Intelligent security linkage method and system integrated with face recognition

Abstract

The invention discloses an intelligent security linkage method and system integrated with face recognition, and belongs to the technical field of intelligent security, and the method comprises the steps: carrying out the illumination compensation and angle correction of a face image collected in a monitoring region, and outputting an enhanced face feature map; inputting the enhanced image into a face recognition model, performing identity recognition in combination with a historical feature vector library, and outputting a face identification vector; based on the vector, extracting an access track and an access control event of the target object, and constructing a linkage trigger feature set of path offset rate, abnormal access time density, authority level deviation and the like; matching is carried out according to the risk rule set, and a security early warning instruction including a linkage level, a time window and a target equipment ID is generated; the instruction is sent to the access controller or the broadcast terminal and a linkage action is executed; according to the method, intelligent identification, dynamic early warning and multi-device linkage control of suspected risk objects can be realized, and the method has the advantages of high identification precision, timely response, adaptive updating of the model and the like.

Description

Technical Field

[0001] This invention relates to the field of intelligent security technology, specifically to an intelligent security linkage method and system integrating facial recognition. Background Technology

[0002] While existing security systems widely utilize facial recognition technology to enhance access control intelligence, significant shortcomings remain in complex real-world environments. Firstly, factors such as varying lighting conditions, poor shooting angles, and occlusion interference result in low-quality facial images, leading to decreased recognition accuracy, particularly in key locations where this can cause omissions or misjudgments of critical targets. Secondly, current systems primarily rely on single-point static recognition, lacking fusion analysis with dynamic information such as historical personnel behavior and access control events, making it difficult to effectively assess potential risk behaviors. Furthermore, existing recognition results are typically used only for alarm alerts, lacking a linkage mechanism with backend security subsystems. This prevents real-time response and handling after recognition, especially when dealing with suspected high-risk individuals, as they cannot automatically trigger access control or broadcast alerts based on behavioral patterns, resulting in insufficient overall response efficiency and security control capabilities. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent security linkage method and system that integrates facial recognition, so as to solve the shortcomings of the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an intelligent security linkage method integrating face recognition, comprising: The system acquires a sequence of facial images collected within the monitored area, performs quality enhancement processing on the image sequence based on a preset illumination compensation model and angle correction model, and outputs an enhanced facial feature map. The enhanced facial feature map is input into the facial recognition model, combined with the historical feature vector library to perform identity recognition, and outputs the facial identification vector of suspected risky target objects; Based on the face identifier vector, the access trajectory and access control event sequence of the corresponding object in the historical time period are extracted, and a set of linkage trigger features including path offset rate, abnormal access time density and permission level deviation is constructed. Determine whether the set of linkage trigger features satisfies any trigger rule in the multidimensional risk indicator set. If it does, generate a security linkage early warning instruction containing linkage level, time window, and target device ID. The security linkage early warning command is sent to the control node corresponding to the target device ID. The control node includes an access control controller or a broadcast terminal, and executes a response action according to the linkage level.

[0005] Preferably, the quality enhancement processing of the image sequence includes: For the acquired raw face image sequence, extract the brightness distribution features of each frame; Based on the brightness distribution characteristics, the illumination compensation model is invoked to perform local histogram equalization processing on the low-exposure areas in the image. Based on the image after illumination compensation, the face rotation angle is estimated using a 3D keypoint fitting model. Based on the estimated angle, the angle correction model is invoked to perform affine transformation, and the face region of the image is forward-lookingly corrected, outputting an enhanced face feature map for recognition.

[0006] Preferably, based on the illumination-compensated image, a 3D keypoint fitting model is used to estimate the face rotation angle, specifically including: The input is a set of standard facial landmark coordinates extracted from the image. A trained regression network is used to predict the three-dimensional pose angle of the face. If the absolute value of any angle is greater than a set threshold, it is judged as a non-forward-looking face image and needs to enter the angle correction process.

