Elevator passenger authority control system based on face recognition

Abstract

This invention relates to the interdisciplinary field of intelligent buildings and facial recognition technology, specifically disclosing an elevator access control system based on facial recognition. The system includes a facial image acquisition device, a remote photoplethysmography (PPG) signal extraction device, a liveness detection engine, a facial recognition comparison module, an access control decision unit, and an elevator control actuator. It extracts weak color change signals related to heartbeats from facial video streams, combines this with an adversarial generative network (GAN) model to determine the authenticity of the liveness detection. After confirming a liveness detection, it performs facial recognition and access control matching, dynamically controlling elevator door opening and closing and target floor registration. This invention, by integrating remote PPG technology with a multi-scale spatiotemporal consistency constraint liveness detection mechanism, can resist highly realistic deception attacks without requiring active user cooperation, improving security, convenience, and real-time system response.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of intelligent buildings and facial recognition, specifically relating to an elevator access control system based on facial recognition. Background Technology

[0002] With the continuous evolution of smart building and vertical transportation management technologies, biometric-based access control schemes have become a core component of modern building security management. In elevator scenarios with high population flow, automating access control to specific floors through high-precision identity authentication not only enhances the building's security level but also optimizes traffic flow efficiency and reduces manual management costs.

[0003] Facial recognition-based elevator access control technology, due to its contactless and convenient nature, has gradually replaced traditional physical identification methods. This technology typically uses visual acquisition terminals deployed in the elevator car or landing doors to capture and compare the facial features of passengers in real time. Combined with authorization information from a backend database, it dynamically triggers elevator operation commands. To ensure the authenticity of the authentication process, the system often integrates liveness detection logic to identify whether the collected information originates from a real biological individual, preventing unauthorized users from gaining elevator access through demonstration attacks.

[0004] Existing liveness detection mechanisms still face challenges in complex application environments. Traditional methods based on texture analysis or depth estimation struggle to identify highly realistic facial images constructed using generative adversarial networks and high-definition video playback, making them vulnerable to deception attacks from photos, screen media, or 3D simulated masks, leading to system security vulnerabilities. Interactive liveness detection methods relying on command actions require users to perform specific cooperative actions, increasing waiting time and reducing the elevator experience. Due to a lack of perception of deep physiological characteristics such as hemodynamic changes, existing solutions often fail to achieve accurate permission determination when there are drastic changes in lighting or when users wear obstructions, making it difficult to balance security and real-time performance. A facial recognition-based elevator access control system is desired. Summary of the Invention

[0005] The purpose of this invention is to provide an elevator access control system based on facial recognition, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The elevator access control system based on facial recognition includes a facial image acquisition device, a remote photoplethysmography (PPG) signal extraction device, a liveness detection engine, a facial recognition comparison module, an access control decision unit, and an elevator control actuator. The facial image acquisition device is configured to capture facial video streams of passengers in real time in the elevator hall door or car entrance area, and transmit the acquired continuous frame image data to the remote PPG signal extraction device and the facial recognition comparison module. The remote photoplethysmography (PPG) pulse wave signal extraction device is configured to analyze the subtle color fluctuations in the facial skin region caused by the periodic changes in blood volume from the facial video stream output by the facial image acquisition device, and then extract the physiological signal sequence reflecting the heart rhythm. The liveness detection engine is configured to fuse the physiological signal sequence output by the remote photoplethysmography pulse wave signal extraction device with the original image features output by the face image acquisition device. It uses a built-in adversarial generative network model to determine the authenticity of the input data in order to identify whether there are deceptive attack behaviors such as photos, high-definition video replays, or 3D simulation masks. The face recognition comparison module is configured to perform structured analysis on the facial features captured by the face image acquisition device after the liveness detection engine confirms that the input is a real live person, and to match and compare them with the registered user face templates pre-stored in the authorized database. The permission decision unit is configured to generate a corresponding elevator access permission instruction based on the matching result of the face recognition comparison module and the preset floor access policy, and send the elevator access permission instruction to the elevator control actuator. The elevator control actuator is configured to receive instructions from the permission decision unit, control the opening and closing status of the elevator doors and the registration operation of the target floor, thereby realizing dynamic management of the access rights of elevator passengers.

[0007] Preferably, the remote photoplethysmography pulse wave signal extraction device adopts a non-contact optical sensing method. By performing frequency domain and time domain joint analysis on the time series of pixel intensity of the red, green and blue channels in the face video stream, it extracts pulse waveform features with stable periodicity. These pulse waveform features can reflect an individual's heart rate and hemodynamic state under normal lighting conditions.

[0008] Furthermore, the liveness detection engine is embedded with an adversarial generative network model trained on a large number of real human faces and various deception samples. This adversarial generative network model can learn the dynamic response patterns of real biological tissues at the microcirculation level and distinguish non-physiological signals formed by static images, screen playback, or silicone material simulation.

