Intelligent access control system

Abstract

The invention relates to the technical field of access control, in particular to an intelligent access control system which comprises a multi-mode sensor group, a data processing module, a model judgment module and an execution output module, the multi-modal sensor group is used for collecting multi-dimensional data when a user passes; the data processing module is used for performing feature extraction and preprocessing on the multi-dimensional data to generate structured features; the model judgment module is internally provided with a pre-training identity verification model and a pre-training stress state detection model which are respectively used for analyzing the structured features and outputting an identity matching score and a stress risk score; and the execution output module is used for executing access control operation according to the identity matching score and the stress risk score. The problem that a traditional intelligent access control system depends on a single biological characteristic, so that the identity verification accuracy is low in a shielding scene is solved, and the security vulnerability that hidden recognition cannot be achieved and emergency response cannot be triggered when the system is stressed is solved.

Description

Technical Field

[0001] This invention relates to the field of access control technology, specifically an intelligent access control system. Background Technology

[0002] Intelligent access control systems, as core devices combining security and convenient access, are widely used in communities, financial institutions, high-end office buildings, and other scenarios. They achieve personnel identity verification and access control through biometric recognition, card swiping, and other methods, aiming to balance security management and access efficiency. However, in existing technologies, when users wear masks, goggles, or have indistinct facial features, such as the elderly or children, the accuracy of single biometric features, such as facial recognition, is easily affected by interference from obscured areas, leading to a significant decrease in recognition accuracy. This cannot meet the identity verification requirements in high-security scenarios. At the same time, when users are coerced, such as by being threatened with a knife and forced to open the door, traditional systems rely on users to actively alarm, such as by entering a specific password or external monitoring intervention. They cannot identify the coercive state in real time and trigger an emergency response in a covert manner, which poses a security risk. Traditional access control systems suffer from numerous problems. For example, they rely on physical access cards, which are easily lost or damaged, and current ID / IC cards are easily cracked and copied, leading to security vulnerabilities. Their compatibility is also poor, supporting only a single physical access card and failing to adapt to modern smart devices such as smartphones and smartwatches, nor supporting the NFC functionality of various smart devices. Furthermore, they lack remote control and remote door opening capabilities, requiring on-site operation by the user, failing to meet temporary authorization needs or flexibly grant temporary visitor permissions. Additionally, the user experience is limited, with fixed prompts that cannot be customized, failing to meet personalized needs. This invention aims to solve these problems by providing a multifunctional, highly compatible, and remotely controllable smart access control terminal. Summary of the Invention

[0003] This application provides an intelligent access control system that solves the problem of low identity verification accuracy in obstructed scenarios caused by the reliance on a single biometric feature in traditional intelligent access control systems, as well as the security vulnerability of being unable to concealed and trigger emergency response when under duress. It improves the verification reliability and proactive protection capabilities in complex scenarios through multimodal data acquisition and fusion analysis.

[0004] To achieve the above objectives, this application discloses the following technical solution: an intelligent access control system, comprising a multimodal sensor group, a data processing module, a model judgment module, and an execution output module; the multimodal sensor group is used to collect multi-dimensional data during user access; the data processing module is used to extract and preprocess the multi-dimensional data to generate structured features; the model judgment module incorporates a pre-trained identity verification model and a pre-trained duress state detection model, which are used to analyze the structured features and output an identity matching score and a duress risk score, respectively; the execution output module executes access control operations based on the identity matching score and the duress risk score.

[0005] The multimodal sensor group includes a visible light imaging device, an infrared imaging device, an acoustic acquisition device, a motion sensing device, and a physiological signal acquisition device. The visible light imaging device acquires visible light images of the user's face; the infrared imaging device acquires infrared thermal images of the user's face; the acoustic acquisition device acquires the user's voice signals; the motion sensing device acquires the user's motion trajectory data; and the physiological signal acquisition device acquires the user's physiological state signals. The data processing module performs registration processing on the visible light and infrared images, identifies occluded areas, and extracts local features of the unoccluded face; it also performs denoising processing on the voice signals and extracts voiceprint features. The motion trajectory data is analyzed to extract gait features; the physiological state signal is filtered to extract physiological features; the model judgment module inputs the unobstructed facial features, the voiceprint features, and the gait features into the pre-trained identity verification model and outputs the identity matching score; the micro-expression features corresponding to the unobstructed facial features, the voice tension features corresponding to the voiceprint features, and the physiological features are input into the pre-trained coercion state detection model and output the coercion risk score; the execution output module performs operations such as opening the access control, generating verification instructions, triggering warnings, or denying passage based on the combination of the identity matching score and the coercion risk score.