[0007] Preferably, the step of performing identity recognition by combining the historical feature vector library includes: The enhanced facial feature map is input into a facial feature extraction model built on a convolutional neural network to extract the facial feature vector of the current target; The feature subset related to the current time period and monitored area is retrieved from the historical feature vector library to narrow down the comparison range; The similarity score between the current feature vector and the historical feature subset is calculated based on cosine similarity, and a similarity threshold is set to determine whether it is a known identity object; If the similarity score is lower than the similarity threshold and the target object appears more than a preset number of times, its corresponding features are marked as suspected risk targets, and a face identifier vector is output.

[0008] Preferably, the construction of the linkage trigger feature set, which includes path offset rate, abnormal access time density, and permission level deviation, includes: Based on the facial recognition vector association access log database, extract the access control records and camera tracking trajectory of the target object within a preset time period; Path fitting is performed on the spatial distribution of each node in the trajectory, and the path offset rate for the normal travel path of the target object is calculated. Count the number of times the target object enters the high-security area during unauthorized time periods and calculate the abnormal access time density index. The system compares the target object's identity information with the permission level of the area it accesses, assesses the permission level deviation, and combines the path offset rate, abnormal access time density, and permission level deviation to form a multi-dimensional feature set for triggering risk linkage.

[0009] Preferably, determining whether the set of linked triggering features satisfies any triggering rule in the multidimensional risk indicator set includes: The linkage trigger feature set is matched according to the logical expressions defined in the rule base; If any risk rule is triggered, an alert label containing the linkage level will be generated based on the risk level of the corresponding rule. Based on the trigger time, a set time length is traced back to form a response time window, and the target device ID bound to it is called according to the region where the target object is located; The linkage level, time window, and target device ID are encapsulated into a security linkage early warning command.

[0010] Preferably, after the linkage action is completed, the behavioral image sequence over several subsequent time periods is collected and analyzed, and the face recognition model parameter weights are updated based on the analysis results.

[0011] Preferably, after the linkage action is completed, the behavior image sequence of the target object is continuously collected in multiple preset time periods, and the face region in the image is dynamically identified and tracked. The facial feature vectors extracted in each time period are compared with historical label data to label the recognition accuracy and select stable high-confidence samples. The high-confidence samples are used to update the parameter weights of the face recognition model using an online transfer learning algorithm.

[0012] This invention also provides an intelligent security linkage system integrating facial recognition, comprising: The image preprocessing module acquires a sequence of facial images collected within the monitoring area, performs quality enhancement processing on the image sequence based on a preset illumination compensation model and angle correction model, and outputs an enhanced facial feature map. The feature extraction module inputs the enhanced facial feature map into the facial recognition model, performs identity recognition in conjunction with the historical feature vector library, and outputs the facial identification vector of the suspected risk target. The feature construction module extracts the access trajectory and access control event sequence of the corresponding object in the historical time period based on the face identification vector, and constructs a set of linkage trigger features including path offset rate, abnormal access time density and permission level deviation. The early warning generation module determines whether the set of linkage trigger features satisfies any trigger rule in the multidimensional risk indicator set. If it does, it generates a security linkage early warning instruction containing linkage level, time window, and target device ID. The instruction issuing module sends the security linkage early warning instruction to the control node corresponding to the target device ID. The control node includes an access control controller or a broadcast terminal, and executes response actions according to the linkage level.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention provides an intelligent security linkage method integrating facial recognition. Through a collaborative processing flow of image enhancement, facial recognition, behavioral trajectory analysis, and multi-dimensional risk feature extraction, it significantly improves the ability to identify abnormal human behavior in complex environments. By introducing multi-dimensional indicators such as path offset rate, abnormal access time density, and permission level deviation, and intelligently matching them with a risk rule base, it achieves accurate identification and early warning of potentially risky individuals, thereby overcoming the limitations of traditional security systems such as single identification, high false alarm rate, and lack of dynamic behavioral correlation.

[0014] 2. This invention employs a linkage-driven response mechanism, forming a closed-loop linkage between facial recognition results and execution devices such as access control controllers and broadcast terminals, thus achieving an automated processing flow from recognition to handling. Through continuous collection of behavioral data and a high-confidence sample feedback mechanism, combined with an online transfer learning strategy to dynamically optimize the recognition model, the system possesses continuous learning and adaptive capabilities, effectively improving recognition accuracy and environmental adaptability under long-term operation. It is particularly suitable for intelligent security control needs in high-security scenarios such as rail transit, park boundaries, and data centers. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the method of the present invention.