[0009] Furthermore, the adversarial generative network model incorporates a multi-scale spatiotemporal consistency constraint mechanism during training, enabling the model to not only focus on the texture realism of a single frame image, but also capture subtle color variation patterns caused by real blood flow across frames, thereby improving its robustness in recognizing highly realistic deception techniques.

[0010] Preferably, the face recognition comparison module adopts a deep convolutional neural network architecture, in which its feature extraction layer shares some of the underlying visual representations with the liveness detection engine, so as to reduce redundant calculations and enhance the overall real-time processing capability of the system.

[0011] Furthermore, the permission decision unit stores a floor access whitelist that is bound to the identity of each registered user. When the face recognition comparison module confirms the identity and the liveness detection engine verifies that it is a real individual, the permission decision unit only opens the floor button function authorized for that registered user, while the other floor buttons remain locked.

[0012] Furthermore, the face image acquisition device is equipped with an adaptive lighting component, which can automatically adjust the light source intensity and color temperature when the ambient light is insufficient or too strong, to ensure that the acquired facial video stream meets the minimum signal-to-noise ratio requirements for remote photoplethysmography signal extraction.

[0013] Preferably, the system further includes an abnormal behavior recording module, configured to automatically capture the current video segment and encrypt and store it together with the timestamp and device location information when the liveness detection engine determines that it is a deception attack multiple times in a row, for subsequent security auditing.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. The elevator access control system based on facial recognition provided by this invention deeply integrates remote photoplethysmography (PPG) technology from the field of medical engineering with generative adversarial networks (GANs) from computer vision to construct a passive liveness detection mechanism that does not require active user cooperation. This passive liveness detection mechanism can extract microvascular blood flow signals driven by real heartbeats from video streams captured by ordinary cameras, avoiding the dependence on explicit action commands such as blinking and opening the mouth in traditional methods, thus improving the smoothness of the elevator riding process and the user experience.

[0015] 2. Because physiological signals are difficult to simulate with photos, videos, or 3D masks, this invention provides the system with strong resistance to various deception attacks, significantly enhancing the security boundary of identity authentication. Furthermore, by sharing underlying visual features and introducing multi-scale spatiotemporal consistency constraints, the system reduces computational overhead while ensuring high-precision liveness detection, meeting the stringent real-time response requirements of elevator scenarios.

[0016] 3. This invention achieves an organic balance between safety, convenience, and robustness, providing reliable technical support for the vertical traffic safety management of smart buildings. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the liveness detection method that integrates remote photoplethysmography (PPG) signals and original image features in this invention. Figure 3 This is a flowchart illustrating the logical process of facial video stream acquisition and heart rhythm physiological signal extraction in this invention. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the face recognition comparison module, the permission decision unit, and the elevator control actuator in this invention. Detailed Implementation

[0018] Example 1: Please refer to the appendix Figure 1 To be continued Figure 4 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0019] The elevator access control system based on facial recognition includes a facial image acquisition device, a remote photoplethysmography (PPG) signal extraction device, a liveness detection engine, a facial recognition comparison module, an access control decision unit, and an elevator control actuator.

[0020] The facial image acquisition device is used to capture real-time facial video streams of passengers in the elevator lobby or car entrance area, and transmits the acquired continuous frame image data to the remote photoplethysmography (PPG) signal extraction device and the facial recognition comparison module. The facial image acquisition device includes a high dynamic range (HDR) image sensor, an adaptive lighting component, and a high-performance video preprocessing chip. The HDR image sensor is configured to have a light sensitivity of at least 100 dB to ensure that even in strong sunlight or backlighting conditions facing the elevator lobby, it can clearly capture subtle color fluctuations in facial skin.

[0021] The adaptive lighting component includes an infrared sensor, a photoresistor, and an array of light-emitting diodes (LEDs). The infrared sensor detects human proximity signals and triggers system wake-up, while the photoresistor monitors the current ambient illuminance in real time. When the ambient illuminance is below a preset minimum lumen threshold, the adaptive lighting component automatically adjusts the output power and color temperature of the array of LEDs to emit visible or near-infrared light with a uniform spectral energy distribution, thereby providing high signal-to-noise ratio raw data for subsequent physiological signal extraction. The high-performance video preprocessing chip performs hardware-level automatic white balance correction, distortion correction, and color space conversion on the raw image, ensuring high chromaticity consistency of the output video frames over time.

[0022] The remote photoplethysmography (PPG) signal extraction device is used to analyze the subtle color fluctuations in the facial skin region caused by periodic changes in blood volume from the facial video stream output by the facial image acquisition device, and then extract the physiological signal sequence reflecting the heart rhythm. The remote PPG signal extraction device integrates a region-of-interest (ROI) dynamic localization unit, a signal spatial dimensionality reduction unit, and a physiological frequency filtering unit.

