specifically includes:

[0164] S51, when the edge device detects potential abnormal behavior, transmits preliminary detection information of the target, including the two-dimensional coordinates, time frame, bounding box size, preliminary behavior classification label and motion trajectory of the target, to the central server;

[0165] S52. After receiving the data, the central server uses the Swin Transformer model to perform multi-scale feature extraction on the two-dimensional image and motion features. The Swin Transformer model applies an adaptive attention mechanism in each window, calculates local features, and merges the features of each window through weights;

[0166] S53, based on the features extracted by Swin Transformer, combined with time series information, conduct spatiotemporal analysis of abnormal behaviors, use the motion trajectory estimation model to analyze the motion changes of the target, and calculate the change in the three-dimensional spatial position of the target at time t and the previous time t-1:

[0167] ΔP t =∥P t -P t-1 ∥;

[0168] Among them, ΔP t represents the three-dimensional position change of the target between time t and time t-1, P t represents the three-dimensional position coordinates of the target at time t, P t-1 Represents the three-dimensional position coordinates of the target at time t-1;

[0169] S54. Optimize the target's motion trajectory using a three-dimensional behavior optimization model based on Bayesian inference, and estimate the target's position in three-dimensional space using the Bayesian inference model:

[0170]

[0171] Among them, P(P t |Z t ) represents the observed data Z at time t t Given the probability of the target's three-dimensional position, P(Z t |P t represents the observation probability of a given target position, P(P t |P t-1) represents the prior probability of the target in the time series, and C represents the normalization constant;

[0172] S55, combining the three-dimensional position information of multiple cameras, and optimizing the three-dimensional target trajectory by fusing the geometric information of multiple cameras:

[0173]

[0174] in, represents the 3D position of the fused target in the mth frame, N p Indicates the number of cameras involved in the fusion, R j represents the pose matrix of the jth camera, P m Represents the three-dimensional position of the feature point in the mth time frame;

[0175] S56. Based on the fused features, make a final judgment on the abnormal behavior:

[0176]

[0177] Among them, F anomaly Represents the binary judgment of abnormal behavior, γ a represents the probability of abnormal behavior output by the Swin Transformer model, represents the 3D position of the fused target in the m-1th frame, and μ represents the preset abnormal behavior detection threshold.

[0178] In this implementation manner, S6 specifically includes:

[0179] S61, historical abnormal behavior data D h The current real-time monitoring data is input into the deep deterministic policy gradient algorithm. The deep deterministic policy gradient algorithm optimizes the abnormal behavior detection strategy through the actor-critic architecture. The actor network generates the optimal detection threshold adjustment strategy according to the current state.

[0180] S62, the system is in the current state S t Perform behavior evaluation and generate the optimal strategy A through the actor network t , used to adjust the abnormal behavior detection threshold η in real time. At the same time, the critic network evaluates the effect of the current strategy through historical data and gives feedback;

[0181] S63, through abnormal behavior feedback, the system continuously updates the policy network parameters and uses the current state S t and the current action A t The feedback information is used to optimize the adjustment strategy of abnormal behavior detection threshold;

[0182] S64. At each time step, the system dynamically adjusts the abnormal behavior detection threshold η based on the target's real-time motion trajectory and historical behavior pattern:

[0183]

[0184] Among them, η t+1 represents the anomaly detection threshold at the next time t+1, η t represents the detection threshold at the current moment, ρ represents the adaptive adjustment coefficient, ΔO t Indicates the target's motion change at the current moment, ΔO h represents the motion change of historical normal behavior, R t Represents the reward value of the current detection strategy, R max Indicates the maximum value of the reward;

[0185] S65. When abnormal behavior is detected, the system adjusts the threshold η according to the adaptive adjustment. t+1 Determine whether to trigger an alarm. If the target's motion change exceeds the threshold and the system's adaptive strategy assesses the probability of abnormal behavior to be high, an abnormal alarm is triggered and the preset countermeasures are executed.