[0017] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1, please refer to Figure 1 As shown in this embodiment, an intelligent security linkage method integrating face recognition includes: The system acquires a sequence of facial images collected within the monitored area, performs quality enhancement processing on the image sequence based on a preset illumination compensation model and angle correction model, and outputs an enhanced facial feature map. The enhanced facial feature map is input into the facial recognition model, combined with the historical feature vector library to perform identity recognition, and outputs the facial identification vector of suspected risky target objects; Based on the face identifier vector, the access trajectory and access control event sequence of the corresponding object in the historical time period are extracted, and a set of linkage trigger features including path offset rate, abnormal access time density and permission level deviation is constructed. Determine whether the set of linkage trigger features satisfies any trigger rule in the multidimensional risk indicator set. If it does, generate a security linkage early warning instruction containing linkage level, time window, and target device ID. The security linkage early warning command is sent to the control node corresponding to the target device ID. The control node includes an access controller or a broadcast terminal, and executes a response action according to the linkage level. After the coordinated action is completed, the behavioral image sequences over several subsequent time periods are collected and analyzed, and the face recognition model parameter weights are updated based on the analysis results.

[0020] In this invention, to improve the recognition accuracy of low-quality face images in a monitoring environment, the acquired original face image sequence needs to undergo quality enhancement processing. This processing mainly includes four consecutive steps: brightness distribution feature extraction, illumination compensation, pose estimation, and angle correction, as detailed below: First, the original face image sequence collected within the monitoring area is analyzed at the pixel level for each frame based on the Y channel (luminance channel) in the YUV color space. The overall brightness distribution of the image is statistically analyzed, and the average brightness value and brightness standard deviation are calculated to determine whether the image is underexposed.

[0021] Specifically, assuming the image frame size is W×H and the pixel brightness value is denoted as Y(x,y), the average image brightness L is calculated as follows: sum the brightness values ​​of all pixels in the image and divide by the total number of pixels, i.e.: x∈[1,W], y∈[1,H]. If the average brightness L is less than the set threshold T1 (e.g., 60), the image is considered to have an underexposure problem, and the illumination compensation step is initiated.

[0022] After identifying insufficient image brightness, the system invokes a pre-defined local illumination compensation model to enhance the brightness of facial areas in the image. This model is based on the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm, suitable for improving the contrast of key facial areas (such as the eyes, bridge of the nose, and mouth), avoiding overexposure caused by overall equalization. Implementation methods include: Centered on facial key points (68 standard key points can be output by the face detection module), several local enhancement regions are constructed; the CLAHE algorithm is applied to each local region, and the contrast gain coefficient of histogram enhancement is limited to between 1.5 and 3.0; finally, a face image with local illumination compensation is output, which improves the clarity of the recognition features of the main facial regions.

[0023] Based on the illumination-compensated image, the orientation angle of the face image is determined. This invention employs a 3D facial landmark fitting model for pose estimation. This model uses deep learning methods to perform 3D fitting of the key points of the face in the image, outputting the pitch, yaw, and roll angles of the face. The model construction method is as follows: The input is a set of standard facial landmark coordinates extracted from the image; a trained regression network (e.g., a lightweight model based on MobileNet or ResNet) is used to predict the 3D pose angles of the face; if the absolute value of any angle is greater than a set threshold T2 (e.g., ...), the face pose angles are determined by the input. If the image is not in front-view, it is determined to be a non-front-view face image and needs to enter the angle correction process.

[0024] After determining that the face deviates from the forward-looking angle, a preset affine transformation angle correction model is invoked to geometrically adjust the image. This model constructs a transformation matrix between the target pose and the current pose to correct the non-forward-looking face image to a standard forward-looking angle, improving the accuracy of subsequent recognition. The specific processing method is as follows: Using five points—eyes, nose tip, and corners of mouth—as references, calculate the affine transformation matrix between the source image and the front-view template image; use image processing frameworks such as OpenCV or TensorFlow to call affine transformation functions to rotate, scale, and translate the entire image; output the enhanced facial feature map with angle correction completed as the input image for subsequent recognition models.