[0023] The region of interest (ROI) dynamic localization unit is configured to use a facial keypoint detection algorithm to lock onto skin areas with dense capillary distribution, such as the forehead and cheeks, in real time, while excluding non-feature areas such as the eyes, mouth, and areas obscured by masks. The signal space dimensionality reduction unit is configured to perform blind source separation processing on the extracted multi-channel color information. Specifically, it maps the pixel intensity time series of the red, green, and blue channels to independent feature spaces, and removes interference signals caused by ambient light fluctuations by identifying the non-correlation between signals, thereby obtaining a pure volumetric pulse wave raw signal.

[0024] The physiological frequency filtering unit is embedded with a bandpass filter whose passband is limited to between 0.75 Hz and 4.0 Hz, corresponding to the human heart rate range of 45 to 240 beats per minute when at rest or during slight exercise. By filtering out high-frequency thermal noise and low-frequency motion artifacts outside the frequency band, it outputs a pulse characteristic waveform with stable periodicity.

[0025] The liveness detection engine is used to fuse the physiological signal sequence output by the remote photoplethysmography (PPG) signal extraction device with the original image features output by the face image acquisition device, and to determine the authenticity of the input data through a built-in generative adversarial network model in order to identify whether there are deceptive attacks such as photos, high-definition video replays, or 3D simulation masks.

[0026] The liveness detection engine includes a multi-source feature fusion layer, an adversarial generative network (GAN) computation unit, and a spatiotemporal consistency verification module. The multi-source feature fusion layer is configured to perform a Hilbert transform on a one-dimensional physiological signal sequence, converting it into a two-dimensional time-frequency spectrum, and then concatenate it with three-dimensional facial spatial texture features in the feature dimension to form a high-dimensional liveness detection vector. The GAN model consists of a generator and a discriminator. In operation, the generator attempts to reconstruct pseudo-physiological signals that conform to biophysical principles based on the input original image, while the discriminator is configured to distinguish between genuinely acquired physiological signals and signals simulated by the generator.

[0027] Through this adversarial learning, the discriminator can capture the dynamic nonlinear response patterns of real biological tissue at the microcirculation level, patterns that are completely absent or exhibit obvious digital artifacts in still photographs or videos played back on a liquid crystal display. The spatiotemporal consistency verification module is configured to perform consistency analysis on the signal phase between consecutive frames, ensuring that the extracted pulse waveform and the dynamic evolution of facial micro-expressions remain logically consistent along the timeline, preventing advanced spoofing attacks that tamper with the video stream through injection techniques.

[0028] The face recognition comparison module is used to perform structured analysis on the facial features captured by the face image acquisition device after the liveness detection engine confirms that the input is a real live subject, and to match and compare them with the registered user face templates pre-stored in the authorization database. The face recognition comparison module adopts a deep convolutional neural network architecture and integrates a feature extraction sharing unit, a high-dimensional vector comparison engine, and a distributed authorization database.

[0029] The feature extraction sharing unit is configured to share the underlying convolutional feature map with the liveness detection engine. This design allows the system to perform preliminary identity feature extraction while conducting liveness detection, reducing the processor's floating-point operations and lowering system power consumption. The high-dimensional vector comparison engine calculates the cosine similarity between the extracted 512-dimensional facial feature vector and the template vector in the database. When the score is greater than a preset security matching threshold, identity verification is considered successful. The distributed authorization database supports multi-index queries, enabling rapid retrieval of user data based on affiliated organization, functional department, or floor permission tags.

[0030] The permission decision unit is used to generate a corresponding elevator access permission instruction based on the matching result of the face recognition comparison module and a preset floor access policy, and send the elevator access permission instruction to the elevator control actuator. The permission decision unit includes a logic reasoning unit, a permission timing management module, and a policy memory. The policy memory stores the set of legally valid destination floors for each registered user, as well as the valid time interval of that permission.

[0031] Access to cleaning staff is configured to be limited to all floors only during their work hours, while access to ordinary residents is restricted to their designated residential floor. The logic inference unit is configured to perform Boolean logic operations based on the identity verification score, the liveness detection result, and the current system time. This logic inference unit will only generate a high-level trigger signal when the identity is valid, the liveness detection result is true, and the current time is within the authorized range. The access control timing management module is responsible for converting this high-level trigger signal into a digital message conforming to the elevator control bus protocol.

[0032] The elevator control actuator receives instructions from the access control unit to control the opening and closing of the elevator doors and the registration of the target floor, thereby achieving dynamic management of passenger access permissions. The elevator control actuator interacts with the elevator main control cabinet in real time via an industrial-grade serial communication interface or a controller area network bus. Upon receiving an access permission instruction, the elevator control actuator sends a target floor selection signal to the main control cabinet and illuminates the corresponding floor button on the control panel inside the elevator car. The actuator also includes a feedback monitoring unit, which acquires the elevator's current operating status, floor position, and door operator status in real time and feeds this information back to the access control unit to achieve closed-loop management of the control logic.