[0186] refer to Figure 2 , the intelligent security monitoring system based on image recognition includes the following modules:

[0187] A data acquisition module, which is used to collect visible light images through an optical camera and obtain multimodal data through an infrared sensor, a depth camera, and a radar;

[0188] The data preprocessing module is used to preprocess the collected multimodal data and integrate the sensor data through the Kalman filter algorithm to generate complete target information in real time;

[0189] Edge computing module, used to deploy lightweight neural network models on edge devices to perform real-time detection and preliminary processing of generated target information;

[0190] The 3D modeling module is used to collect data from multiple cameras based on the improved visual SLAM technology and build a 3D scene model of the monitoring area;

[0191] The anomaly detection module is used to transmit relevant data to the central server when the edge device detects potential anomalies, and perform multi-scale feature extraction and abnormal behavior analysis on the data through the Swin Transformer model;

[0192] Adaptive learning module, which is used to adaptively learn historical data through a deep deterministic policy gradient algorithm and dynamically adjust the threshold for abnormal behavior detection;

[0193] The alarm and control module is used to trigger differential privacy protection and end-to-end encryption technology after detecting abnormal behavior, automatically control security equipment, and issue real-time alerts to managers.

[0194] Embodiment 1:

[0195] In order to verify the feasibility of the present invention in implementation, the present invention was applied to a public transportation hub in a city center for a three-month application test. The average daily passenger flow of this transportation hub exceeds 500,000 people, and the flow of people is complex and the scenes are diverse. Especially during peak hours in the morning and evening, congestion and chaos are more common. Traditional monitoring systems are difficult to achieve efficient abnormal behavior detection and real-time response due to problems such as insufficient light, many monitoring blind spots, and insufficient equipment computing power. In this context, the intelligent security monitoring system of the present invention was used to conduct a comprehensive monitoring coverage test on the area, with the aim of verifying the system's target detection, abnormal behavior analysis and real-time response capabilities in a complex environment.

[0196] The system first deployed multimodal sensor acquisition equipment, including optical cameras, infrared sensors, depth cameras and radars, which are distributed at various entrances and exits of the hub, platforms, ticket areas and areas with dense pedestrian traffic. Optical cameras are used to collect conventional visible light images, infrared sensors are responsible for thermal imaging at night and in poor light environments, depth cameras obtain depth information of target objects, and radars are used to obtain distance and speed information of target objects. During peak hours, due to the dense crowds, it is difficult for a single camera to monitor every detail. Through the fusion processing of multimodal sensors, the system can make up for visual blind spots and continue to monitor efficiently in low light and blocked environments.

[0197] The system's preprocessing module denoises and standardizes the data collected by the multimodal sensor, and uses the Kalman filter algorithm to predict and fuse the state of the target. After three months of testing, the system's target detection accuracy reached 95% in low light or complex occlusion conditions, an increase of 20 percentage points compared to the 75% accuracy of a traditional single optical camera. Especially in low light conditions at night, the combination of infrared sensors and depth cameras effectively compensates for the shortcomings of optical cameras, allowing the system to accurately identify targets in these environments.

[0198] In the edge computing part, the system deploys a lightweight neural network model for real-time target detection and preliminary analysis. In actual tests, the system can process more than 25 frames of image data per second on the edge device, with an average delay of only 150 milliseconds, which is a 50% reduction in delay compared to traditional centralized processing solutions. In this way, the system can complete processing directly on the edge device when normal behavior is detected, avoiding unnecessary data transmission. Only when potential abnormal behavior is detected will the system transmit relevant data to the central server for more in-depth analysis. This processing mode effectively reduces the burden on the central server. Under high-traffic data transmission during peak hours, the processing delay of the central server is reduced from 800 milliseconds of the traditional system to 400 milliseconds, and the real-time response capability of the system is significantly improved.

[0199] Through improved visual SLAM technology, the system has achieved three-dimensional scene modeling in the monitoring area inside the transportation hub. The multiple cameras deployed by the system work together to generate a three-dimensional scene model of the monitoring area, covering the key areas of the entire transportation hub. Especially in complex scenes such as platforms and entrances and exits, the system can effectively solve the occlusion problem. In the test, the system can accurately track the target through three-dimensional modeling. Even if the target is partially blocked or leaves the field of view of a camera, the system can still use the three-dimensional information of other cameras for continuous tracking. Compared with traditional two-dimensional plane monitoring, the system's target tracking accuracy has increased by 15%.