[0025] In this invention, to achieve accurate recognition of enhanced facial feature maps and effective screening of potential risk targets, the processed image needs to be input into the feature extraction and comparison module and combined with a historical feature vector database for recognition. This part includes the following sequential steps: The image, after quality enhancement, is input into a face feature extraction model to extract the facial feature vector of the current target. This model is built upon a deep convolutional neural network, preferably using FaceNet, ArcFace, or a lightweight MobileFaceNet architecture. It performs multi-layer convolution and normalization processing on the facial regions of the image, outputting a fixed-length vector (typically 128-dimensional or 512-dimensional) representing the deep semantic features of the current face. The feature vector is represented as... This forms the basis for subsequent comparisons with data in the historical database.

[0026] To improve the efficiency and real-time performance of face comparison, a subset of features is dynamically selected from a pre-set historical feature vector library, based on the current monitoring time period and data collection location information. The selection rules for this subset include: Time filtering: Select facial feature records that have appeared in this area within the past 48 hours; Region filtering: Filters records that match the current monitoring point ID; Status filtering: Removes high-frequency, resident personnel characteristics that the system marks as "trusted objects".

[0027] This step can be achieved by configuring an indexed database (such as Annoy or FAISS) to enable fast subset queries and approximate vector retrieval.

[0028] Call the similarity calculation module to perform the calculation on the current feature vector. Each feature vector in the historical feature subset Calculate cosine similarity The calculation formula is: Set a similarity threshold T3 (recommended value is 0.65). If the highest similarity score is lower than T3, it means that the current target has no clear matching identity in the historical database, and proceed to the next step of risk screening.

[0029] When an unknown object is identified, the current facial recognition vector is recorded, and its frequency of appearance is tracked in subsequent monitoring frames. If the target object appears more than a preset frequency threshold (e.g., N=3 times) within a set time window (e.g., 5 minutes), the system determines the object as a "suspected risk target." The judgment criteria here are: unauthorized objects frequently appearing in specific areas may exhibit risky behaviors such as tailing, probing, or loitering; and risk scoring is also combined with other dimensions such as area access permissions and time periods.

[0030] Finally, the feature vectors of the objects are marked as "to be handled" and used as the input basis for subsequent behavior analysis and security linkage, ensuring the proactiveness and targeting of the security system's response.

[0031] Based on the facial recognition vector, the access trajectory and access control event sequence of the corresponding object within a historical time period are extracted, and a set of linkage trigger features including path offset rate, abnormal access time density, and permission level deviation is constructed. The specific implementation steps are as follows: The system associates the currently identified facial recognition vector with the access log database, retrieving access control records and video tracking trajectories of the target object within a preset time period (e.g., the last 72 hours). Access control record data includes timestamps, access control device numbers, entry / exit directions, and card swiping or recognition methods. The video tracking trajectory consists of location information automatically recorded by cameras, combined with camera location numbers and timestamps to form a spatial-temporal sequence of trajectory points. This step quickly retrieves complete access behavior data of the target by matching the unique facial identifier field in the database.

[0032] The acquired trajectory point sequence is smoothed and fitted using a Bézier curve or Kalman filter, and then compared with the historical normal behavior path of the target object to calculate the path offset rate.

[0033] Path offset rate is defined as the spatial deviation between the current path of a target object and its regular path within the same time period and region. The specific calculation method is as follows: For each trajectory, point sets are resampled in chronological order to ensure a uniform dimension; the Euclidean distance between corresponding points on the current trajectory and historical trajectories is calculated, and the average value is taken as the offset rate; if the offset rate exceeds a preset threshold... If the distance is less than 5 meters, it is considered an abnormal path. This indicator is used to identify suspicious behaviors such as potential detours, probing movements, and attempts to evade monitoring.

[0034] Clustering and statistical analysis are performed on the timestamps of each entry and exit event in the target object's access records to identify the frequency of its entry into high-security areas during unauthorized time periods. The implementation is as follows: Define the authorized access time range based on the target's identity (e.g., 08:00–18:00); classify the security level of each access control device in the system according to the area level configuration file; count the number of times the target object enters level 2 or higher security areas during unauthorized time periods, and divide this number by the total length of unauthorized time periods to obtain the abnormal access time density value (unit: times / hour); when this indicator exceeds a threshold... When the frequency is 0.2 times per hour, the system will classify the behavior as an abnormal access behavior.