[0033] Furthermore, during processing, the remote photoplethysmography (PPG) signal extraction device performs spatial domain mean pooling on each frame of the image, aggregating selected skin region pixels into a single intensity value and establishing a time series across the red, green, and blue color dimensions. The device is configured to employ adaptive cancellation technology, calculating the global illuminance variation trend in non-skin regions and subtracting this global illuminance variation component from the skin region signal, thereby eliminating interference from indoor lighting flicker or light and shadow changes during elevator operation on physiological signal extraction.

[0034] Furthermore, the adversarial generative network model in the liveness detection engine incorporates structural similarity loss and perceptual loss functions during the training phase. The structural similarity loss function is configured to constrain the generated signal to maintain a high degree of morphological consistency with the real pulse wave, while the perceptual loss function utilizes a pre-trained deep network to extract high-level semantic features, ensuring the system can identify signal distortions caused by subtle differences in absorbance and scattering rates between the 3D mask material and real skin. Through this multi-dimensional constraint, the liveness detection engine exhibits extremely high robustness against highly realistic deception techniques such as silicone masks and ultra-thin latex masks.

[0035] The system also includes an abnormal behavior recording module. This module is configured to work in real-time with the liveness detection engine. When the liveness detection engine repeatedly determines that an input sample is a spoofing attack and the confidence level exceeds a preset alarm threshold, the abnormal behavior recording module automatically triggers a high-frequency sampling mode, capturing high-definition video clips 10 seconds before and after the attack, and encrypting and encapsulating the data using advanced encryption standard algorithms. This data, along with the attack timestamp, unique device identifier, geographic location coordinates, and identified forgery type tags, is uploaded to the smart building's management cloud via a secure transport layer protocol, providing strong support for subsequent security tracing and evidence chain preservation.

[0036] The access control decision unit also features multi-factor dynamic weight adjustment. During peak elevator usage hours, to ensure efficient passage, the unit is configured to appropriately reduce the sliding window length for liveness detection, achieving millisecond-level rapid response. At night or during sensitive periods, the unit automatically increases the accuracy level, requiring multiple consecutive sets of physiological signal sequences to meet liveness criteria before granting access. This dynamic adjustment mechanism enables the system to achieve an optimal balance between security levels and user experience.

[0037] The facial image acquisition device also features intelligent field-of-view adjustment, which automatically drives its internal micro-panel or digital zoom algorithm based on the detected face position, ensuring that the face's proportion in the image remains within a preset golden ratio range. This ensures that the number of facial skin pixels acquired is sufficient to support high-precision photoplethysmography (PPG) analysis, maintaining consistent recognition performance even when children or adults of varying heights are riding elevators.

[0038] The elevator control actuator employs optocoupler isolation technology at the hardware level, physically isolating its input and output terminals from the elevator's original circuitry to prevent potential damage to the recognition system from cable surges or electromagnetic interference. The actuator also incorporates a heartbeat detection circuit. If the facial recognition system or liveness detection engine malfunctions or experiences a software error, this circuit automatically switches to bypass mode, restoring the elevator's normal manual operation. This ensures that the elevator's safe and normal operation as a fundamental lifeline facility is not compromised under any extreme circumstances.

[0039] Example 2: Based on Example 1, this example provides an elevator access control system based on face recognition with edge computing and distributed collaborative architecture. This access control system is optimized for multi-elevator collaboration scenarios in large-scale commercial complexes, and focuses on enhancing the system's processing capacity under high concurrency requests and the redundancy of physiological signal extraction.

[0040] The system includes edge sensing terminals deployed on the elevator car side, local processing arrays deployed on floor shaft nodes, and a central computing cluster deployed in the building's computer room.

[0041] The edge-sensing terminal, whose core function corresponds to the face image acquisition device and permission execution feedback mechanism in Embodiment 1, integrates a binocular stereo vision camera array. The binocular stereo vision camera array consists of two infrared-enhanced sensors with identical parameters, maintaining a preset baseline distance between them. This configuration allows the edge-sensing terminal to generate corresponding high-precision depth map information while acquiring the facial video stream. The edge-sensing terminal is configured to synchronously encapsulate the original video stream and depth information stream, and transmit them to the local processing array via an internal high-speed bus.

[0042] The local processing array is responsible for performing the remote photoplethysmography (PPG) signal extraction and preliminary liveness filtering tasks described in Example 1. Each local processing array is configured to simultaneously manage the acquisition terminals of multiple elevators, and its internal architecture includes a multi-core parallel computing architecture. The signal processing module in the local processing array is configured to use a distributed spatiotemporal correlation algorithm to not only extract the physiological signals of a single face, but also eliminate the global motion vector caused by elevator car swaying through cross-camera joint calibration. After completing the feature extraction of the physiological signals locally, the local processing array only uploads the compressed physiological feature vectors and key facial feature point data to the central computing cluster, without uploading massive amounts of raw video data, thus alleviating the bandwidth pressure on the building's internal network.

[0043] The central computing cluster corresponds to the liveness detection engine, face recognition comparison module, and permission decision unit in Embodiment 1. Due to the use of a high-performance GPU server cluster, the central computing cluster is capable of running much larger-scale deep adversarial generative network models.

