Firewall emergency passage control system based on biometric identification

Abstract

This application provides a biometric-based emergency passage control system for fire-resistant curtain walls, relating to the field of building fire safety technology. It performs time-stamp synchronization and alignment, and multispectral spatial coordinate registration on multi-source sensor data. Subsequently, it quantifies and evaluates smoke concentration and ambient temperature to obtain an environmental noise index. Based on this index, it performs adaptive defogging reconstruction and contrast enhancement on visual feature regions to extract refined feature vectors. In the modal fusion stage, the system dynamically allocates the weights of each modality. It solves the high false rejection rate problem caused by feature distortion in fire environments through feature uncertainty quantification and nonlinear kernel mapping similarity reconstruction, and introduces cross-modal consistency calibration penalty logic to address the risk of false matches caused by environmental interference. Finally, it dynamically adjusts the decision threshold based on the environmental noise index, and drives the fire-resistant curtain wall's electrically controlled push rod in conjunction with the fire alarm trigger signal, ensuring reliable opening of escape routes while maintaining the tightness of access control in safe areas.

Description

Technical Field

[0001] This application relates to the field of building fire safety technology, and more specifically, to a fire-resistant curtain wall emergency passage control system based on biometric recognition. Background Technology

[0002] Fire-resistant curtain walls, as key physical partitions in building fire protection systems, play a crucial role in preventing the spread of fire and smoke during a fire. However, after a fire-resistant curtain wall is closed, the emergency exits within its coverage area need to provide rapid evacuation routes for authorized personnel while maintaining the integrity of the building's safety zones. This requires the passage control system to accurately identify and authorize access in extreme environments. Traditional emergency exit controls often rely on mechanical switches or simple access control systems, which face problems such as incomplete identity verification or delayed response in fire scenarios, making it difficult to balance safety control and escape efficiency.

[0003] However, fire environments pose a severe challenge to the reliability of biometric recognition systems. The non-uniform smoke obscuring and high-temperature thermal radiation generated by a fire cause non-linear drift in biometric features within the mapping space through physical scattering and optical distortion. For example, facial geometric feature points shift in position under dense smoke interference, resulting in non-linear distortion of the extracted feature vector relative to the registered template. If the system uses a fixed linear metric for similarity comparison, it cannot distinguish between feature distortion induced by the environment and increased feature distance due to identity discrepancies, easily leading to a high probability of false rejections and causing escape routes to become deadlocked at critical moments. Furthermore, in multimodal fusion recognition, if the system processes each modality score in isolation while ignoring cross-modal consistency correlations between visual and non-visual modalities, when fire flashes cause false feature matching by the visual sensor, the system will struggle to distinguish false recognition results caused by environmental interference, creating an irreconcilable contradiction between ensuring escape efficiency and maintaining safety control.

[0004] Therefore, an optimized biometric-based emergency access control system for fire-resistant curtain walls is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a biometric-based fire-resistant curtain wall emergency passage control system, comprising: The data synchronization and registration module is used to perform timestamp synchronization and multispectral spatial coordinate registration on the acquired sensor data sources to obtain synchronization data packets. The sensor data sources include video streams, thermal imaging signals, smoke concentration, and ambient temperature. The environmental interference factor quantification and evaluation module is used to quantify the environmental interference factors of smoke concentration and ambient temperature in the synchronization data packet to obtain the environmental noise index. The biometric robustness preprocessing and enhancement module is used to perform biometric robustness preprocessing and enhancement on visual feature regions in synchronization data packets based on the environmental noise index to obtain refined feature vectors. The modal weight dynamic allocation module is used to map the associated weight coefficients based on the environmental noise index to obtain the modal weight distribution, and to calculate the modal component similarity between the refined feature vector and the pre-stored registered template features based on the environmental noise index to obtain the original modal component set. The modal weight distribution is used to perform multi-dimensional confidence weighted fusion on the original modal component set to obtain the fusion confidence score. The adaptive decision threshold dynamic calculation module is used to dynamically calculate the decision threshold based on the preset basic safety threshold according to the environmental noise index to obtain the decision threshold. The access authorization and control feedback module is used to determine the personnel access authorization status based on the threshold exceedance judgment result of the fusion confidence score relative to the decision threshold, and to generate control feedback for driving the fireproof curtain wall electric control push rod in combination with the fire alarm trigger signal.

[0006] Compared with existing technologies, the biometric-based fire-resistant curtain wall emergency passage control system provided in this application first performs time-stamp synchronization and multispectral spatial coordinate registration on multi-source sensor data to form unified spatiotemporal reference data. Then, it quantifies and evaluates smoke concentration and ambient temperature to generate an environmental noise index, which serves as a global environmental perception parameter throughout subsequent processing stages. The system utilizes the environmental noise index to drive adaptive defogging reconstruction and contrast enhancement, completing robust preprocessing of damaged features, and extracting refined feature vectors through a convolutional neural network. In the modal fusion stage, the system introduces feature uncertainty quantification and Gaussian kernel mapping mechanisms to replace traditional linear metrics, solving the problems of high false rejection rate and passage deadlock caused by nonlinear distortion of features in fire environments. Simultaneously, it filters false matching results through cross-modal consistency calibration penalty logic and dynamically adjusts modal weights based on the environmental noise index to complete confidence fusion. Finally, the system adaptively adjusts the decision threshold, combines fire alarm signals to determine access authorization, and drives the fire-resistant curtain wall's electronic control actuator to perform actions, achieving coordinated protection of escape efficiency and safety management in extreme environments. Attached Figure Description

[0007] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0008] Figure 1This is a system block diagram of a biometric-based fireproof curtain wall emergency passage control system according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow of a biometric-based fireproof curtain wall emergency passage control system according to an embodiment of this application; Figure 3 This is a block diagram of the biometric robustness preprocessing and enhancement module in a biometric recognition-based fireproof curtain wall emergency passage control system according to an embodiment of this application; Figure 4 This is a block diagram of the modal weight dynamic allocation module in the biometric recognition-based fireproof curtain wall emergency passage control system according to an embodiment of this application; Figure 5 This is a block diagram of the modal component similarity calculation processing unit in the biometric identification-based fireproof curtain wall emergency passage control system according to an embodiment of this application; Figure 6 This is a block diagram of the adaptive decision threshold dynamic calculation module in the biometric recognition-based fireproof curtain wall emergency passage control system according to an embodiment of this application. Detailed Implementation

[0009] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0010] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0011] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0012] This application uses system block diagrams and data flow diagrams to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0013] Currently, in the identification scenario of emergency passages in fire-resistant curtain walls, the non-uniform smoke and high-temperature heat radiation generated by fire can cause non-linear distortions in biometrics. Traditional linear measurement methods cannot distinguish between environmentally induced feature drift and differences in true identity, easily leading to high false rejection rates and deadlock of escape routes. Furthermore, the isolated processing of multimodal scores ignores cross-modal consistency correlations, making it difficult for the system to identify false matching results under environmental interference. Therefore, the technical solution of this application proposes a fire-resistant curtain wall emergency passage control system based on biometric recognition. Specifically, the system first performs time-stamp synchronization and multispectral spatial coordinate registration on multi-source sensor data such as video streams, thermal imaging signals, smoke concentration, and ambient temperature. Then, it quantitatively evaluates smoke concentration and ambient temperature to obtain an environmental noise index, which serves as an environmental perception benchmark throughout the entire process, driving the adaptive adjustment of subsequent stages. The system performs adaptive defogging reconstruction and contrast enhancement on visual features based on the environmental noise index to extract refined feature vectors. During the modality fusion stage, it dynamically allocates the weights of each modality. It reduces the probability of false rejection by incorporating feature nonlinear offset through feature uncertainty quantification and nonlinear kernel mapping similarity reconstruction. It also introduces cross-modal consistency calibration penalty logic to filter false matching results. Finally, it dynamically adjusts the decision threshold based on the environmental noise index and drives the fireproof curtain wall electric control push rod in combination with the fire alarm trigger signal. This ensures the reliable opening of escape routes while maintaining the tight access control of safe areas.