[0006] This invention relates to an intelligent access control system that collects multi-dimensional data, including visible light images, infrared thermal imaging, voice signals, motion trajectories, and physiological states, using a multi-modal sensor array. Combined with NFC device reading of identification information, this achieves multi-source complementary acquisition of user identity information, improving the robustness of identity verification in complex scenarios. The data processing module extracts structured data such as unobstructed facial features, voiceprints, gait, physiological characteristics, and NFC features through preprocessing steps including image registration, denoising, trajectory analysis, and signal filtering, effectively enhancing the recognizability and reliability of these features. The model judgment module utilizes a pre-trained identity verification model and a coercion state detection model to respectively assess the... The system integrates structured features for analysis, outputting identity matching scores and coercion risk scores. This dual-dimensional evaluation mechanism avoids misjudgments based on single features, improving the accuracy of security assessments. The execution output module performs operations such as opening access control, generating verification commands, triggering alerts, or denying access based on the score combination, covering all scenarios including normal access, temporary verification, coercion alerts, and abnormal rejection, enhancing the system's adaptability. The addition of near-field communication reading devices expands authentication methods, supporting contactless rapid identification. Offline mode operation ensures the availability of basic functions during network interruptions, while local log storage and recovery mechanisms provide data support for subsequent auditing and optimization. Through multi-module collaboration and multi-dimensional optimization, this design achieves highly secure, highly adaptable, and highly reliable intelligent access control. Attached Figure Description

[0007] Figure 1 This is a system block diagram according to an embodiment of the present invention;

[0008] Figure 2 This is a module interaction diagram of an embodiment of the present invention. Detailed Implementation

[0009] Specific embodiments of the invention will now be described in detail. Although the invention is described in conjunction with these specific embodiments, it should be understood that the invention is not intended to be limited to these specific embodiments. Rather, these embodiments are intended to cover alternative, modified, or equivalent embodiments that may be included within the spirit and scope of the invention as defined by the claims. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. The invention may be practiced without some or all of these specific details. In other instances, well-known processes have not been described in detail so as not to unnecessarily obscure the invention.

[0010] When used in conjunction with the terms "comprising," "method comprising," or similar language in this specification and appended claims, the singular forms "a," "some," and "the" include plural references unless the context clearly indicates otherwise. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0011] Application Overview: In existing technologies, intelligent access control systems mostly rely on single biometric features such as facial recognition for identity verification. When users are wearing masks, goggles, or have obscured facial features, such as the elderly or children, this single feature is easily interfered with, leading to a significant decrease in recognition accuracy. Furthermore, the system cannot covertly identify users in a coerced state, such as when threatened with a knife, and can only rely on user-initiated alarms or external monitoring intervention, resulting in a security response delay vulnerability. If these issues are not addressed, identity verification failures in obscured scenarios may lead to unauthorized entry, while the inability to promptly trigger alarms in coerced scenarios will threaten user safety.

[0012] To address the aforementioned issues, this application first considers compensating for the limitations of single features through multimodal data acquisition. It designs a multimodal sensor array encompassing visible light imaging, infrared imaging, acoustic acquisition, motion sensing, physiological signal acquisition, and near-field communication devices to simultaneously acquire multi-dimensional data such as facial images, speech, motion trajectories, physiological states, and NFC tags. Subsequently, a data processing module performs image registration to extract unobstructed facial features, denoises speech to extract voiceprints, analyzes motion trajectories to extract gait, and filters physiological signals to extract physiological features, combining these with NFC tags to generate structured features. Furthermore, a pre-trained identity verification model is used to fuse multi-dimensional features to calculate identity matching scores, and a pre-trained coercion state detection model analyzes micro-expressions, speech tension, and physiological signals to calculate coercion risk scores. Finally, an execution output module performs operations such as opening access control, generating verification commands, triggering warnings, or denying passage based on the combination of the two scores, achieving high-accuracy verification and covert coercion state identification in occluded scenarios, thus improving system security and adaptability.