[0035] By comparing the permission level in the target object's identity configuration with the difference between its actual region level in its access records, an assessment is made as to whether any permission overreach has occurred. The specific calculation method is as follows: assign security level labels (e.g., levels 1-5, with higher numbers indicating higher security levels) to all system regions; extract the highest permissible access level from the target object's identity configuration. ; In the statistics of their access behavior, the regional permission level is greater than The number and proportion of events form a permission deviation factor; if the deviation proportion exceeds a threshold... If it is 10%, it will be marked as unauthorized access.

[0036] The path offset rate, abnormal access time density, and permission level deviation are used as joint features input into the linkage rule engine to construct a linkage trigger feature set. This set is used to match and judge with preset risk patterns to determine whether to enter the linkage response process.

[0037] The method uses a weighted combination to comprehensively score the three indicators by default, and the weight of each indicator can be configured according to the actual scenario (e.g., path offset rate 40%, time density 30%, permission deviation 30%). This method takes into account three behavioral characteristics: spatial anomalies, temporal anomalies, and permission violations, thereby improving the ability to identify complex risk behaviors.

[0038] After constructing the linkage trigger feature set, this invention determines whether the linkage trigger feature set satisfies any trigger rule in the multi-dimensional risk indicator set. If it does, a security linkage early warning instruction containing the linkage level, time window, and target device ID is generated. This process includes the following four consecutive steps: This step inputs the constructed set of linkage trigger features into the preset risk rule matching engine, executing conditional judgments based on logical expressions. The rule base defines several risk rule items in a structured manner, each rule item consisting of logical conditions and corresponding response parameters. An example rule structure is as follows: Rule Number: R001; Logical Expression: Path offset rate ≥ 5 and permission level deviation ≥ 2; Linkage Level: Level II; Target Device Type: Access Control Controller + Voice Broadcast; Response Window: 10 minutes.

[0039] The matching process substitutes specific values ​​from the linked feature set (such as a path offset rate of 6.2 and a permission level deviation of 3) into a logical expression and uses Boolean calculations to determine whether the rule conditions are met. This process can be implemented based on a rule tree or a rule engine written in a DSL (Domain-Specific Language) (such as Drools or nools), and supports dynamic loading and hot updates.

[0040] When any risk rule is successfully triggered, a corresponding linkage level label is generated based on the preset risk level parameters of that rule, such as "Level I Linkage," "Level II Linkage," or "Early Warning Mode." Level I linkage indicates a high-risk state, requiring multi-device responses such as lockdown, broadcasting, and image tracking; Level II linkage represents medium risk, typically triggering access control restrictions and local alerts; Early Warning Mode is used for early warning scenarios where behavior is suspicious but does not constitute a substantial threat. This label will serve as the core content of the "Linkage Level" field in subsequent instructions.

[0041] After identifying the risk trigger time, the system automatically traces back based on the response time parameters configured in the rule to generate a fixed-length response time window, such as "within 10 minutes" or "within 30 minutes". This window is used to interface with subsystems such as device control strategies and video backtracking analysis. Simultaneously, combined with the spatial ID of the target object's current location, the system extracts the device IDs requiring a response from the area-device binding table, including access control controller IDs, broadcast terminal IDs, and warning light IDs. For example, if the object is in "Area A2", the system will retrieve device IDs such as "Access Control A2-03" and "Broadcast A2-01" from the binding table. Finally, the system integrates all the above elements to form a structured security linkage early warning instruction, in the following format: {"Linkage Level":"Level II","Time Window":"2025-11-03 14:20:00 to 14:30:00","Target Device ID":["DMJ-A2-03","BCT-A2-01"],"Trigger Rule Number":"R001"}. The instruction is encapsulated in JSON or XML format and pushed to the central control module or device middleware platform. It is then distributed in real time via WebSocket or MQTT protocol to ensure timely response and synchronous execution of the linkage action.