[0014] Figure 1 This is a system block diagram of a fireproof curtain wall emergency passage control system based on biometric recognition according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in a biometric-based fire-resistant curtain wall emergency passage control system according to an embodiment of this application. Figure 1 and Figure 2As shown, the biometric-based fireproof curtain wall emergency passage control system 100 according to an embodiment of this application includes: a data synchronization and registration module 110, used to perform timestamp synchronization alignment and multispectral spatial coordinate registration on the acquired sensor data sources to obtain a synchronization data packet, the sensor data sources including video streams, thermal imaging signals, smoke concentration, and ambient temperature; an environmental interference factor quantification and evaluation module 120, used to perform environmental interference factor quantification and evaluation on the smoke concentration and ambient temperature in the synchronization data packet to obtain an environmental noise index; a biometric robustness preprocessing and enhancement module 130, used to perform biometric robustness preprocessing and enhancement on the visual feature regions in the synchronization data packet based on the environmental noise index to obtain a refined feature vector; and a modal weight dynamic allocation module 14. 0, used to map the associated weight coefficients based on the environmental noise index to obtain the modal weight distribution, and to calculate the modal component similarity between the refined feature vector and the pre-stored registered template features based on the environmental noise index to obtain the original modal component set, and to perform multi-dimensional confidence weighted fusion on the original modal component set using the modal weight distribution to obtain the fusion confidence score; the adaptive decision threshold dynamic calculation module 150, used to perform adaptive decision threshold dynamic calculation on the preset basic safety threshold based on the environmental noise index to obtain the decision threshold; the access authorization and control feedback module 160, used to determine the personnel access authorization status based on the threshold over-limit judgment result of the fusion confidence score relative to the decision threshold, and to generate control feedback for driving the fireproof curtain wall electric control push rod in combination with the fire alarm trigger signal.

[0015] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the data synchronization and registration module 110 is used to perform timestamp synchronization alignment and multispectral spatial coordinate registration on the acquired sensor data sources to obtain synchronization data packets. The sensor data sources include video streams, thermal imaging signals, smoke concentration, and ambient temperature. It should be noted that, due to the involvement of multiple heterogeneous sensor data sources such as video streams, thermal imaging signals, smoke concentration, and ambient temperature in the fire-resistant curtain wall emergency passage control scenario, each sensor has inherent differences in sampling frequency, clock reference, and coordinate space. Directly sending unaligned multi-source data into the subsequent processing link will lead to temporal misalignment and spatial mismatch during feature extraction and fusion judgment. Therefore, the technical solution of this application first performs timestamp synchronization alignment and multispectral spatial coordinate registration on the acquired sensor data sources to obtain synchronization data packets. Through the above processing, multi-source sensing data under different sampling rhythms and coordinate systems can be unified into the same spatiotemporal reference frame, providing a consistent data foundation for subsequent environmental assessment and biometric recognition.

[0016] More specifically, in a concrete example of this application, the generation process of the aforementioned synchronization data packet includes the following processing stages. First, hardware-level interrupt capture and frame encapsulation processing is performed on each sensor data source to obtain a buffer stream array. In a fire emergency scenario, a visible light camera outputs a video stream at a rate of 30 frames per second, a thermal imaging sensor outputs an infrared thermal image at a rate of 10 frames per second, and smoke concentration sensors and temperature sensors output scalar sampled values ​​at fixed periods. The arrival times and formats of each data source are different. In this stage, the raw output signal of each sensor is captured through a hardware interrupt mechanism. In the interrupt service routine, a timestamp of the local hardware clock is appended to each frame of data, and it is encapsulated into a data frame of a uniform format and written to the corresponding circular buffer, thereby forming a buffer stream array containing the data frame sequences of each sensor.

[0017] Subsequently, a time alignment matrix is ​​obtained by correcting the clock reference time offset of the buffer stream array based on the master clock. Since there is a drift deviation between the local clocks of each sensor and the master clock of the main control processor, the timestamps of each frame in the buffer stream array need to be uniformly corrected. In this stage, using the master clock as a reference, the offset of each sensor's local clock relative to the master clock is calculated, the timestamps of each frame in the buffer stream array are linearly compensated, and the data frames are aligned to the same time tick according to the corrected timestamps, forming a time alignment matrix indexed by a unified time axis.

[0018] Furthermore, a pre-calibrated affine transformation matrix is ​​used to perform multispectral spatial coordinate registration between the visible light pixel coordinates and the thermal imaging pixel coordinates in the time alignment matrix to obtain spatiotemporal registration data. Due to differences in installation location, focal length, and resolution, the pixel coordinate systems of the output images from the visible light camera and the thermal imaging sensor are not consistent; the same physical location corresponds to different pixel coordinates in the two images. In this stage, an affine transformation matrix, pre-calibrated using a checkerboard calibration board, is used to perform an affine transformation on the pixel coordinates of the thermal imaging image, establishing a pixel-by-pixel spatial correspondence with the pixel coordinates of the visible light image. This unifies the time-aligned multispectral image data into the same spatial coordinate system, forming spatiotemporal registration data.

[0019] Finally, the image pixel matrix and physical values ​​in the spatiotemporal registration data are stitched and encapsulated to generate a synchronization data packet. In this stage, the spatially registered visible light pixel matrix and thermal imaging pixel matrix are stitched together along the channel dimension, and the smoke concentration and ambient temperature values ​​at the same time frame are written as additional fields into the data packet header, forming a synchronization data packet containing multispectral image information and environmental physical parameters, which can then be used in subsequent environmental interference factor quantitative assessment and biometric extraction stages.

[0020] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the environmental interference factor quantification assessment module 120 is used to quantify and assess the smoke concentration and ambient temperature in the synchronization data packet to obtain the environmental noise index. It should be noted that since changes in smoke concentration and ambient temperature in a fire scenario directly affect the quality of biometric data acquisition, the degree of interference caused to visual sensors and thermal imaging sensors varies depending on the intensity of the fire. Without a quantitative characterization of the current environmental degradation, subsequent feature preprocessing, modal fusion, and decision threshold adjustment will lack a basis for adaptive control. Therefore, the technical solution of this application further quantifies and assesses the smoke concentration and ambient temperature in the synchronization data packet to obtain the environmental noise index. Through the above processing, discrete environmental physical parameters can be transformed into unified quantified indicators, providing a consistent environmental perception benchmark for all links in the entire chain.