[0013] Example

[0014] An intelligent access control system includes a multimodal sensor group, a data processing module, a model judgment module, and an execution output module. The multimodal sensor group is used to collect multi-dimensional data during user access. The data processing module is used to extract and preprocess the multi-dimensional data to generate structured features. The model judgment module has a built-in pre-trained identity verification model and a pre-trained duress state detection model, which are used to analyze the structured features and output identity matching scores and duress risk scores, respectively. The execution output module executes access control operations based on the identity matching scores and the duress risk scores.

[0015] In this embodiment, multi-dimensional verification and security control are achieved through multi-module collaboration. A multimodal sensor array is deployed at the access control front end. When a user approaches, the data acquisition process is automatically triggered, simultaneously acquiring data on facial features, voice, motion, physiological data, and near-field communication. The data processing module acts as an intermediary. After receiving the sensor data, it performs standardized preprocessing, transforming unstructured image and voice data into analyzable structured features before transmitting it to the model judgment module. The pre-trained identity verification model in the model judgment module performs identity comparison by fusing multiple feature classes, while the pre-trained coercion state detection model mines potential coercion signals from the features. The scores output by both models serve as the basis for execution. The execution output module establishes a stable connection with the cloud server via a 4G / 5G network. It supports both local execution based on scores and response to remote commands. When a user initiates a remote door opening request via a mobile app, the command is encrypted and transmitted to the execution output module via the cloud. After verifying the request's legitimacy, the module controls the access control to open. The entire system's collaborative process involves sensor acquisition, data processing, model judgment, and operation execution. Each step achieves low-latency communication through an internal bus, ensuring that the entire process from user triggering to access control response is completed in a short time, balancing security and access efficiency.

[0016] This solution further proposes that the multimodal sensor group includes a visible light imaging device, an infrared imaging device, an acoustic acquisition device, a motion sensing device, and a physiological signal acquisition device. The visible light imaging device is used to acquire visible light images of the user's face; the infrared imaging device is used to acquire infrared thermal images of the user's face; the acoustic acquisition device is used to acquire the user's voice signals; the motion sensing device is used to acquire the user's motion trajectory data; and the physiological signal acquisition device is used to acquire the user's physiological state signals. The data processing module performs registration processing on the visible light images and the infrared images, identifies occluded areas, and extracts local features of the unoccluded face; it also performs denoising processing on the voice signals and extracts voiceprint features. The system performs trajectory analysis on the motion trajectory data and extracts gait features; it also performs signal filtering on the physiological state signals and extracts physiological features; the model judgment module inputs the unobstructed facial local features, the voiceprint features, and the gait features into the pre-trained identity verification model and outputs the identity matching score; it inputs the micro-expression features corresponding to the unobstructed facial local features, the voice tension features corresponding to the voiceprint features, and the physiological features into the pre-trained coercion state detection model and outputs the coercion risk score; the execution output module performs operations such as opening the access control, generating verification instructions, triggering warnings, or refusing passage based on the combination of the identity matching score and the coercion risk score.

[0017] In this embodiment, the various devices in the multimodal sensor group have clearly defined roles and work collaboratively. The visible light imaging device captures facial details under natural or supplemental lighting conditions, while the infrared imaging device is unaffected by light and can clearly acquire facial thermal distribution features even in backlit or dim environments. After the images acquired by both are registered by the data processing module, areas obscured by masks, sunglasses, etc., can be accurately located, and only unobscured local features such as the forehead and corners of the eyes are extracted for identity verification. The acoustic acquisition device captures the user's voice in real time, such as the "open the door" command or natural conversation. After noise reduction processing to remove environmental noise, voiceprint features are extracted—voiceprints are unique and can complement facial features. The motion sensing device continuously tracks the user's movement trajectory from approach to stopping, analyzing gait features such as stride length and stride frequency. Gait, as a dynamic biometric feature, is difficult to imitate, enhancing the reliability of identity verification. The physiological signal acquisition device acquires signals such as the user's heart rate and skin conductance through contact or non-contact methods. Fluctuations in these signals can reflect stress states, providing a basis for stress detection.