[0044] In this embodiment, the adversarial generative network model is configured in a multi-task parallel mode, capable of analyzing the blood perfusion distribution map of elevator passengers while performing liveness detection. This blood perfusion map includes not only heart rate information but also trends in blood oxygen saturation and subtle indicators of emotional arousal. The discriminator compares these multidimensional physiological indicators with historical authorized samples. If it finds a significant deviation between the current individual's physiological response pattern and their physiological baseline during registration with the system, the system automatically increases the security verification weight for this identification, triggering additional voiceprint verification or CAPTCHA interaction.

[0045] The access control unit in the central computing cluster is configured with predictive scheduling capabilities. By analyzing historical pedestrian flow data across the entire building and currently identified user identities, it anticipates pedestrian flow pressure on different floors and generates optimized elevator group control commands, which are then sent to the elevator control actuators. For example, when the system detects that the core management team of a high-rise company is continuously entering the elevator lobby, it automatically allocates the nearest empty elevator and pre-registers their frequently used floors, achieving a seamless and ultra-fast elevator experience.

[0046] The system also incorporates an environmental self-balancing calibration mechanism. Since the lighting conditions inside the elevator car change drastically as the elevator moves between different floors, the local processing array embeds a light environment simulator. This simulator uses real-time acquired ambient light parameters to generate a compensation function in the digital domain. This compensation function is deconvolved with the original signal output from the image acquisition device, thus pre-canceling the pseudo-pulse signal caused by ambient light fluctuations before the physiological signal is extracted. This allows the system to maintain continuous and accurate liveness detection capabilities even during frequent elevator door opening and closing.

[0047] In this embodiment, the remote photoplethysmography (PPG) signal extraction device further integrates adaptive frequency tracking technology. Considering the significant fluctuations in heart rate among individuals under stress, after exercise, or at rest, this technology is configured to estimate the dominant frequency component of the signal in real time within a sliding window using short-time Fourier transform. When a sudden change in the dominant frequency is detected, the system automatically adjusts the filter bandwidth and increases the sampling weight in that frequency band, thereby ensuring accurate identification of the true heart rate even under extreme physiological conditions and avoiding misjudgment of a living individual due to sudden heart rate changes.

[0048] The liveness detection engine also incorporates an adversarial training strategy based on synthetic data. By simulating the interaction of millions of facial skin models with light sources of different wavelengths in a virtual environment, a virtual rPPG dataset is generated, containing various pathological features, makeup occlusions, and special lighting conditions. With the assistance of this massive amount of synthetic data, the adversarial generative network model possesses stronger generalization capabilities and can recognize highly realistic biological masks made from novel polymer materials that have never been seen before.

[0049] In this embodiment, the facial recognition comparison module employs a federated learning framework. While protecting user privacy, recognition nodes deployed in different buildings or different areas within the same group can collaboratively train and optimize the facial feature extraction model without exchanging underlying image data. This ensures that the system maintains extremely high recognition accuracy and security even when facing groups of people from different regions and ethnicities riding the elevator.

[0050] The permission decision-making unit also includes a multi-level policy arbitrator. After receiving feedback from the liveness detection engine and the facial recognition module, the arbitrator dynamically adjudicates based on the security level of the current floor. For example, on floors leading to the vault or core server room, the arbitrator is configured to require a signal-to-noise ratio of physiological signals greater than 15 dB and a heartbeat cycle consistency probability greater than 98% before issuing an operation command; while for floors leading to the cafeteria or public activity areas, a more lenient discrimination standard is adopted to ensure rapid passage in high-traffic scenarios.

[0051] The elevator control actuator in this embodiment features dual redundancy and hot backup. Internally, it contains two completely independent logic processing circuits and communication interfaces, with the two circuits maintaining state synchronization via synchronization pulses. In the event of a hardware failure in the main control circuit, the backup circuit can seamlessly take over control within 10 microseconds and send a self-check and maintenance request to the management backend, ensuring that the access control of the building's vertical transportation system never goes offline.

[0052] The abnormal behavior recording module in this embodiment incorporates blockchain traceability technology. When an unauthorized intrusion attempt or liveness detection alarm occurs, the system not only encrypts and stores the evidence locally, but also packages and stores key metadata, including the alarm event hash value, processing result, and decision evidence chain, on the building's private blockchain. This tamper-proof recording mechanism provides the most credible electronic credentials for handling subsequent legal disputes or security audits.

[0053] In this embodiment, the adaptive illumination component is refined into a multispectral light source system. This system comprises three independently controlled light sources in three bands: red, green, and near-infrared. The remote photoplethysmography (PPG) signal extraction device is configured to dynamically combine the illumination intensity of these three bands based on the reflectivity of users with different skin tones. For example, for skin with high melanin content, the system automatically increases the proportion of near-infrared illumination, utilizing the penetrating power of near-infrared light into deep tissues to extract blood flow signals. This overcomes the technical bottleneck of decreased physiological signal extraction accuracy in dark-skinned individuals using traditional visible light imaging.