[0021] More specifically, in a specific example of this application, the generation process of the aforementioned environmental noise index includes the following processing stages. First, the synchronization data packet is parsed and extracted based on a preset bit offset to obtain the original environmental sampling value composed of the smoke concentration value and the ambient temperature value. The synchronization data packet is encapsulated in a predefined binary structure, wherein the image pixel matrix occupies the front area of ​​the data packet, and the smoke concentration value and the ambient temperature value are stored in the additional fields of the data packet according to a fixed byte offset. In this stage, the starting address of the smoke concentration field and the ambient temperature field is located according to the preset bit offset, and the corresponding length of byte data is read and decoded according to the agreed data type to obtain the smoke concentration value expressed as a percentage and the ambient temperature value expressed in degrees Celsius, which together constitute the original environmental sampling value.

[0022] Subsequently, the original environmental sampling values ​​are normalized to reflect abnormal environmental conditions based on preset limit values ​​to obtain normalized interference features including normalized smoke factor and normalized temperature factor. Smoke concentration and ambient temperature have different physical dimensions and numerical ranges. Smoke concentration ranges from 0 to 100 percent, while ambient temperature in a fire scenario can rise from normal temperature to several hundred degrees Celsius, making direct weighted calculation impossible. In this stage, the smoke concentration value is divided by a preset smoke concentration limit value, and the ambient temperature value is divided by a preset temperature limit value, mapping both to a normalization interval of 0 to 1. A normalization result tending towards 0 indicates the environment is close to normal, while a tendency towards 1 indicates environmental deterioration approaching a preset extreme boundary. This yields the normalized smoke factor and normalized temperature factor, which together constitute the normalized interference features.

[0023] Furthermore, the normalized interference features are weighted and synthesized using the environmental noise index to obtain the environmental noise index. The interference mechanisms of smoke concentration and ambient temperature on biometric recognition differ. Smoke primarily affects the imaging clarity of the visible light channel, while high temperature affects the signal-to-noise ratio of the thermal imaging channel; their contributions to recognition accuracy differ. In this stage, preset weighting coefficients are assigned to the normalized smoke factor and the normalized temperature factor, respectively, and a weighted summation is performed to obtain the environmental noise index, which ranges from 0 to 1. A higher index value indicates a stronger overall interference from the current fire environment on biometric recognition. Subsequent feature enhancement intensity, modality weight allocation, and decision threshold adjustment are all driven by this index.

[0024] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the biometric robustness preprocessing and enhancement module 130 is used to perform biometric robustness preprocessing and enhancement on the visual feature regions in the synchronization data packet based on the environmental noise index to obtain refined feature vectors. It should be noted that, given that smoke obscuring in a fire environment reduces the clarity of visible light images and high-temperature thermal radiation compresses the dynamic range of thermal imaging signals, a higher environmental noise index means more severe degradation of biometric information in the visual feature regions. If feature extraction is directly performed on the damaged original image data, the resulting feature vector will carry a large amount of environmental noise components, leading to a decrease in the reliability of subsequent modality fusion and similarity comparison. Based on this, the technical solution of this application further performs robust preprocessing and enhancement of the visual feature regions in the synchronization data packets based on the environmental noise index to obtain refined feature vectors. Specifically, the environmental noise index drives the adaptive adjustment of the dehazing reconstruction intensity and histogram equalization parameters, performing dehazing restoration on visible light video frames and contrast enhancement on thermal imaging signals. A pre-trained convolutional neural network is then used to perform feature mapping and attention-weighted pooling on the enhanced multimodal images, compressing high-dimensional image information into compact, refined feature vectors. Through the above processing, biometric information degraded by smoke and high temperatures can be adaptively recovered under different fire intensities, making the extracted refined feature vectors robust to environmental interference and providing high-quality feature input for subsequent modality weight allocation and confidence fusion.

[0025] Figure 3 This is a block diagram of the biometric robustness preprocessing and enhancement module in a biometric-based fire-resistant curtain wall emergency passage control system according to an embodiment of this application. Figure 3As shown, the biometric robustness preprocessing and enhancement module 130 includes: an environment-adaptive dehazing reconstruction unit 131, used to perform environment-adaptive dehazing reconstruction on video frames in the synchronous data packet based on the environmental noise index to obtain a repaired intermediate data packet; an adaptive histogram equalization processing unit 132, used to dynamically adjust the clipping limit threshold in the repaired intermediate data packet based on the environmental noise index, and perform contrast-limited adaptive histogram equalization processing on the thermal imaging signal to obtain an enhanced multimodal image group; and a feature mapping and attention-weighted pooling processing unit 133, used to perform feature mapping and attention-weighted pooling processing on the enhanced multimodal image group using a pre-trained convolutional neural network to obtain a refined feature vector.

[0026] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the environmental adaptive defogging reconstruction unit 131 is used to perform environmental adaptive defogging reconstruction on video frames in the synchronous data packet based on the environmental noise index to obtain a repaired intermediate data packet. It should be noted that, because smoke generated by a fire forms a non-uniform scattering and masking layer in visible light video frames, the contrast and detail information of biometric regions such as facial contours and iris textures are suppressed by the forward scattering of smoke particles. Furthermore, the higher the smoke concentration, the more severe the degradation of these regions. If a fixed-parameter defogging strategy is used, it may introduce over-enhanced artifacts in low-smoke scenarios and insufficient defogging power in high-smoke scenarios. Therefore, the technical solution of this application further performs environmental adaptive defogging reconstruction on video frames in the synchronous data packet based on the environmental noise index to obtain a repaired intermediate data packet. Through the above processing, the defogging intensity can be dynamically adjusted according to the actual deterioration of the current fire environment, and the recognizability of biometric regions in video frames can be restored evenly under different smoke concentration conditions.

[0027] More specifically, in a concrete example of this application, the process of generating the aforementioned repair intermediate data packet includes the following processing stages. First, the visible light video frame corresponding to the current time frame is extracted from the synchronization data packet, and the transmittance of the video frame is estimated based on the dark channel prior principle. For each pixel position in the video frame, the minimum value of each color channel is taken within its local neighborhood window to construct a dark channel map. Then, the initial transmittance at each pixel position is estimated based on the pixel intensity distribution in the dark channel map. This transmittance characterizes the proportion of scene light at that position that penetrates the smoke layer and reaches the sensor.

[0028] Subsequently, the initial transmittance is adaptively corrected using the environmental noise index. A higher environmental noise index indicates more severe smoke obscuration, requiring stronger compensation for the transmittance estimation. In this stage, the environmental noise index is mapped to a lower limit constraint parameter for transmittance. When the environmental noise index is high, the lower limit value of transmittance is increased to avoid numerical overflow and noise amplification during the defogging process in dense smoke areas where transmittance tends to 0. When the environmental noise index is low, the lower limit constraint is relaxed to retain more space for defogging repair, thus obtaining an adaptively corrected transmittance map.