[0018] The model judgment module utilizes features in a layered logic: the identity verification model focuses on integrating unobstructed facial features, voiceprint features, and gait features, cross-verifying identity from both static and dynamic dimensions; the coercion state detection model focuses on facial micro-expressions, such as involuntary muscle twitches, vocal tension, and physiological characteristics, such as increased pitch, erratic speech rate, and sudden increases in heart rate. These signals are typically difficult to actively control, ensuring the concealment of coercion identification. When the execution output module performs operations based on the score combination, it links with near-field communication data. If the user simultaneously presents an NFC device, its identification information will serve as an auxiliary feature to improve verification efficiency, especially suitable for scenarios where holding objects with both hands makes biometric verification inconvenient.

[0019] This solution further proposes that the image acquisition device in the multimodal sensor group has a resolution and frame rate suitable for acquiring user facial image information, the acoustic acquisition device has a sampling rate suitable for acquiring user voice signals, the motion sensing device has a preset detection range and angular resolution, and the physiological signal acquisition device transmits encrypted data to the data processing module via wireless communication.

[0020] The design of the parameters of each device in the multimodal sensor array is centered on adapting to the actual scene. The resolution and frame rate of the image acquisition device need to meet the requirements for capturing facial details, clearly identifying subtle features such as wrinkles around the eyes, while avoiding motion blur caused by a low frame rate, ensuring that effective images can still be acquired when users pass by quickly. The sampling rate of the acoustic acquisition device needs to match the frequency range of human speech to accurately capture speech signals. At the same time, directional sound pickup design reduces background noise interference, so that the user's voiceprint can still be clearly extracted even in noisy office areas or community environments.

[0021] The motion sensing device's detection range covers the entire process of a user approaching the access control from 3 meters away until they stop in front of the door. Its angular resolution ensures it can distinguish between different users standing side-by-side, avoiding trajectory confusion. When the physiological signal acquisition device uses wireless communication, it employs a dedicated encryption protocol to encrypt the data end-to-end, preventing interception or tampering during transmission and protecting the user's physiological privacy.

[0022] As a supplement to the multimodal sensor group, the near-field communication (NFC) reader supports multiple devices by being compatible with various protocols: when recognizing the transit card signal of an Apple phone, it parses the identification information through Apple's Core NFC framework; when reading the virtual access card of an Android phone, it adapts to its HCE-based analog signal; for smart bracelets and watches, it ensures stable recognition by adjusting the sensing distance, thus solving the problem of signal strength differences between different devices.

[0023] This solution further proposes that when the data processing module performs registration processing on the visible light image and the infrared image, it adopts a feature point matching algorithm and ensures that the superposition error is within a preset range through feature extraction and mismatch point removal algorithms; when performing noise reduction processing on the speech signal, it adopts a processing algorithm to suppress environmental noise; when performing trajectory analysis on the motion trajectory data, it adopts a feature extraction algorithm to obtain gait-related parameters; and when performing signal filtering on the physiological state signal, it adopts a filtering algorithm to extract physiological feature parameters.

[0024] The data processing module prioritizes accuracy and robustness in processing various types of data. When registering visible light and infrared images, it first identifies common facial key points in both images, such as pupils and the tip of the nose, using a feature point matching algorithm. Then, it uses a false matching point removal algorithm to eliminate false matches caused by occlusion or lighting conditions. Ultimately, this keeps the superposition error of the two images within an acceptable range, ensuring that feature extraction from unoccluded areas is accurate.

[0025] When denoising speech signals, an algorithm for suppressing environmental noise is adopted: first, the environmental noise characteristics during silent periods are analyzed and a noise model is established; then, the noise component corresponding to the model is subtracted from the speech signal, while preserving the original characteristics of the speech, making the processed voiceprint features purer and improving the stability of identity verification.

[0026] When analyzing motion trajectory data, the continuous trajectory is decomposed into gait cycles through feature extraction algorithms, such as the time from the start of one step to the start of the next, stride length, horizontal distance between two steps, and number of steps per unit time. These parameters are standardized to form gait feature vectors, so that even if the user is wearing different shoes or carrying different items, the user's identity can still be matched through the core parameters.

[0027] When filtering physiological state signals, a filtering algorithm is used to filter out high-frequency interference caused by breathing and movement, and to extract low-frequency features such as heart rate variability and skin conductance. The changing trends of these features can effectively reflect the user's emotional state and provide a reliable basis for stress detection.