[0054] The system also possesses self-evolving anomaly detection capabilities. By accumulating long-term norms of physiological signal distribution among normal users within a building, the liveness detection engine can automatically learn the "statistical fingerprint" of real biological signals in that specific environment. When an attack sample, despite being morphologically perfect, exhibits physiological characteristics that significantly deviate from the statistical patterns of the population, the system identifies it as a potential deepfake attack and adds it to a blacklist for continuous monitoring. This defense strategy based on statistical anomalies provides a forward-looking technical means to address potential future dynamic video spoofing techniques based on generative adversarial networks.

[0055] In summary, this embodiment achieves ultimate security management of elevator access permissions in complex environments by leveraging deep collaboration between the edge and cloud, utilizing more refined multispectral sensing and distributed computing power. While ensuring recognition accuracy, it also enhances the system's adaptability to extreme environments and high-concurrency scenarios.

[0056] Example 3: Based on the above examples, this example further discloses an elevator access control system based on face recognition that integrates mobile terminal linkage and multi-biometric feature fusion, aiming to solve the problem of access determination for users wearing heavily obstructed objects or in extremely low light conditions.

[0057] In addition to all the core components described in the previous embodiments, the system also includes a wireless near-field sensing module and a voiceprint texture analysis unit.

[0058] The wireless near-field sensing module is configured to detect authorized mobile terminals in the elevator waiting area in real time via Bluetooth Low Energy protocol or ultra-wideband positioning technology. When the face image acquisition device cannot acquire complete facial features due to visual occlusion, the permission decision unit automatically triggers multi-factor association logic. The logic is configured to: obtain the device identity identifier reported by the wireless near-field sensing module, and use it as an auxiliary index to retrieve the corresponding user's historical physiological baseline in the database.

[0059] The voiceprint texture analysis unit is configured to acquire the user's voice signal through a pickup array deployed in the elevator lobby door. The voiceprint recognition model integrated within the system extracts acoustic features such as spectral envelope and Mel-frequency cepstral coefficients from the voice and compares them with voiceprint templates stored in the database. Simultaneously, the remote photoplethysmography (PPG) signal extraction device does not cease operation but is configured to enter a "micro-exposed area scanning mode," specifically targeting the user's potentially exposed skin on their hands, behind their ears, or on their neck for high-magnification digital amplification and signal capture.

[0060] In this embodiment, the liveness detection engine is configured in a cross-modal fusion discrimination mode. It not only processes the physiological pulse signal from the visual side but also performs cross-modal consistency checks on the extracted heart rhythm and the fundamental frequency fluctuations and breathing rhythm in the voiceprint. According to bioacoustic principles, there is a complex physiological coupling relationship between the fluctuations in breathing and the heartbeat when a person speaks. By analyzing this cross-modal coherence, the discriminator can fundamentally identify artificially synthesized speech or recordings played through a speaker. Even if more than 80% of the user's face is obscured, the system can still provide a highly confident permission determination result through a triple dimension of "micro-skin region rPPG plus voiceprint liveness detection plus near-field device authentication."

[0061] Upon receiving a judgment command for such a complex scenario, the elevator control actuator will display a friendly reminder on the interactive screen inside the elevator car, informing the user that multi-factor authentication has been successfully completed, and automatically opening the door to the target floor. This design expands the system's applicability, enabling it to provide unbiased and safe passage services even in specialized workplaces or medical institutions.

[0062] The remote photoplethysmography (PPG) signal extraction device employs a super-resolution reconstruction algorithm when processing such minute skin areas. This algorithm utilizes a deep neural network to enhance texture and restore details in low signal-to-noise ratio images of minute skin, amplifying blurred color variations into clear waveform curves. Through this digital signal enhancement method, the system can maintain the continuity of physiological signals even under extremely unfavorable acquisition conditions, ensuring that the determination time for liveness detection does not increase as the acquisition area shrinks.

[0063] The system also includes environmental temperature and humidity compensation logic. The permission decision unit is configured to connect to the building's automated management system to obtain the current indoor and outdoor temperature and humidity. This is because ambient temperature affects the microcirculation of human skin. Based on temperature parameters, the system automatically adjusts the gain coefficient of the physiological signal extraction device and the sensitivity threshold of the liveness detection engine, eliminating systematic deviations in recognition accuracy caused by seasonal changes.

[0064] In this embodiment, the face recognition comparison module also integrates an attention mechanism network. This attention mechanism network is configured to automatically identify occluded parts of the face and concentrate computational weights on the unoccluded feature regions. Through this localized robust feature matching, the system can maintain a false recognition rate almost identical to that of full-face recognition when dealing with common scenarios involving wearing masks or sunglasses.