[0029] Furthermore, scene radiometric restoration is performed on the video frame using a modified transmittance map based on an atmospheric scattering model. The brightest pixel region in the dark channel image is selected to estimate the global atmospheric light value. This atmospheric light value, along with the modified transmittance map, is then substituted into the inverse operation of the atmospheric scattering model to perform dehazing restoration on each pixel of the video frame, stripping away the smoke scattering component and restoring the original radiometric information of the scene. After the inverse operation, the restored result undergoes pixel value truncation and normalization to ensure that the pixel values ​​of the output image are within the effective dynamic range, thereby generating a dehazed and restored visible light image frame. This restored visible light image frame, along with the undehazed thermal imaging signal from the synchronization data packet, is encapsulated to form a restored intermediate data packet for subsequent contrast enhancement and feature extraction stages.

[0030] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the adaptive histogram equalization processing unit 132 is used to dynamically adjust the shearing limitation threshold in the repair intermediate data packet based on the environmental noise index, and to perform contrast-limited adaptive histogram equalization processing on the thermal imaging signal to obtain an enhanced multimodal image group. It should be noted that, due to the high-temperature thermal radiation in a fire scene causing the grayscale distribution of the thermal imaging signal to tend to concentrate, the contrast between facial thermal features and the background heat source is compressed. If a histogram equalization strategy with a fixed shearing limitation threshold is adopted, the enhancement strength is insufficient to separate biological thermal features from environmental thermal noise when the environmental noise index is high, and may introduce false edges due to over-enhancement when the environmental noise index is low. Based on this, the technical solution of this application further dynamically adjusts the shearing limitation threshold in the repair intermediate data packet based on the environmental noise index, and performs contrast-limited adaptive histogram equalization processing on the thermal imaging signal to obtain an enhanced multimodal image group. Through the above processing, the contrast enhancement amplitude can be adaptively controlled according to the interference intensity of the current fire environment, effectively highlighting the contours of biological thermal features in the thermal imaging signal.

[0031] More specifically, in a concrete example of this application, a thermal imaging infrared grayscale image is first extracted from the repair intermediate data package. An environmental noise index is linearly mapped to a preset base shearing limit value to obtain a dynamic shearing limit threshold. The higher the environmental noise index, the larger the threshold, thus enhancing the separation capability of facial thermal features obscured by a high-temperature background. Subsequently, the thermal imaging infrared grayscale image is divided into multiple local sub-blocks. The grayscale histogram of each sub-block is truncated and uniformly redistributed using the dynamic shearing limit threshold. Then, a cumulative distribution function mapping is performed to achieve local contrast equalization. After processing each sub-block, bilinear interpolation is used to smooth the transition of boundary regions to eliminate grayscale jumps, thereby obtaining a contrast-enhanced thermal imaging image. Finally, this thermal imaging image is combined and encapsulated with the dehazed and reconstructed visible light image frame from the repair intermediate data package to form an enhanced multimodal image group, which is then used in subsequent feature mapping and attention-weighted pooling stages.

[0032] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the feature mapping and attention-weighted pooling processing unit 133 is used to perform feature mapping and attention-weighted pooling processing on the enhanced multimodal image group using a pre-trained convolutional neural network to obtain refined feature vectors. It should be noted that since the visible light image frames and thermal imaging images in the enhanced multimodal image group still exist in the form of high-dimensional pixel matrices, directly using the original pixel data for subsequent similarity comparison will face problems of excessive dimensionality and redundant information. Furthermore, the contribution of pixels at different spatial locations to identity recognition varies, and the feature information density of key facial regions is much higher than that of background regions. Based on this, the technical solution of this application further utilizes a pre-trained convolutional neural network to perform feature mapping and attention-weighted pooling processing on the enhanced multimodal image group to obtain refined feature vectors. Through the above processing, high-dimensional multimodal image information can be compressed into a compact and discriminative low-dimensional feature representation. Simultaneously, the attention mechanism focuses the feature vectors on local regions with strong identity representation capabilities, suppressing interference from background noise and irrelevant regions.

[0033] More specifically, in a concrete example of this application, visible light image frames and thermal imaging images from the enhanced multimodal image set are respectively input into a pre-trained convolutional neural network. Layer-by-layer feature mapping through multiple convolutional kernels extracts high-level semantic feature maps corresponding to each modality. Each spatial location of each feature map encodes the local feature response within the corresponding receptive field. Subsequently, attention-weighted pooling is performed on the high-level semantic feature maps of each modality. An attention branch network calculates an attention weight for each spatial location of the feature map. Areas with rich textures, such as the periorbital region and nasal alae, affected by residual smoke from fires receive higher attention weights, while edge areas obscured by smoke or blurred by high temperatures receive lower attention weights. These attention weights are then used to perform weighted spatial pooling on the feature maps, compressing the two-dimensional feature maps of each modality into one-dimensional feature vectors. Finally, the one-dimensional feature vectors of each modality are concatenated sequentially to form a refined feature vector containing feature components from both the visible light and thermal imaging modalities, which is then used for subsequent dynamic allocation of modality weights and similarity calculation.

[0034] Specifically, the pre-trained convolutional neural network is a deep convolutional neural network model pre-trained on a large-scale face recognition dataset and an infrared thermal imaging human body feature dataset. Its network structure includes multiple cascaded convolutional layers, batch normalization layers, and nonlinear activation layers, used to extract multi-scale feature representations from the input image, ranging from low-level texture edges to high-level semantic structures. Before being deployed to the fireproof curtain wall emergency passage control scenario, the pre-trained convolutional neural network has undergone parameter optimization on diverse samples containing different lighting conditions, occlusion levels, and pose variations, enabling it to generalize the extraction capabilities of facial texture features and human body thermal radiation distribution features. During the actual inference phase, the network parameters of the pre-trained convolutional neural network remain frozen and do not participate in online updates. The visible light image frames and thermal imaging images in the enhanced multimodal image group are forward-propagated through the convolutional layers of the network. Each convolutional layer performs feature response calculations within the local receptive field of the input image using the convolutional kernels obtained through pre-training, and abstracts high-level semantic feature maps related to identity recognition layer by layer. Each spatial location of the feature map encodes local discriminative information within the corresponding receptive field. Subsequently, the attention-weighted pooling layer evaluates the importance of each spatial location of the feature map and performs weighted compression, finally outputting a compact one-dimensional feature vector.