[0028] This solution further proposes that the input of the pre-trained identity verification model is the unobstructed facial local feature vector, the voiceprint feature vector, the gait feature vector, and the NFC feature vector, and the identity matching score is output after feature concatenation; the input of the pre-trained coercion state detection model is the micro-expression feature vector, the voice tension feature vector, and the physiological feature vector, and the coercion risk score is output after feature weighting.

[0029] The feature fusion of the pre-trained authentication model employs a hierarchical processing logic. Unobstructed facial local feature vectors, voiceprint feature vectors, gait feature vectors, and NFC feature vectors are each processed through independent feature encoding layers, transforming features of different dimensions into vectors of uniform length. These vectors are then combined into a comprehensive feature vector through a feature concatenation layer. During concatenation, the weights of each feature vector are dynamically adjusted according to the scenario. For example, when the user's face is severely obscured, the weights of voiceprint and gait features are increased; when the user presents an NFC device, the weights of the NFC feature are increased, ensuring that the model can still accurately output identity matching scores in complex scenarios.

[0030] The feature weighting of the pre-trained stress state detection model is based on the reliability of the features: micro-expression feature vectors, reflecting subtle facial movements, have relatively high weights; voice tension feature vectors, such as pitch fluctuations and pause frequencies, and physiological feature vectors, such as heart rate fluctuations and skin conductivity, serve as auxiliary features with slightly lower weights. The weighting process is implemented through fully connected layers, with in-layer parameters determined based on stress sample data. This accurately captures the collaborative patterns of micro-expression abnormalities, voice tension, and physiological fluctuations, and the output stress risk score can effectively distinguish between natural and real stress states.

[0031] NFC feature vector processing needs to be compatible with multiple devices: the identification information of Apple mobile phone transit cards is directly used as the feature vector input after being encrypted and transmitted; the simulated identification of Android virtual access cards needs to be verified for signature legality before being converted into feature vectors; the identification information of smart bracelets and watches needs to be matched with the encryption rules corresponding to their device models to ensure that only the feature vectors of authorized devices can be recognized by the model.

[0032] This solution further proposes that the identity matching score is calculated based on the preset weights of the unobstructed facial features, the voiceprint features, the gait features, and the NFC features; and the coercion risk score is calculated based on the preset weights of the micro-expression features, the voice tension features, and the physiological features.

[0033] The preset weighting of identity matching scores is based on the stability and uniqueness of features. Unobstructed facial features, containing rich identity information, have the highest weight and can effectively distinguish different individuals; voiceprint features, being difficult to imitate, have the next highest weight and are particularly suitable for scenarios with occluded faces; gait features, as dynamic features, have a lower weight but can compensate for the shortcomings of static features; NFC features, as actively presented identifiers, have a moderate weight and can quickly improve verification efficiency. This weighting allocation ensures the core role of biometrics while enhancing convenience through NFC features.

[0034] The pre-defined weighting of the coercion risk score takes into account the concealment of features. Micro-expression features are difficult to actively control and have the highest weight, as subtle facial muscle contractions can still be captured even if the user deliberately conceals them; voice tension features are next, as changes in a user's voice under coercion are difficult to completely hide; physiological features have a weight comparable to voice tension features, as signals such as heart rate and skin conductance are not subject to subjective control and can serve as objective evidence of a coercive state.

[0035] The weights are not fixed. The system supports updating model parameters via the cloud. When it is found that the reliability of a certain type of feature decreases in a specific scenario, such as when users wear masks more often in winter, which reduces the usability of facial features, the weights can be adjusted remotely to increase the proportion of voiceprint and gait features, ensuring that the model always adapts to actual needs.

[0036] This solution further proposes that the operation of the execution output module to open the access control includes: when the identity matching score reaches a first preset threshold and the coercion risk score is lower than a second preset threshold, controlling the access control lock to open and triggering a welcome prompt; the operation of generating verification instructions includes: when the identity matching score is in a first preset range and the coercion risk score is lower than a second preset threshold, generating verification information and sending it to the user registration terminal; the operation of triggering an early warning includes: when the identity matching score is in a second preset range and the coercion risk score reaches a third preset threshold, sending an early warning message to the monitoring center and triggering the access control to display a confusing prompt; the operation of denying passage includes: when the identity matching score is lower than a third preset threshold and the coercion risk score is lower than a second preset threshold, displaying a verification failure prompt and recording an abnormal log; when the identity matching score is lower than a third preset threshold and the coercion risk score reaches a third preset threshold, simultaneously triggering an alarm and a monitoring center early warning.