[0042] After facial recognition and risk rule matching are completed, this invention generates a structured security linkage early warning command. This command needs to be sent to the target control node device within the designated area, and different intensity response actions are triggered according to the linkage level in the command. The function of this module is to establish a linkage link between the intelligent recognition system and the execution control device, realizing closed-loop control of automatic response and on-site handling. The specific implementation method is as follows: The control nodes are execution devices deployed in key areas, and their types include, but are not limited to, the following two categories: Access control controller: Used to control the opening and closing of entrance and exit channels, supporting multiple door opening modes such as face recognition, card swiping, and remote authorization, and has the ability to receive remote commands; Broadcasting terminals include fixed voice broadcasting equipment and area broadcasting loudspeakers, which can be used to publish warning audio or instruction information and have linkage triggering interfaces.

[0043] All control nodes are connected to the security IoT platform and maintain long-term communication with the central control system via MQTT (Message Queuing Telemetry Transport) or WebSocket protocols, supporting real-time command reception and feedback reporting.

[0044] After the rule is triggered, the IP address, communication protocol type, and control command format template of the target device are parsed from the device mapping table based on the target device ID.

[0045] The instruction encapsulation format is as follows: {"Device ID":"DMJ-A2-03","Control Type":"Access Control","Linkage Level":"Level II","Action":"Lock Access","Effective Time":"2025-11-03 14:22:00","Duration":"10 minutes"}. The "Action" field is automatically assigned based on the linkage level, such as "Emergency Lock + Voice Warning" for Level I and "Lock Only" for Level II. The duration is defined in the warning rules, typically ranging from 5 to 30 minutes. After processing by a unified linkage instruction scheduling module, the instruction is sent to the target device control node via communication middleware.

[0046] After receiving the instruction, the target control node parses the control content and executes the corresponding action. The execution logic is as follows: For access control controllers: after parsing the "lock" instruction, the corresponding entrance and exit channels are immediately closed, prohibiting all unauthorized passage; if the instruction is accompanied by a "priority identification list", then whitelisted individuals are allowed to pass through the local facial recognition terminal, balancing control and passage efficiency; the controller will transmit the execution status back to the central platform in real time, marking it as "closed" or "unlocked".

[0047] For broadcast terminals: call the voice template library of the corresponding area and synthesize audio content according to the warning level; start the broadcast task, and the broadcast frequency and volume are automatically adjusted according to the linkage level, such as repeating the broadcast every minute at level I with the maximum volume; if the object has been captured and tracked, the broadcast information may contain a description of the target's appearance or operation prompts (such as: "Please do not enter the area ahead").

[0048] All control nodes must return an acknowledgment frame within 5 seconds of receiving the instruction. The returned frame must include the device ID, reception time, instruction status, and execution status code (e.g., 200 for successful execution, 503 for device offline). If there is no response within 10 seconds or multiple failures occur, the system will trigger the backup device control scheme or issue a manual intervention notification.

[0049] In the implementation of this invention, after the linkage action is completed, the behavioral image sequences over several subsequent time periods are collected and analyzed, and the face recognition model parameter weights are updated based on the analysis results. The specific implementation steps are as follows: After the linkage response is completed, based on the identified target object's identity vector, image frame sequences of the target object appearing in the monitoring screen are continuously acquired within N set time periods (such as the next three 5-minute windows). Camera resources deployed in the target area are invoked, and combined with object trajectory information, the face dynamic detection and trajectory tracking module is activated. This module, based on multi-target tracking algorithms (such as SORT or Deep SORT), combines target appearance features and motion information to perform continuous intra-frame tracking of the same object. Simultaneously, a face detection model (such as MTCNN or RetinaFace) is used to identify face regions in the image frames, and face images are extracted for subsequent analysis and processing.

[0050] The acquired behavioral image sequence is processed for facial feature extraction. The same deep feature extraction model (such as ArcFace) as the original recognition process is used to extract the feature vector of each frame of facial image.

[0051] Subsequently, the facial feature vectors extracted in each time period are compared with the identifier vectors output by the target object during the recognition phase, and the cosine similarity algorithm is used to calculate the degree of matching: If the similarity is higher than the set threshold If a sample has a similarity score of 0.75 and appears stably in consecutive frames, it is considered a "high-confidence sample." The recognition accuracy is calculated for each time period, and frames with low confidence or high noise interference are removed to ensure the quality of subsequent training samples. This step can use a sliding window strategy to smooth similarity changes and avoid occasional errors caused by pose or occlusion.