[0035] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the modal weight dynamic allocation module 140 is used to map the associated weight coefficients based on the environmental noise index to obtain the modal weight distribution, and to calculate the modal component similarity between the refined feature vector and the pre-stored registered template features based on the environmental noise index to obtain the original modal component set. The modal weight distribution is then used to perform multi-dimensional confidence weighted fusion of the original modal component set to obtain a fusion confidence score. It should be noted that, given the different interference mechanisms of fire environment on visible light and thermal imaging modes, the reliability of each mode varies dynamically under different fire intensities. If modal fusion and similarity comparison are performed using fixed weights and fixed linear metric criteria, it cannot adapt to the real-time evolution of the environment, nor can it distinguish between environmentally induced feature nonlinear distortions and differences in true identity. Furthermore, the isolated processing of each modal score ignores cross-modal consistency correlation, and lacks means to verify and suppress false matching results under interference such as fire and light flicker. Based on this, the technical solution of this application further maps the correlation weight coefficients based on the environmental noise index to obtain the modal weight distribution, and calculates the modal component similarity between the refined feature vector and the pre-stored registration template features based on the environmental noise index to obtain the original modal component set. The modal weight distribution is then used to perform multi-dimensional confidence weighted fusion of the original modal component set to obtain the fusion confidence score. Through the above processing, the fusion contribution of each modality can be dynamically allocated according to the environmental noise index, reducing the probability of false rejections caused by feature distortion in a fire environment. Simultaneously, cross-modal mutual trust verification filters out false matching results, ensuring that the fusion confidence score balances access reliability and access control rigor.

[0036] Figure 4 This is a block diagram of the modal weight dynamic allocation module in a biometric-based fire-resistant curtain wall emergency passage control system according to an embodiment of this application. Figure 4 As shown, the modal weight dynamic allocation module 140 includes: an associated weight coefficient mapping unit 141, used to perform associated weight coefficient mapping in a preset weight matrix based on the environmental noise index to obtain a modal weight distribution containing visual modal weights and non-visual modal weights; a modal component similarity calculation processing unit 142, used to perform modal component similarity calculation processing on refined feature vectors and pre-stored registered template features based on the environmental noise index to obtain an original modal component set; and a multi-dimensional confidence weighted fusion unit 143, used to perform multi-dimensional confidence weighted fusion on the original modal component set using the modal weight distribution to obtain a fusion confidence score.

[0037] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the associated weight coefficient mapping unit 141 is used to perform associated weight coefficient mapping in a preset weight matrix based on the environmental noise index to obtain a modal weight distribution including visual modal weights and non-visual modal weights. It should be noted that, since smoke primarily degrades the imaging quality of the visible light channel while high temperatures compress the signal-to-noise ratio of the thermal imaging channel, the credibility of the identity information carried by each modality varies dynamically under different fire intensities. If a fixed weight is assigned to each modality during the fusion stage, the severely damaged modalities will make an excessively high negative contribution to the fusion result. Therefore, the technical solution of this application further performs associated weight coefficient mapping in a preset weight matrix based on the environmental noise index to obtain a modal weight distribution including visual modal weights and non-visual modal weights. Through the above processing, the contribution ratio of each modality in subsequent fusion can be adjusted in real time according to the current environmental degradation level, allowing modalities with higher credibility to gain greater fusion influence.

[0038] More specifically, in a concrete example of this application, a preset weight matrix uses the discretized interval of the environmental noise index as the row index and the visual modality weight and non-visual modality weight as the column index, pre-storing the weight coefficients corresponding to each modality under different environmental degradation levels. During the mapping process, the current environmental noise index is first matched with the row index interval in the preset weight matrix to determine the environmental level interval it falls into. Then, the visual modality weight and non-visual modality weight in the corresponding row of that interval are read. When the environmental noise index is high, smoke severely obstructs the visible light channel, and the visual modality weight is mapped to a lower value while the non-visual modality weight is mapped to a higher value, making the fusion process rely more on the thermal imaging modality less affected by smoke. When the environmental noise index is low, the imaging quality of the visible light channel is better, and the visual modality weight is mapped to a higher value to fully utilize the rich texture details in the visible light image. Normalization constraints are applied to the mapped visual modality weight and non-visual modality weight to ensure that their sum is 1, thus forming a modality weight distribution for subsequent multi-dimensional confidence-weighted fusion.

[0039] In the aforementioned biometric-based fireproof curtain wall emergency passage control system 100, the modal component similarity calculation processing unit 142 is used to perform modal component similarity calculation processing on the refined feature vector and the pre-stored registered template features based on the environmental noise index to obtain the original modal component set. It should be noted that, given that the non-uniform smoke obscuring and high-temperature thermal radiation generated by a fire will cause nonlinear drift of the refined feature vector relative to the pre-stored registered template features through physical scattering and optical distortion, traditional linear comparison based on fixed geometric measurement criteria such as cosine similarity cannot distinguish between environmentally induced feature distortion and the increase in feature distance caused by identity discrepancies. Furthermore, the isolated processing of each modal score ignores the cross-modal consistency correlation between visual and non-visual modalities, and lacks effective verification methods when the visual sensor generates false feature matching due to fire flash. Based on this, the technical solution of this application further performs modal component similarity calculation on the refined feature vector and the pre-stored registered template features based on the environmental noise index to obtain the original modal component set. Specifically, it quantifies feature uncertainty by applying exponential weight gain to the local entropy values ​​of the feature space components, uses a Gaussian kernel function for nonlinear mapping alignment to accommodate feature offsets caused by the environment, and introduces cross-modal consistency penalty logic to smooth and correct the modal scores. Through the above processing, it is possible to reduce the false rejection probability caused by feature nonlinear distortion in extreme fire environments, while filtering out false matching results caused by environmental interference, so that the output original modal component set takes into account both recognition passability and security.

[0040] Figure 5 This is a block diagram of the modal component similarity calculation processing unit in a biometric-based fire-resistant curtain wall emergency passage control system according to an embodiment of this application. Figure 5 As shown, the modal component similarity calculation processing unit 142 includes: an exponential weight gain processing subunit 1421, used to perform exponential weight gain processing on the local entropy values ​​of the feature space components of the refined feature vector based on the environmental noise index to obtain a feature uncertainty matrix; a nonlinear mapping alignment processing subunit 1422, used to perform nonlinear mapping alignment processing on the refined feature vector using the feature uncertainty matrix as the kernel radius adjustment parameter of the Gaussian kernel function to obtain a preliminary calibration similarity; and a smoothing correction processing subunit 1423, used to perform smoothing correction processing on the preliminary calibration similarity based on cross-modal consistency penalty logic based on the difference between the visual modal scores and non-visual modal scores in the preliminary calibration similarity to obtain the original modal component set.

[0041] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the exponential weight gain processing subunit 1421 is used to perform exponential weight gain processing on the local entropy values ​​of the feature space components of the refined feature vector based on the environmental noise index to obtain a feature uncertainty matrix. It should be noted that, given the uneven distribution of smoke concentration at a fire scene, the degree of environmental interference varies among different dimensions of the feature components in the refined feature vector. Some dimensions of the feature components lose their effective representation ability for identity information due to smoke obscuring or high-temperature radiation. If all dimensions of the feature components are treated equally in subsequent similarity calculations, contaminated feature components will participate in the comparison calculation, introducing misleading distance biases. Based on this, the technical solution of this application further performs exponential weight gain processing on the local entropy values ​​of the feature space components of the refined feature vector based on the environmental noise index to obtain a feature uncertainty matrix. By quantifying the degree of disorder of each dimension of the feature components during the extraction process and dynamically amplifying this disorder using the environmental noise index, it can perceive in real time which feature components are no longer representative due to smoke or high-temperature interference, providing a reliability benchmark for subsequent similarity calculations. Through the above processing, high uncertainty regions can be identified and marked, avoiding misleading decisions based on contaminated data, and providing a dimensional level of reliability reference for subsequent nonlinear similarity reconstruction based on kernel mapping.