[0037] The operation logic of the output module covers verification needs across all scenarios. When the identity matching score reaches the first preset threshold and the coercion risk score is lower than the second preset threshold, it is determined to be normal access. The access control lock control unit immediately unlocks the door, and at the same time, the prompt unit plays a preset welcome voice message, which supports user-customized content. The entire process requires no additional user operation.

[0038] When the identity matching score falls within the first preset range and the coercion risk score is below the second preset threshold, the system determines the identity to be verified, generates a verification command, and sends it to the user's registration terminal, such as a mobile app, via a communication unit. The verification command can be in the form of a dynamic verification code. After the user enters the correct code, the command is synchronized to the execution output module via the cloud, triggering the access control to open. If the user does not respond in time, the command automatically expires after 5 minutes to ensure security.

[0039] When the identity matching score is in the second preset range and the coercion risk score reaches the third preset threshold, the system determines that there may be coercion. The communication unit immediately sends an early warning message to the monitoring center, including real-time images, time, and location. At the same time, the access control screen displays a confusing prompt, such as "System failure, please try again later," to avoid provoking the coercer and to buy time for the monitoring center to intervene.

[0040] When the identity matching score is lower than the third preset threshold and the coercion risk score is lower than the second preset threshold, it is determined that the identity does not match. The prompt unit displays the verification failure information and records the abnormal log, including the collected features and time. The log is uploaded to the cloud periodically for administrator auditing.

[0041] When the identity matching score is lower than the third preset threshold and the coercion risk score reaches the third preset threshold, it is determined to be a combination of abnormality and coercion. The execution output module immediately triggers an audible and visual alarm and sends an emergency warning to the monitoring center. The access control remains locked to prevent unauthorized entry.

[0042] The execution of remote door opening commands requires additional verification. After the user initiates a request through the mobile app, the cloud server first verifies the user's permissions, then encrypts and transmits the command to the execution output module. After the module verifies that the command signature is correct, it controls the door access to open according to the above logic. The entire process is completed within 10 seconds.

[0043] Temporary access control is implemented in this stage as follows: when a visitor uses a temporary authorization code, the output module combines the authorization code with the identity matching score for judgment. If the authorization code is valid and the score meets the requirements, the access control can be triggered to open. The authorization code will automatically expire after it expires.

[0044] This solution further proposes that the intelligent access control system includes a near-field communication reading device, which is set in the multimodal sensor group to sense the signal of the user's near-field communication device and read its identification information. After being processed by an encryption algorithm, a near-field communication feature vector is generated and input into the pre-trained authentication model.

[0045] The near-field communication (NFC) reader is integrated into a multi-modal sensor array and achieves multi-device compatibility through a built-in NFC module. Its workflow is as follows: when a user brings the device close to the sensing area, the reader automatically activates, emits a specific frequency radio frequency signal, wakes up the NFC chip within the device, and establishes communication.

[0046] For transit cards on Apple phones, the reader directly reads the transit card's unique identification information by adapting to Apple's NFC protocol. This information is then encrypted using a hash algorithm to generate an NFC feature vector, preventing the original identification from being leaked.

[0047] For virtual access cards on Android phones, the device reads and verifies the digital signature of the virtual card to ensure it comes from a legitimate app. Then, it extracts the identification information from the card, encrypts it, and generates a feature vector to prevent counterfeit virtual cards from being identified.

[0048] For smart bracelets and watches, the reading device adjusts communication parameters according to the bracelet model, such as sensing distance and signal strength, and reads the built-in access control identifier. After encryption, the identifier is input into the pre-trained identity verification model.

[0049] The encryption algorithm uses a dynamic key mechanism: the communication key between the reading device and the data processing module is automatically updated every 24 hours, and the key information is synchronized through the cloud to ensure that the NFC feature vector cannot be cracked during transmission. It also supports device whitelist management, and users can add frequently used devices to the whitelist through the APP to improve recognition speed.