[0200] In terms of abnormal behavior detection, the system uses the Swin Transformer model to accurately analyze the behavior patterns of people through multi-scale feature extraction and spatiotemporal sequence analysis. During peak hours in the morning and evening, when the system detects a sudden gathering of people or unusual rapid movement of individuals, the system can quickly determine the abnormal situation and dynamically adjust the threshold through an adaptive learning mechanism. During the test period, the system detected more than 500 abnormal behaviors, most of which were crowds or pedestrians running fast. Through adaptive learning, the system gradually reduced the number of false alarms, and eventually controlled the false alarm rate to less than 2%, while the false alarm rate of traditional systems was as high as 8%. It is worth noting that in an emergency, the system successfully detected a suspicious person who stayed on the platform for a long time and suddenly ran after many people left the platform. The system promptly issued an alarm and notified security personnel to take action.

[0201] Data privacy protection is also one of the core advantages of this system. Throughout the testing process, the system ensured the security of all monitoring data through differential privacy protection and end-to-end encryption technology. Since the transportation hub is a public place, involving a large number of people's images and behavioral data, privacy issues are particularly important. The system is able to encrypt when processing and storing data to ensure that no personal information will be leaked or abused. During these three months, the system did not experience any data leakage incidents, which fully complied with local privacy protection policy requirements.

[0202] Table 1 Performance comparison between traditional monitoring system and intelligent security monitoring system of the present invention

[0203]

[0204] Table 1 provides a detailed performance analysis of the traditional monitoring system and the intelligent security monitoring system of the present invention. In terms of target detection accuracy, the traditional monitoring system is only 75%, while the intelligent monitoring system of the present invention significantly improves the detection accuracy to 95% through multimodal sensor fusion and deep learning technology. Especially in night environments, the accuracy of traditional monitoring systems is only 60%, which is limited by light conditions; while the present invention achieves a nighttime detection accuracy of 90% through the combination of infrared sensors and depth cameras, effectively solving the shortcomings of traditional monitoring systems in low light conditions.

[0205] In low light or occlusion situations, the detection rate of the traditional system drops to 50%, while the detection accuracy of the system of the present invention is increased to 80% through multimodal data fusion and Kalman filtering technology. In terms of data processing delay, the traditional system has a processing delay of up to 300 milliseconds due to the heavy load of the central server; while the system of the present invention reduces the delay to 150 milliseconds through the introduction of edge computing technology, improving the real-time response capability of the system.

[0206] In terms of false alarm rate, the traditional monitoring system has a false alarm rate of up to 8% due to the lack of adaptive learning mechanism. The present invention effectively reduces the false alarm rate to 2% through adaptive learning and deep deterministic policy gradient algorithm, greatly reducing the possibility of false alarms and missed alarms. In terms of target tracking accuracy, the traditional system is 85%, while the system of the present invention achieves 100% accurate tracking through 3D modeling and multi-camera collaboration, especially in occluded or complex environments.

[0207] In addition, during the peak period of data transmission delay, the traditional system has a delay of 800 milliseconds, while the intelligent security system of the present invention reduces it to 400 milliseconds, further improving the response speed and efficiency. Most importantly, in terms of data privacy protection, the traditional monitoring system has experienced two data leakage incidents, while the present invention uses differential privacy protection technology and end-to-end encryption technology to ensure data security, and no data leakage incidents occurred during the test.

[0208] Overall, after three months of testing, the intelligent security monitoring system of the present invention has demonstrated excellent performance in a complex transportation hub environment. The system can not only achieve accurate target recognition and tracking in low light and complex environments, but also greatly improve the real-time and responsiveness of data processing through efficient collaboration between edge computing and central servers. In addition, the accuracy of abnormal behavior detection and the privacy protection capabilities of the system have also been fully verified. Through this embodiment, it is proved that the present invention can effectively solve the shortcomings of traditional monitoring systems in dynamic and complex environments, and has significant application value.