[0042] More specifically, in a concrete example of this application, the refined feature vector output from the upstream step is first received. Local entropy values ​​are calculated for each feature component of each modality to characterize the information disorder of that dimension during feature extraction. Higher entropy values ​​indicate greater instability of the feature components in that dimension. Subsequently, the environmental noise index is used as a gain operator to exponentially weight the local entropy values ​​of each dimension. Higher environmental noise indices amplify the entropy, further highlighting the uncertainty of the damaged dimensions under high fire intensity conditions, thereby generating the feature uncertainty matrix corresponding to each modality. This process can be represented as: in, Let be the characteristic uncertainty matrix of the i-th mode. This refers to the environmental noise index. This represents the j-th eigenvalue of the i-th mode in the refined eigenvector. The total dimension of the feature vector. This represents the natural exponential function. Let represent the natural logarithm function. In the above formula, The local entropy value of the i-th modal feature space component was partially calculated. The higher the entropy value, the more uniform the distribution of the feature components in each dimension, i.e. the more dispersed the information, the greater the disorder in the feature extraction process. The environmental noise index is partially transformed into a gain coefficient greater than or equal to 1 using a natural exponential function. As the environmental noise index approaches 0, the gain coefficient approaches 1, meaning no additional amplification is applied to the entropy value. As the environmental noise index approaches 1, the gain coefficient grows exponentially, thus marking the feature dimension with high disorder as a high uncertainty state. In fire evacuation routes, when the left side of an evacuee's face is partially obscured by dense smoke while the right side is relatively clear, the dimension component encoding the left facial region in the corresponding visible light modality feature vector will exhibit a high entropy value. After the exponential gain of the environmental noise index, this dimension in the feature uncertainty matrix will acquire a high uncertainty label, while the dimension component encoding the clear right side will maintain a low uncertainty value. This provides a dimensional-level reliability distinction for subsequent kernel mapping similarity reconstruction.

[0043] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the nonlinear mapping alignment processing subunit 1422 is used to perform nonlinear mapping alignment processing on the refined feature vector using the feature uncertainty matrix as the kernel radius adjustment parameter of the Gaussian kernel function to obtain a preliminary calibration similarity. It should be noted that, given that traditional linear comparisons based on fixed geometric metrics such as cosine similarity cannot accommodate nonlinear feature shifts caused by fire, the positional shift of facial geometric feature points under dense smoke interference causes nonlinear distortion of the refined feature vector relative to the registered template. Therefore, a nonlinear mapping method with high tolerance is needed to replace rigid distance metrics. Based on this, the technical solution of this application further uses the feature uncertainty matrix as the kernel radius adjustment parameter of the Gaussian kernel function to perform nonlinear mapping alignment processing on the refined feature vector to obtain a preliminary calibration similarity. By dynamically injecting the feature uncertainty matrix into the kernel radius adjuster to expand the tolerance radius of the judgment kernel, it can accommodate micro-shifts in feature components caused by light refraction in smoke-filled evacuation passages. Even when the facial features of an escapee are physically distorted due to environmental influences, the consistency of their identity can still be captured. The above processing can reduce the false rejection rate in extreme fire environments and eliminate the risk of deadlock caused by channel recognition failure through an adaptive matching strategy.

[0044] More specifically, in a concrete example of this application, firstly, pre-stored registration template features corresponding to the current person to be identified are retrieved from the database. Then, the tested feature components of each modality in the refined feature vector are paired modally by modality with the corresponding template components in the registered template features. Subsequently, a Gaussian kernel function is introduced as a similarity measurement tool. The feature uncertainty matrix generated in the previous step is dynamically injected into the kernel radius adjustment parameter of the Gaussian kernel function. For modal dimensions with high uncertainty, the kernel radius is correspondingly increased to improve the tolerance for feature shift in that dimension. For modal dimensions with low uncertainty, the kernel radius is maintained near the basic bandwidth to maintain the accuracy of the judgment. Based on the adjusted kernel radius, the Euclidean distance between the tested feature components and the template components is calculated and substituted into the Gaussian kernel function for nonlinear mapping, transforming the original linear distance into a similarity score within the interval of 0 to 1, thereby obtaining the preliminary calibrated similarity corresponding to each modality. This process can be expressed as: in, To initially calibrate the i-th modal score in the similarity, To refine the measured feature components in the feature vector, Features of the pre-stored registration template The characteristic uncertainty matrix, The preset base core bandwidth, Let be the squared Euclidean distance between the feature component to be tested and the registered template feature. This represents the natural exponential function. In the above formula, the kernel radius is determined by... The decision is made when the characteristic uncertainty matrix... When the value is large, the kernel radius increases accordingly, and the decay rate of the Gaussian kernel function on the feature distance slows down. This ensures that even if there is a large Euclidean distance between the feature component to be tested and the template component, the output similarity score can still be maintained within a reasonable range and will not drop sharply to approach 0. When the value is small, the kernel radius is close to the base bandwidth. The Gaussian kernel function recovers a relatively strict judgment criterion. In fire evacuation routes, when the feature components of the left side of a person's face are shifted due to dense smoke, the feature uncertainty matrix value corresponding to this mode is high. After the kernel radius is enlarged, the Gaussian kernel function can accommodate this shift and output a reasonable similarity score. Meanwhile, the uncertainty value corresponding to the clear right side of the face is low, and the kernel radius is maintained near the basic bandwidth to maintain accurate judgment. Thus, it achieves adaptive tolerance to nonlinear feature drift induced by the environment, avoiding false rejections caused by a sharp drop in the overall similarity score due to damage to local features.

[0045] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the smoothing correction processing subunit 1423 is used to perform smoothing correction processing on the preliminary calibration similarity based on cross-modal consistency penalty logic, based on the difference between the visual modality score and the non-visual modality score in the preliminary calibration similarity, to obtain the original modal component set. It should be noted that, given that multiple modalities of biometrics should have logical consistency under the same identity, severe cross-modal conflicts indicate sensor failure or environmental deception. While the aforementioned nonlinear similarity reconstruction based on kernel mapping solves the false rejection rate problem caused by feature distortion, it still processes each modal score in isolation. In a fire scene, the glare from the flames may cause visual sensors to generate false feature matches. Without cross-modal verification logic, it will be difficult to identify such false recognition results, thereby threatening the reliability of access control in secure areas. Based on this, the technical solution of this application further applies a smoothing correction process based on cross-modal consistency penalty logic to the preliminary calibration similarity, based on the difference between the visual modal scores and non-visual modal scores, to obtain the original modal component set. This process constructs a cross-modal mutual trust verification mechanism, filtering out false matching results caused by environmental interference, thereby improving the access control security of fire-resistant curtain wall emergency passages in complex and extreme environments. Through the above processing, the access control tightness of safe areas within the building can be ensured while maintaining escape efficiency.