[0050] This solution further proposes that the pre-trained identity verification model and the pre-trained stress state detection model built into the model judgment module are both trained neural network models. The pre-trained identity verification model is trained using labeled identity matching sample data, and the pre-trained stress state detection model is trained using labeled stress state sample data.

[0051] Both the pre-trained authentication model and the pre-trained stress state detection model are based on deep learning neural network models, and the training process emphasizes data diversity and scenario coverage.

[0052] The training data for the pre-trained identity verification model includes facial, voiceprint, gait, and NFC tag samples from different groups in various scenarios, with each sample labeled with real identity information. During training, deep representations of each feature are first extracted through a basic network, and then the discriminative power of the features is optimized through comparative learning, enabling the model to accurately match different features of the same person while distinguishing similar features between different people.

[0053] The training data for the pre-trained stress state detection model includes micro-expressions, vocal tension, and physiological feature samples under both normal and stress states. The stress samples cover user data with different stress types and response patterns. During training, a multi-label classification strategy is employed, enabling the model to recognize mixed states. The output stress risk score quantifies the likelihood of stress. The model periodically receives new sample data from the cloud for incremental training, continuously improving its ability to identify novel stress scenarios.

[0054] Both types of models are deployed on local edge computing units, supporting real-time inference, while retaining the model parameter update interface, enabling remote upgrades via incremental model packages pushed from the cloud, without the need for on-site maintenance.

[0055] This solution further proposes that the execution output module includes an access control lock control unit, a prompting unit, and a communication unit. The access control lock control unit is used to control the opening or closing of the access control lock, the prompting unit is used to output prompting information, and the communication unit is used to communicate with the monitoring center or user terminal.

[0056] The three units of the execution output module work together to realize access control and interaction functions. The access control lock control unit is connected to the electromagnetic lock or motor lock through a relay. After receiving the control command, it first verifies the legality of the command, whether it comes from the model judgment module or an authorized remote request, and then drives the lock to perform the opening or closing operation. During the operation, the lock status is fed back to the system in real time.

[0057] The prompting unit includes a voice module and a display screen. The voice module allows users to customize prompts via a mobile app: users record or upload voice files, which are then reviewed in the cloud and synchronized to the module. The module then plays the corresponding voice file according to different scenarios, and the volume can be adjusted via the app. The display screen outputs text prompts, complementing the voice prompts and suitable for users with hearing impairments or in noisy environments.

[0058] The communication unit is the core of the remote function. It establishes a TCP / IP connection with the cloud server via 4G and 5G modules and uses the MQTT protocol for data transmission. When remotely opening the door, the user's APP command is encrypted by the cloud and sent to the communication unit. After decryption, the unit verifies the user ID, timestamp, and other information in the command. Once confirmed, it sends an opening command to the access control lock control unit and feeds the operation result back to the cloud, which is then synchronously updated to the user's APP.

[0059] Temporary permission management is achieved through communication unit and cloud collaboration: users set the permission time or number of uses for visitors in the APP, the cloud generates a unique temporary authorization code, and synchronizes it to the local cache of the execution output module; when a visitor uses the device, the module verifies the validity of the authorization code, performs the operation based on the identity verification result, and automatically clears the cache after the permission expires, ensuring that temporary permissions are controllable.

[0060] The communication unit also supports log upload functionality, which periodically uploads access control operation records to the cloud. Users can view historical records through the APP, enabling full traceability of access control usage.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation methods of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. An intelligent access control system, characterized in that, It includes a multimodal sensor group, a data processing module, a model judgment module, and an execution output module; the multimodal sensor group is used to collect multi-dimensional data when users pass through; the data processing module is used to extract and preprocess the multi-dimensional data to generate structured features; the model judgment module has built-in pre-trained identity verification model and pre-trained coercion state detection model, which are used to analyze the structured features and output identity matching score and coercion risk score, respectively. The execution output module performs access control operations based on the identity matching score and the coercion risk score.