[0065] The system also features an emergency mode self-trigger function. When a violent fall, violent conflict, or a user's prolonged abnormal heart rate is detected inside the elevator, the access control unit immediately bypasses the regular floor restriction commands, forcing the elevator to the nearest emergency floor or the ground floor lobby. It automatically connects to the property management alarm center and emergency medical services, and simultaneously sends the user's identity information and a snapshot of their current physiological indicators to the rescue terminal. This function elevates the access control system from a simple security tool to a core component of a smart building system with life safety monitoring capabilities.

[0066] The system in this embodiment supports remote online evolution. All liveness detection models and facial feature extraction algorithms support incremental updates. Whenever a new liveness attack method emerges globally, the cloud-based defense center distributes the latest adversarial features to each elevator edge. The system's internal self-updating module is configured to automatically perform cold start switching of models during low-power periods, ensuring that the system always has the technical reserves to defend against the most cutting-edge security threats.

[0067] In summary, this embodiment constructs a robust elevator access control platform that is available in all times and all scenarios through deep coupling of multimodal physiological characteristics and enhanced adaptability in complex environments, achieving a high degree of unity between safety, reliability and humanistic care.

[0068] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An elevator access control system based on facial recognition, characterized in that, It includes a face image acquisition device, a remote photoplethysmography (PPG) pulse wave signal extraction device, a liveness detection engine, a face recognition comparison module, an access control decision unit, and an elevator control actuator; The face image acquisition device is configured to capture facial video streams of passengers in the elevator entrance area and transmit the image data to the remote photoplethysmography signal extraction device and the face recognition comparison module. The remote photoplethysmography (PPG) signal extraction device is configured to analyze color fluctuations in the facial skin region from the facial video stream and extract physiological signal sequences reflecting heart rhythm. The liveness detection engine is configured to fuse the physiological signal sequence with the original image features and identify, through an adversarial generative network model, whether there are deceptive attack behaviors such as photo and video replays of three-dimensional simulation masks. The face recognition comparison module is configured to perform structured parsing of the facial features and match them with a pre-stored face template after confirming that the input is a real live person. The permission decision unit is configured to generate elevator access permission instructions based on the matching results and preset strategies, and send them to the elevator control actuator. The elevator control actuator is configured to receive the elevator access permission command and control the opening and closing status of the elevator doors and the registration operation of the target floor.

2. The elevator access control system based on facial recognition according to claim 1, characterized in that, The face image acquisition device includes a high dynamic range image sensor, an adaptive fill light component, and a video preprocessing chip. The high dynamic range image sensor is configured to have a light sensitivity tolerance above a preset decibel threshold, and is used to capture color fluctuation data of facial skin in strong sunlight or backlight environments. The adaptive lighting component includes an infrared sensor, a photoresistor, and an array of light-emitting diodes; The infrared sensor is used to detect signals from the human body and trigger the system to wake up; the photoresistor is used to monitor ambient illuminance in real time. When the ambient illuminance is lower than the preset lumen threshold, the adaptive fill light component adjusts the output power and color temperature of the array of light-emitting diodes to emit fill light with uniform spectral energy distribution to the facial area; The video preprocessing chip is used to perform automatic white balance correction, distortion correction, and color space conversion on the original image to ensure that the output video frames have color consistency in the time dimension.

3. The elevator access control system based on facial recognition according to claim 1, characterized in that, The remote photoplethysmography pulse wave signal extraction device integrates a region of interest dynamic positioning unit, a signal space dimensionality reduction unit, and a physiological frequency filtering unit. The region of interest dynamic localization unit is configured to use a facial key point detection algorithm to lock the skin area of ​​the forehead and cheeks, and exclude non-feature areas such as the eyes, mouth and areas covered by occlusions; The signal space dimensionality reduction unit is configured to perform blind source separation processing on the extracted multi-channel color information, map the pixel intensity time series of the red, green and blue channels to independent feature spaces, and remove the interference signals generated by ambient light fluctuations by finding the non-correlation between signals to obtain the pure volumetric pulse wave original signal. The physiological frequency filtering unit has an embedded bandpass filter whose passband is limited to the preset heart rate frequency range under human rest and slight exercise conditions. It is used to filter out high-frequency thermal noise and low-frequency motion artifacts outside the frequency band and output a pulse characteristic waveform with stable periodicity.

4. The elevator access control system based on facial recognition according to claim 1, characterized in that, The liveness detection engine includes a multi-source feature fusion layer, an adversarial generative adversarial network computing unit, and a spatiotemporal consistency verification module. The multi-source feature fusion layer is configured to convert a one-dimensional physiological signal sequence into a two-dimensional time-frequency spectrum through Hilbert transform, and then concatenate and splice it with three-dimensional facial spatial texture features in the feature dimension to form a high-dimensional liveness detection vector; the adversarial generative network model consists of a generator and a discriminator. The generator is used to reconstruct pseudo-physiological signals that conform to the laws of biophysics based on the input original image, while the discriminator is configured to distinguish between real physiological signals and signals simulated by the generator, and to capture the dynamic nonlinear response patterns of real biological tissues at the microcirculation level through adversarial learning. The spatiotemporal consistency verification module is configured to perform consistency analysis on the signal phase between consecutive frames to ensure that the extracted pulse waveform and the dynamic evolution of facial micro-expressions remain logically consistent on the time axis.