[0046] More specifically, in a concrete example of this application, the scores corresponding to the visual and non-visual modalities in the preliminary calibration similarity are first extracted, and the absolute value of the difference between them is calculated to quantify the degree of cross-modal consistency deviation. Then, a logical consistency penalty term is constructed based on this absolute value of difference. Using a hyperbolic tangent function, the absolute value of the difference is scaled by a consistency sensitivity coefficient and mapped to a penalty factor in the range of 0 to 1. The larger the difference, the closer the penalty factor tends to 1; the smaller the difference, the closer the penalty factor tends to 0. Finally, the penalty factor is applied to the original scores of each modality in the preliminary calibration similarity in a multiplicative manner. The score of each modality is multiplied by the result of subtracting the penalty factor, thus obtaining the final corrected score after cross-modal consistency calibration. The final corrected scores of each modality together constitute the original modal component set. This process can be expressed as: in, The final corrected score is the value generated from the original modal component set. To initially calibrate the original scores in the similarity, and These are the preliminary similarity scores for the visual modality and the non-visual modality, respectively. The consistency sensitivity coefficient, It is the hyperbolic tangent function. This represents the absolute value of the difference between the visual modality score and the non-visual modality score. In the above formula, Cross-modal differences were analyzed using a sensitivity coefficient. After magnification, the values ​​are compressed to the 0-1 range using the hyperbolic tangent function as a penalty factor. When the scores of the visual and non-visual modalities are highly consistent, the absolute value of the difference tends to 0, the penalty factor tends to 0, and the correction coefficient... When the scores of two modes tend towards 1, the original scores of each mode are almost unaffected. When there is a sharp conflict between the scores of two modes, the absolute value of the difference increases, the penalty factor tends towards 1, the correction coefficient tends towards 0, and the final corrected scores of each mode are suppressed synchronously. In fire evacuation routes, when the flashing of fire light causes the visible light sensor to generate false feature matching on the face of the escaping person, resulting in a high score for the visual mode and a low score for the thermal imaging mode due to the failure to detect the matching thermal feature distribution, the absolute value of the difference between the two is large. After mapping with the hyperbolic tangent function, the penalty factor tends towards 1, the correction coefficient tends towards 0, and the final corrected score is quickly suppressed, thereby avoiding false authorization caused by fire light interference and ensuring the tightness of access control in the safe area.

[0047] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the multi-dimensional confidence weighted fusion unit 143 is used to perform multi-dimensional confidence weighted fusion of the original modal component set using modal weight distribution to obtain a fusion confidence score. It should be noted that since the final corrected scores of each modality in the original modal component set have undergone cross-modal consistency calibration, but the confidence levels of each modality still differ under the current fire environment, simply summing or averaging the modal scores with equal weights will not reflect the dynamic changes in the confidence levels of each modality driven by the environmental noise index, potentially leading to unreasonable contributions from damaged modalities to the final judgment result. Therefore, the technical solution of this application further utilizes modal weight distribution to perform multi-dimensional confidence weighted fusion of the original modal component set to obtain a fusion confidence score. Through the above processing, the scores of each modality after uncertainty quantification, kernel mapping reconstruction and cross-modal consistency calibration can be fused according to the weight ratio of environmental adaptation, and a scalar score that comprehensively reflects the confidence level of multimodal identity matching can be output for subsequent decision threshold determination.

[0048] More specifically, in a concrete example of this application, the visual modal weights and non-visual modal weights in the modal weight distribution are multiplied modally by the final corrected scores of the corresponding modalities in the original modal component set, and then the weighted scores of each modality are summed to obtain the fusion confidence score. In evacuation scenarios with high fire intensity, the environmental noise index is high, and the non-visual modal weights account for a large proportion in the modal weight distribution. The fusion confidence score is more dominated by the corrected scores of the thermal imaging modality, which is less affected by smoke. In scenarios where smoke has not yet spread in the early stages of a fire, the environmental noise index is low, and the visual modal weights account for a large proportion. The fusion confidence score relies more on the rich facial texture feature matching results in the visible light modality. Thus, the fusion confidence score comprehensively represents the overall confidence level of multimodal identity matching in the form of a single scalar. Its value ranges from 0 to 1, with a higher value indicating a higher degree of identity matching between the person to be identified and the registered template.

[0049] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the adaptive decision threshold dynamic calculation module 150 is used to adaptively calculate the decision threshold based on the environmental noise index and a preset basic safety threshold to obtain the decision threshold. It should be noted that, given the dynamic evolution of fire environment intensity over time, if a fixed basic safety threshold is used as the criterion for access authorization, safety control may be relaxed due to a low threshold in the early stages of a fire when environmental interference is relatively mild, while the threshold may be too high during the intense fire phase when smoke concentration and temperature rise sharply, hindering the emergency evacuation of legitimate personnel. A dynamic balance between safety control and escape efficiency cannot be achieved. Therefore, the technical solution of this application further uses the environmental noise index to adaptively calculate the decision threshold based on the preset basic safety threshold to obtain the decision threshold. Specifically, the environmental noise index and the basic safety threshold are modified and mapped using an exponential decay function through an escape urgency coefficient to obtain a dynamic decay threshold, and a maximum value screening function is used to anchor the dynamic decay threshold to the preset minimum safety lower limit threshold. Through the above processing, the decision threshold can be adaptively lowered as the fire environment deteriorates, thereby reducing the access threshold for authorization and prioritizing escape efficiency. At the same time, the boundary anchoring of the minimum safety lower limit threshold ensures that the decision threshold will not be lowered indefinitely, maintaining the basic safety baseline of access control.

[0050] Figure 6 This is a block diagram of the adaptive decision threshold dynamic calculation module in a biometric-based fire-resistant curtain wall emergency passage control system according to an embodiment of this application. Figure 6As shown, the adaptive decision threshold dynamic calculation module 150 includes: a correction mapping unit 151, used to correct and map the environmental noise index and the basic safety threshold based on the escape urgency coefficient using an exponential decay function to obtain a dynamic decay threshold; and a boundary anchoring processing unit 152, used to perform boundary anchoring processing on the dynamic decay threshold and the preset minimum safety lower limit threshold using a maximum value screening function to obtain a decision threshold.

[0051] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the correction mapping unit 151 is used to correct and map the environmental noise index and the basic safety threshold using an exponential decay function based on the escape urgency coefficient to obtain a dynamic decay threshold. It should be noted that since there is a positive correlation between the degree of deterioration of the fire environment and the urgency of personnel escape, a higher environmental noise index means a greater fire intensity and a more pressing evacuation window. Maintaining a high safety threshold at this time would increase the probability of false rejection due to decreased recognition accuracy, hindering the emergency evacuation of legitimate personnel. Therefore, a mapping mechanism that can smoothly lower the threshold according to the degree of environmental deterioration is needed. Based on this, the technical solution of this application further corrects and maps the environmental noise index and the basic safety threshold using an exponential decay function based on the escape urgency coefficient to obtain a dynamic decay threshold. Through the above processing, the safety threshold can be non-linearly decayed as the fire environment deteriorates. In the early stages of a fire, when the environmental noise index is low, the threshold is close to the basic safety threshold to maintain strict safety control. During the intense fire phase, when the environmental noise index rises, the threshold is lowered more rapidly to prioritize personnel evacuation efficiency.