[0209] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An intelligent security monitoring method based on image recognition, characterized in that: The steps include: S1, collects visible light images through optical cameras, and obtains multimodal data through infrared sensors, depth cameras and radars; S2, pre-process the acquired multi-modal data, and integrate the sensor data in real time through the Kalman filter algorithm to generate complete target information; S3, deploy lightweight neural network models on edge devices to detect the generated complete target information in real time and perform preliminary processing on conventional scenarios; S4, based on the improved visual SLAM technology, multiple cameras work together to build a three-dimensional scene model of the monitoring area and obtain three-dimensional information of multiple cameras; S5. When the edge device detects a potential abnormal situation, it transmits the data to the central server, uses the SwinTransformer model to analyze the data for abnormal behavior, and combines the three-dimensional information of multiple cameras to optimize the detection results; S6. Use the deep deterministic policy gradient algorithm to adaptively learn historical data, dynamically adjust the threshold of abnormal behavior detection, and optimize behavior recognition and response strategies; S7. After detecting an abnormal situation, the data is processed through differential privacy protection and end-to-end encryption technology, while automatically controlling security equipment and issuing real-time alerts to managers.

2. The intelligent security monitoring method based on image recognition according to claim 1 is characterized in that: The S2 specifically includes: S21, performing data synchronization and time alignment on the visible light image collected by the optical camera, the thermal imaging data of the infrared sensor, the depth image data of the depth camera, and the distance and speed information of the radar; S22, performing denoising processing on the synchronized data, including applying a Gaussian filter algorithm to the visible light image to filter out high-frequency noise, applying a median filter algorithm to the depth image to remove isolated noise points, and applying a mean smoothing process to the infrared data to reduce random noise in the thermal imaging; S23, performing standardization processing on the denoised data, mapping the data of different sensors to the same value range; S24, applying the Kalman filter algorithm to the standardized data for fusion processing, and predicting the target state and covariance matrix at each time step t: in, represents the predicted state at time step t, A represents the state transfer matrix, x t-1 represents the state at time step t-1, B represents the control input matrix, u t-1 represents the control input at time step t-1; in, represents the prediction covariance matrix at time step t, P t-1 represents the covariance matrix at time step t-1, A T represents the transposed matrix of the state transfer matrix, Q represents the process noise covariance matrix; S25, through the Kalman gain K t Update the state estimate: Among them, H represents the observation matrix, H T represents the transposed matrix of the observation matrix, and R represents the observation noise covariance matrix; S26. Finally, the fused target state is calculated through the updated estimate value and the updated covariance of the Kalman filter: Among them, x t new represents the updated state estimate at time step t, z t Represents the observation value of the current time step; Among them, P t new represents the updated covariance matrix at time step t, and I represents the identity matrix.

3. The intelligent security monitoring method based on image recognition according to claim 1 is characterized in that: The S3 specifically includes: S31, inputting the complete target information generated by step S2 into the lightweight neural network model in the edge device, wherein the lightweight neural network model is a pre-trained improved adaptive frequency domain decomposition model; S32, perform spectrum transformation on the input image data, convert the spatiotemporal domain image into frequency domain representation, and use discrete Fourier transform to convert the image I(x,y) into frequency domain representation I(u,v): Among them, I(u,v) represents the frequency domain representation, M represents the width of the image, N represents the height of the image, u and v represent the frequency components, and j represents the imaginary unit; S33, performing adaptive frequency selection on the converted frequency domain data, dynamically decomposing the frequency components and applying different weights to each frequency band, setting a trainable adaptive threshold λ f : Among them, W f (u,v) represents the adaptive frequency weight matrix, and Z represents the normalization factor; S34, applying an adaptive scale interpolation algorithm to the filtered frequency domain feature data to dynamically adjust the feature scale of each frequency band: F rescaled (u,v)=F(u,v)·γ(u,v); Among them, F rescaled (u,v) represents the adjusted frequency domain features, F(u,v) represents the input frequency domain features, and γ(u,v) represents the frequency adaptive scaling factor; S35, after the frequency domain features are adaptively screened and scaled, the features are returned to the spatiotemporal domain and processed by a sparse spatiotemporal convolutional network. The sparse spatiotemporal convolutional network dynamically adjusts the sparsity of the convolution kernel S k , controls the computational complexity of feature extraction: Among them, F i represents the input features, W k represents the convolution weight of the kth layer, θ represents the adaptive threshold, δ(·) represents the sparsity function, K represents the total number of convolutional layers, and ∥·∥1 represents the L1 norm; When the feature sparsity meets the set threshold θ, some convolution calculations are skipped to reduce the computational burden of edge devices; S36, the final output includes the category probability p of each target i and the spatiotemporal bounding box (x i ,y i ,t i ,w i ,h i ), where t i Represents the boundary information in the time dimension, (x i ,y i ) represents the spatial coordinate, w i Indicates the width of the bounding box, h i Indicates the height of the bounding box.