[0052] The face image samples marked with high confidence are input into the model update module to perform online transfer learning.

[0053] The online transfer learning is implemented based on the following strategies: using a frozen base network structure (e.g., keeping the ArcFace backbone unchanged), only fine-tuning the parameters of the last two fully connected layers; employing mini-batch gradient descent (Mini-batch SGD) for model training, with no more than 16 samples per batch to avoid overfitting; setting the learning rate to a low value (e.g., 0.0001) and using L2 regularization constraints to ensure smooth model weight updates; and aiming to minimize the vector distance between the target sample and its historical vectors, constructing a target recognition subtask to improve the model's adaptability to the features of this type of target.

[0054] After training is completed, the updated face recognition model weight file is synchronized to the inference engine (such as TensorRT or OpenVINO environment) deployed on the face recognition terminal to ensure higher robustness in subsequent recognition under similar lighting, angle and scene conditions.

[0055] The update strategy employs a soft update mechanism, which integrates the new model weights proportionally into the currently deployed model. For example, the new weight ratio is α (suggested value: 0.2), and the old model weight ratio is... The weighted fusion formula is: . For the new weights after merging, For the old model weights, These are temporary updated weights obtained after training with new samples.

[0056] Example 2, please refer to Figure 2 As shown in the figure, the intelligent security linkage system integrating face recognition described in this embodiment includes: The image preprocessing module acquires a sequence of facial images collected within the monitoring area, performs quality enhancement processing on the image sequence based on a preset illumination compensation model and angle correction model, and outputs an enhanced facial feature map. The feature extraction module inputs the enhanced facial feature map into the facial recognition model, performs identity recognition in conjunction with the historical feature vector library, and outputs the facial identification vector of the suspected risk target. The feature construction module extracts the access trajectory and access control event sequence of the corresponding object in the historical time period based on the face identification vector, and constructs a set of linkage trigger features including path offset rate, abnormal access time density and permission level deviation. The early warning generation module determines whether the set of linkage trigger features satisfies any trigger rule in the multidimensional risk indicator set. If it does, it generates a security linkage early warning instruction containing linkage level, time window, and target device ID. The instruction issuing module sends the security linkage early warning instruction to the control node corresponding to the target device ID. The control node includes an access control controller or a broadcast terminal, and executes response actions according to the linkage level.

[0057] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A smart security linkage method integrating facial recognition, characterized in that: include: The system acquires a sequence of facial images collected within the monitored area, performs quality enhancement processing on the image sequence based on a preset illumination compensation model and angle correction model, and outputs an enhanced facial feature map. The enhanced facial feature map is input into the facial recognition model, combined with the historical feature vector library to perform identity recognition, and outputs the facial identification vector of suspected risky target objects; Based on the face identifier vector, the access trajectory and access control event sequence of the corresponding object in the historical time period are extracted, and a set of linkage trigger features including path offset rate, abnormal access time density and permission level deviation is constructed. Determine whether the set of linkage trigger features satisfies any trigger rule in the multidimensional risk indicator set. If it does, generate a security linkage early warning instruction containing linkage level, time window, and target device ID. The security linkage early warning command is sent to the control node corresponding to the target device ID. The control node includes an access control controller or a broadcast terminal, and executes a response action according to the linkage level.

2. The intelligent security linkage method integrating face recognition according to claim 1, characterized in that: in, The quality enhancement processing of the image sequence includes: For the acquired raw face image sequence, extract the brightness distribution features of each frame; Based on the brightness distribution characteristics, the illumination compensation model is invoked to perform local histogram equalization processing on the low-exposure areas in the image. Based on the image after illumination compensation, the face rotation angle is estimated using a 3D keypoint fitting model. Based on the estimated angle, the angle correction model is invoked to perform affine transformation, and the face region of the image is forward-lookingly corrected, outputting an enhanced face feature map for recognition.