5. The elevator access control system based on facial recognition according to claim 1, characterized in that, The face recognition comparison module includes a feature extraction sharing unit, a high-dimensional vector comparison engine, and a distributed authorization database; The feature extraction sharing unit is configured to share the underlying convolutional feature map with the liveness detection engine, so as to complete identity feature extraction while performing liveness detection. The high-dimensional vector comparison engine is configured to perform cosine similarity calculation between the extracted facial feature vectors of a preset dimension and the template vectors in the database, and determine that the identity verification is successful when the score is greater than the preset security matching threshold. The distributed authorization database supports multi-index queries and is used to quickly retrieve user data based on the organization, functional department, or floor permission tags.

6. The elevator access control system based on facial recognition according to claim 1, characterized in that, The permission decision-making unit includes a logic reasoning unit, a permission timing management module, and a policy memory. The policy memory stores the set of legally valid floors corresponding to each registered user, as well as the valid interval of that permission on the time axis. The logic reasoning unit is configured to perform Boolean logic operations on the comprehensive identity verification score, the liveness determination result and the current system time, and generate a high-level trigger signal only when the identity is valid, the liveness determination result is true and the current time is within the authorized interval. The permission timing management module is responsible for converting the high-level trigger signal into a digital message that conforms to the elevator control bus protocol. The permission decision unit also has a multi-factor dynamic weight adjustment function, which is configured to reduce the sliding window length for liveness detection during peak elevator usage periods, and increase the detection accuracy level during preset sensitive periods.

7. The elevator access control system based on facial recognition according to claim 1, characterized in that, The elevator control actuator interacts with the elevator main control cabinet through an industrial-grade serial communication interface or a controller local area network bus. The elevator control actuator includes a feedback monitoring unit, which is used to acquire the current operating status of the elevator, the floor position, and the door operator opening and closing status in real time, and feed the information back to the permission decision unit to achieve closed-loop management. The elevator control actuator uses optocoupler isolation technology at the hardware level to physically isolate the input and output terminals from the original elevator circuit system. The elevator control actuator has a built-in heartbeat detection circuit. Once a software anomaly is detected, the heartbeat detection circuit automatically switches to bypass mode, restoring the elevator's manual operation function.

8. The elevator access control system based on facial recognition according to claim 1, characterized in that, It also includes an abnormal behavior recording module; The abnormal behavior recording module is configured to automatically capture the current video segment and encrypt and encapsulate the data using an advanced encryption standard algorithm when the liveness detection engine determines that a spoofing attack has been detected multiple times in a row. The abnormal behavior recording module introduces blockchain traceability technology, which packages and stores key metadata, including alarm event hash value, processing result and decision evidence chain, on a private blockchain. The encrypted data, along with timestamp, device unique identifier, geographical coordinates and identified forgery type label, is uploaded to the management terminal through a secure transmission layer protocol.

9. The elevator access control system based on facial recognition according to claim 1, characterized in that, The system adopts an edge computing and distributed collaborative architecture, including edge sensing terminals deployed on the elevator car side, local processing arrays deployed on floor shaft nodes, and a central computing power cluster deployed in the computer room. The edge sensing terminal integrates a binocular stereo vision camera array consisting of two infrared-enhanced sensors, configured to generate high-precision depth map information while acquiring facial video streams, and synchronously transmit the video stream and depth information stream to the local processing array. The local processing array is responsible for performing remote photoplethysmography pulse wave signal extraction and preliminary liveness filtering, and uses a distributed spatiotemporal algorithm to eliminate the global motion vector caused by elevator car swaying. The central computing cluster is equipped with the liveness detection engine, the face recognition comparison module, and the permission decision unit. It is configured to run an adversarial generative network model in a multi-task parallel mode. While performing liveness detection, it analyzes the blood perfusion distribution map of elevator passengers. The blood perfusion distribution map includes heart rate information and blood oxygen saturation variation trends.

10. The elevator access control system based on facial recognition according to claim 1, characterized in that, It also includes a wireless near-field sensing module, a voiceprint texture analysis unit, and environmental temperature and humidity compensation logic; The wireless near-field sensing module is configured to detect the identity of authorized mobile terminals in the waiting area in front of the elevator via Bluetooth protocol or ultra-wideband positioning technology, and use them as an auxiliary index to retrieve the user's historical physiological baseline in the database. The voiceprint texture analysis unit is configured to acquire the user's voice signal through a pickup array and extract acoustic features, and to perform cross-modal consistency checks on the extracted heart rhythm and the voice fundamental frequency fluctuation and breathing rhythm in the voiceprint. The environmental temperature and humidity compensation logic is connected to the building automation management system to obtain environmental temperature and humidity parameters, and corrects the gain coefficient of the physiological signal extraction device and the sensitivity threshold of the liveness detection engine based on the parameters.

Images