4. The intelligent security monitoring method based on image recognition according to claim 1 is characterized in that: The S4 specifically includes: S41, collecting two-dimensional image data in the monitoring area from different angles through multiple cameras i (x, y, t), where i represents the i-th camera, x and y represent pixel coordinates, and t represents the time frame; S42, using the improved visual SLAM algorithm, extract feature points f in the view of each camera i (x,y,t), and based on the dynamic sparsity adaptive mechanism, the feature points are screened: Among them, S(f i ) represents the feature point f i The sparsity of n represents the total number of feature points in the current feature point set, ∥·∥1 represents the L1 norm, and W k Represents the weight matrix of the k-th layer feature, θ k represents the adaptive threshold; S43, using the matching of feature points between multiple cameras, and using an adaptive multi-view geometric transformation model based on geometric optimization, the spatial correlation between feature points is calculated: Among them, T ij Represents the feature point f i and feature point f j The three-dimensional geometric transformation matrix between i ,f j ) represents the distance difference between two points, R ij represents the rotation matrix, t ij represents the translation vector, ∥·∥2 represents the L2 norm; S44, based on the improved deep learning pose estimation network, using the multi-view feature point matching results, the pose T of all cameras is estimated through a global optimization algorithm i (t) Perform adaptive estimation: Among them, X v Represents the feature points in three-dimensional space, P i and P j represents the projection matrix of the camera, τ represents the regularization coefficient, m represents the total number of feature points, ∥P i X v -P j X v ∥ 2 represents the error after projection between camera i and camera j, A regularization constraint that represents the change in camera pose in adjacent time frames; S45, by using the acquired camera position and feature point depth information, using the adaptive 3D reconstruction algorithm based on graph optimization, 3D modeling of the monitoring area is performed: Among them, Z(f i ) represents the feature point f i The three-dimensional depth information, B represents the camera baseline, f represents the focal length, d(f i ) represents parallax; S46. Finally, the three-dimensional information of all cameras is combined and a complete three-dimensional scene model is output through global optimization. The three-dimensional scene model includes the spatial position, contour and structure information of the objects in the monitoring area.