3. The intelligent security linkage method integrating face recognition according to claim 2, characterized in that: Based on the illumination-compensated image, a 3D keypoint fitting model is used to estimate the face rotation angle, specifically including: The input is a set of standard facial landmark coordinates extracted from the image. A trained regression network is used to predict the three-dimensional pose angle of the face. If the absolute value of any angle is greater than a set threshold, it is judged as a non-forward-looking face image and needs to enter the angle correction process.

4. The intelligent security linkage method integrating face recognition according to claim 1, characterized in that: in, The process of performing identity recognition by combining a historical feature vector library includes: The enhanced facial feature map is input into a facial feature extraction model built on a convolutional neural network to extract the facial feature vector of the current target; The feature subset related to the current time period and monitored area is retrieved from the historical feature vector library to narrow down the comparison range; The similarity score between the current feature vector and the historical feature subset is calculated based on cosine similarity, and a similarity threshold is set to determine whether it is a known identity object; If the similarity score is lower than the similarity threshold and the target object appears more than a preset number of times, its corresponding features are marked as suspected risk targets, and a face identifier vector is output.

5. The intelligent security linkage method integrating face recognition according to claim 1, characterized in that: in, The constructed set of linkage trigger features includes path offset rate, abnormal access time density, and permission level deviation, including: Based on the facial recognition vector association access log database, extract the access control records and camera tracking trajectory of the target object within a preset time period; Path fitting is performed on the spatial distribution of each node in the trajectory, and the path offset rate for the normal travel path of the target object is calculated. Count the number of times the target object enters the high-security area during unauthorized time periods and calculate the abnormal access time density index. The system compares the target object's identity information with the permission level of the area it accesses, assesses the permission level deviation, and combines the path offset rate, abnormal access time density, and permission level deviation to form a multi-dimensional feature set for triggering risk linkage.

6. The intelligent security linkage method integrating face recognition according to claim 1, characterized in that: in, Determining whether the set of linked triggering features satisfies any triggering rule in the multidimensional risk indicator set includes: The linkage trigger feature set is matched according to the logical expressions defined in the rule base; If any risk rule is triggered, an alert label containing the linkage level will be generated based on the risk level of the corresponding rule. Based on the trigger time, a set time length is traced back to form a response time window, and the target device ID bound to it is called according to the region where the target object is located; The linkage level, time window, and target device ID are encapsulated into a security linkage early warning command.

7. The intelligent security linkage method integrating face recognition according to claim 1, characterized in that: After the coordinated action is completed, the behavioral image sequences over several subsequent time periods are collected and analyzed, and the face recognition model parameter weights are updated based on the analysis results.

8. The intelligent security linkage method integrating face recognition according to claim 7, characterized in that: After the linkage action is completed, the system continuously collects the behavior image sequence of the target object within multiple preset time periods, and performs dynamic recognition and trajectory tracking of the face region in the image. The facial feature vectors extracted in each time period are compared with historical label data to label the recognition accuracy and select stable high-confidence samples. The high-confidence samples are used to update the parameter weights of the face recognition model using an online transfer learning algorithm.

9. An intelligent security linkage system integrating facial recognition, used to implement the intelligent security linkage method integrating facial recognition as described in any one of claims 1-8, characterized in that: include: The image preprocessing module acquires a sequence of facial images collected within the monitoring area, performs quality enhancement processing on the image sequence based on a preset illumination compensation model and angle correction model, and outputs an enhanced facial feature map. The feature extraction module inputs the enhanced facial feature map into the facial recognition model, performs identity recognition in conjunction with the historical feature vector library, and outputs the facial identification vector of the suspected risk target. The feature construction module extracts the access trajectory and access control event sequence of the corresponding object in the historical time period based on the face identification vector, and constructs a set of linkage trigger features including path offset rate, abnormal access time density and permission level deviation. The early warning generation module determines whether the set of linkage trigger features satisfies any trigger rule in the multidimensional risk indicator set. If it does, it generates a security linkage early warning instruction containing linkage level, time window, and target device ID. The instruction issuing module sends the security linkage early warning instruction to the control node corresponding to the target device ID. The control node includes an access control controller or a broadcast terminal, and executes response actions according to the linkage level.

Images