2. The intelligent access control system according to claim 1, characterized in that, The multimodal sensor group includes a visible light imaging device, an infrared imaging device, an acoustic acquisition device, a motion sensing device, and a physiological signal acquisition device. The visible light imaging device is used to acquire visible light images of the user's face, the infrared imaging device is used to acquire infrared thermal images of the user's face, the acoustic acquisition device is used to acquire the user's voice signals, the motion sensing device is used to acquire the user's motion trajectory data, and the physiological signal acquisition device is used to acquire the user's physiological state signals. The data processing module performs registration processing on the visible light images and the infrared images, identifies occluded areas, and extracts local features of the unoccluded face. The speech signal is denoised and voiceprint features are extracted. The motion trajectory data is analyzed to extract gait features; The physiological state signal is filtered and physiological features are extracted; the model judgment module inputs the unobstructed facial local features, the voiceprint features and the gait features into the pre-trained identity verification model and outputs the identity matching score; the micro-expression features corresponding to the unobstructed facial local features, the voice tension features corresponding to the voiceprint features and the physiological features are input into the pre-trained stress state detection model and output the stress risk score; The execution output module performs operations such as opening the access control, generating a verification command, triggering an early warning, or denying passage based on the combination result of the identity matching score and the coercion risk score.

3. The intelligent access control system according to claim 1, characterized in that, The image acquisition device in the multimodal sensor group has a resolution and frame rate suitable for acquiring user facial image information, the acoustic acquisition device has a sampling rate suitable for acquiring user voice signals, the motion sensing device has a preset detection range and angular resolution, and the physiological signal acquisition device transmits encrypted data to the data processing module via wireless communication.

4. The intelligent access control system according to claim 1, characterized in that, When the data processing module performs registration processing on the visible light image and the infrared image, it uses a feature point matching algorithm and a feature extraction and mismatch point removal algorithm to ensure that the superposition error is within a preset range; when performing noise reduction processing on the speech signal, it uses an environmental noise suppression processing algorithm; when performing trajectory analysis on the motion trajectory data, it uses a feature extraction algorithm to obtain gait-related parameters. When filtering the physiological state signal, a filtering algorithm is used to extract physiological feature parameters.

5. The intelligent access control system according to claim 1, characterized in that, The input to the pre-trained identity verification model is the unobstructed facial local feature vector, the voiceprint feature vector, the gait feature vector, and the NFC feature vector. After feature concatenation, the identity matching score is output. The input to the pre-trained coercion state detection model is the micro-expression feature vector, the voice tension feature vector, and the physiological feature vector. After feature weighting, the coercion risk score is output.

6. The intelligent access control system according to claim 1, characterized in that, The identity matching score is calculated based on preset weights of the unobstructed facial features, the voiceprint features, the gait features, and the NFC features; the coercion risk score is calculated based on preset weights of the micro-expression features, the voice tension features, and the physiological features.

7. The intelligent access control system according to claim 1, characterized in that, The execution output module performs the following operations to open the access control system: when the identity matching score reaches a first preset threshold and the coercion risk score is lower than a second preset threshold, it controls the access control lock to open and triggers a welcome prompt; it performs the following operations to generate verification instructions: when the identity matching score is within a first preset range and the coercion risk score is lower than a second preset threshold, it generates verification information and sends it to the user registration terminal; it performs the following operations to trigger an early warning: when the identity matching score is within a second preset range and the coercion risk score reaches a third preset threshold, it sends an early warning message to the monitoring center and triggers the access control system to display a confusing prompt; it performs the following operations to refuse passage: when the identity matching score is lower than a third preset threshold and the coercion risk score is lower than a second preset threshold, it displays a verification failure prompt and records an abnormal log; when the identity matching score is lower than a third preset threshold and the coercion risk score reaches a third preset threshold, it simultaneously triggers an alarm and a monitoring center early warning.

8. The intelligent access control system according to claim 1, characterized in that, The intelligent access control system includes a near-field communication reading device, which is set in the multimodal sensor group to sense the signal of the user's near-field communication device and read its identification information. After being processed by an encryption algorithm, a near-field communication feature vector is generated and input into the pre-trained authentication model.

9. The intelligent access control system according to claim 1, characterized in that, The pre-trained identity verification model and the pre-trained stress state detection model built into the model judgment module are both trained neural network models. The pre-trained identity verification model is trained using labeled identity matching sample data, and the pre-trained stress state detection model is trained using labeled stress state sample data.

10. The intelligent access control system according to claim 1, characterized in that, The execution output module includes an access control lock control unit, a prompting unit, and a communication unit. The access control lock control unit is used to control the opening or closing of the access control lock. The prompting unit is used to output prompt information. The communication unit is used to communicate with the monitoring center or user terminal.

Images