5. The intelligent security monitoring method based on image recognition according to claim 1 is characterized in that: The S5 specifically includes: S51, when the edge device detects potential abnormal behavior, transmits preliminary detection information of the target, including the two-dimensional coordinates, time frame, bounding box size, preliminary behavior classification label and motion trajectory of the target, to the central server; S52. After receiving the data, the central server uses the Swin Transformer model to perform multi-scale feature extraction on the two-dimensional image and motion features. The Swin Transformer model applies an adaptive attention mechanism in each window, calculates local features, and merges the features of each window through weights; S53, based on the features extracted by Swin Transformer, combined with time series information, conduct spatiotemporal analysis of abnormal behaviors, use the motion trajectory estimation model to analyze the motion changes of the target, and calculate the change in the three-dimensional spatial position of the target at time t and the previous time t-1: ΔP t =∥P t -P t-1 ∥; Among them, ΔP t represents the three-dimensional position change of the target between time t and time t-1, P t represents the three-dimensional position coordinates of the target at time t, P t-1 Represents the three-dimensional position coordinates of the target at time t-1; S54. Optimize the target's motion trajectory using a three-dimensional behavior optimization model based on Bayesian inference, and estimate the target's position in three-dimensional space using the Bayesian inference model: Among them, P(P t |Z t ) represents the observed data Z at time t t Given the probability of the target's three-dimensional position, P(Z t |P t ) represents the observation probability of a given target position, P(P t |P t-1 ) represents the prior probability of the target in the time series, and C represents the normalization constant; S55, combining the three-dimensional position information of multiple cameras, and optimizing the three-dimensional target trajectory by fusing the geometric information of multiple cameras: in, represents the 3D position of the fused target in the mth frame, N p Indicates the number of cameras involved in fusion, R j represents the pose matrix of the jth camera, P m Represents the three-dimensional position of the feature point in the mth time frame; S56. Based on the fused features, make a final judgment on the abnormal behavior: Among them, F anomaly Indicates the binary judgment of abnormal behavior, γ a represents the probability of abnormal behavior output by the Swin Transformer model, represents the 3D position of the fused target in the m-1th frame, and μ represents the preset abnormal behavior detection threshold.

6. The intelligent security monitoring method based on image recognition according to claim 1 is characterized in that: The S6 specifically includes: S61, historical abnormal behavior data D h The current real-time monitoring data is input into the deep deterministic policy gradient algorithm. The deep deterministic policy gradient algorithm optimizes the abnormal behavior detection strategy through the actor-critic architecture. The actor network generates the optimal detection threshold adjustment strategy according to the current state. S62, the system is in the current state S t Perform behavior evaluation and generate the optimal strategy A through the actor network t , used to adjust the abnormal behavior detection threshold η in real time. At the same time, the critic network evaluates the effect of the current strategy through historical data and gives feedback; S63, through abnormal behavior feedback, the system continuously updates the policy network parameters and uses the current state S t and the current action A t The feedback information is used to optimize the adjustment strategy of abnormal behavior detection threshold; S64. At each time step, the system dynamically adjusts the abnormal behavior detection threshold η based on the target's real-time motion trajectory and historical behavior pattern: Among them, η t+1 represents the anomaly detection threshold at the next time t+1, η t represents the detection threshold at the current moment, ρ represents the adaptive adjustment coefficient, ΔO t Indicates the target's motion change at the current moment, ΔO h represents the motion change of historical normal behavior, R t Represents the reward value of the current detection strategy, R max Indicates the maximum value of the reward; S65. When abnormal behavior is detected, the system adjusts the threshold η according to the adaptive adjustment. t+1 Determine whether to trigger an alarm. If the target's motion change exceeds the threshold and the system's adaptive strategy assesses the probability of abnormal behavior to be high, an abnormal alarm is triggered and the preset countermeasures are executed.

7. An intelligent security monitoring system based on image recognition, executing the intelligent security monitoring method based on image recognition according to any one of claims 1 to 6, characterized in that: Includes the following modules: A data acquisition module, which is used to collect visible light images through an optical camera and obtain multimodal data through an infrared sensor, a depth camera, and a radar; The data preprocessing module is used to preprocess the collected multimodal data and integrate the sensor data through the Kalman filter algorithm to generate complete target information in real time; Edge computing module, used to deploy lightweight neural network models on edge devices to perform real-time detection and preliminary processing of generated target information; The 3D modeling module is used to collect data from multiple cameras based on the improved visual SLAM technology and build a 3D scene model of the monitoring area; The anomaly detection module is used to transmit relevant data to the central server when the edge device detects potential anomalies, and perform multi-scale feature extraction and abnormal behavior analysis on the data through the Swin Transformer model; Adaptive learning module, which is used to adaptively learn historical data through a deep deterministic policy gradient algorithm and dynamically adjust the threshold for abnormal behavior detection; The alarm and control module is used to trigger differential privacy protection and end-to-end encryption technology after detecting abnormal behavior, automatically control security equipment, and issue real-time alerts to managers.

Images