[0052] More specifically, in a specific example of this application, a preset basic safety threshold and an escape urgency coefficient are first obtained. The basic safety threshold is the minimum confidence score required for passage authorization under normal environmental conditions. The escape urgency coefficient is a pre-calibrated positive real number parameter used to control the rate at which the threshold decreases with the degree of environmental deterioration. The larger the escape urgency coefficient, the more sensitive the threshold is to changes in the environmental noise index. Subsequently, the product of the environmental noise index and the escape urgency coefficient is used as the independent variable of the exponential decay function, and the basic safety threshold is used as the initial amplitude of the decay function to perform exponential decay calculation. When the environmental noise index tends to 0, the output of the exponential decay function tends to the basic safety threshold itself, and the threshold hardly decays. As the environmental noise index gradually increases, the escape urgency coefficient amplifies the effect of the environmental noise index, and the output of the exponential decay function decreases at an accelerated rate in a non-linear manner, causing the dynamic decay threshold to fall below the basic safety threshold, with the reduction increasing as the environmental noise index increases. In fire evacuation routes, as smoke concentration and ambient temperature rise rapidly, causing the environmental noise index to increase quickly from a low value, the dynamic attenuation threshold decreases accordingly. This makes it easier for the fusion confidence score to exceed the threshold, triggering access authorization and shortening the waiting time for evacuees in front of the route. The resulting dynamic attenuation threshold is then used in subsequent boundary anchoring processing.

[0053] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the boundary anchoring processing unit 152 is used to perform boundary anchoring processing on the dynamic attenuation threshold and the preset minimum safety lower limit threshold using a maximum value screening function to obtain a decision threshold. It should be noted that since the dynamic attenuation threshold continuously decreases with the increase of the environmental noise index, in extreme scenarios with extremely high fire intensity, the dynamic attenuation threshold may attenuate to a level close to 0. If no lower limit constraint is set for it, almost any confidence score can pass the threshold judgment, thereby triggering access authorization and allowing unverified personnel to enter the safe area, thus losing the basic protective capability of access control. Based on this, the technical solution of this application further utilizes a maximum value screening function to perform boundary anchoring processing on the dynamic attenuation threshold and the preset minimum safety lower limit threshold to obtain a decision threshold. Through the above processing, it can be ensured that the decision threshold is not lower than the preset safety bottom line under any fire intensity condition, maintaining the basic safety constraints of access control while prioritizing escape efficiency.

[0054] More specifically, in a concrete example of this application, the dynamic attenuation threshold and the preset minimum safety lower limit threshold are simultaneously input into the maximum value filtering function, and the larger of the two is taken as the final decision threshold. The minimum safety lower limit threshold is a fixed constant pre-calibrated according to building safety management regulations, representing the minimum identity matching confidence requirement allowed for access authorization under any environmental conditions. When the dynamic attenuation threshold is higher than the minimum safety lower limit threshold, the maximum value filtering function outputs the dynamic attenuation threshold itself, and the decision threshold maintains adaptive adjustment characteristics as the environmental noise index changes; when the fire intensity is extremely high, causing the dynamic attenuation threshold to decay to below the minimum safety lower limit threshold, the maximum value filtering function outputs the minimum safety lower limit threshold, and the decision threshold is anchored at the safety baseline and no longer decreases. In fire evacuation routes, even if the smoke concentration and ambient temperature reach extreme levels, the decision threshold is still anchored above the confidence threshold specified by the minimum safety lower limit threshold, ensuring that only personnel with a fusion confidence score exceeding this threshold can be granted access authorization. The resulting decision threshold is used in subsequent access authorization and control feedback processes.

[0055] In the aforementioned biometric-based fire-resistant curtain wall emergency passage control system 100, the access authorization and control feedback module 160 is used to determine the personnel access authorization status based on the threshold exceedance judgment result of the fusion confidence score relative to the decision threshold, and to generate control feedback for driving the fire-resistant curtain wall's electric control push rod in conjunction with the fire alarm trigger signal. It should be noted that since the calculation of both the fusion confidence score and the decision threshold has been completed, the comparison result needs to be converted into a clear access authorization command to ultimately drive the physical actuator of the fire-resistant curtain wall to complete the passage opening and closing action. Furthermore, in emergency situations where a fire alarm has been triggered but biometric recognition has not yet been completed or has failed, an emergency escape route must still be maintained. Without this fallback mechanism, the escape route may remain closed during a fire due to a failure in the recognition link. Based on this, the technical solution of this application further determines the personnel access authorization status based on the threshold exceedance judgment result of the fusion confidence score relative to the decision threshold, and generates control feedback for driving the fire-resistant curtain wall's electric control push rod in conjunction with the fire alarm trigger signal. Through the above processing, the fusion judgment result of multimodal identity recognition can be logically integrated with the fire emergency signal to output a control signal that can directly drive the physical actuator, and the reliable execution of the push rod action can be ensured through closed-loop monitoring.

[0056] More specifically, in a concrete example of this application, the above-mentioned control feedback generation process includes the following processing stages. First, a sign function is used to determine the polarity of the difference between the fusion confidence score and the decision threshold to obtain the authorization logic state. The difference is obtained by subtracting the decision threshold from the fusion confidence score. This difference is then input into the sign function for polarity determination. When the difference is positive, i.e., the fusion confidence score is higher than the decision threshold, the sign function outputs a positive value, and the authorization logic state is set to the authorization pass state. When the difference is negative or zero, i.e., the fusion confidence score has not reached the decision threshold, the sign function outputs a non-positive value, and the authorization logic state is set to the authorization rejection state.

[0057] Subsequently, a logical OR operation is performed on the authorization logic state and the fire alarm trigger signal to obtain the final access decision. The fire alarm trigger signal is a binary signal from the fire alarm controller; it is valid when a fire has been triggered and invalid when no fire has been triggered. The authorization logic state and the fire alarm trigger signal are logically ORed. If either is valid, the final access decision is to allow passage; otherwise, it is to deny passage. Even if the biometric link fails to output an authorization status due to extremely harsh environmental conditions in fire evacuation routes, as long as the fire alarm controller has issued a fire alarm trigger signal, the final access decision will still be to allow passage, ensuring that escape routes are not blocked due to identification failure.

[0058] Furthermore, a pulse width modulation (PWM) signal is generated based on the final access decision, and closed-loop monitoring is performed using the push rod stroke current to generate control feedback. When the final access decision is to allow passage, a PWM signal with a duty cycle corresponding to the full extension of the push rod is generated to drive the fire-resistant curtain wall's electrically controlled push rod to perform an opening action; when the final access decision is to deny passage, a PWM signal with a duty cycle corresponding to the retraction of the push rod is generated to maintain the closed state of the fire-resistant curtain wall. During the push rod's action, the push rod stroke current is collected in real time as a feedback signal, and the actual stroke current is compared with the expected current curve. When the stroke current deviates from the expected range, the push rod action is judged to be abnormal and a retry or alarm is triggered. When the stroke current matches the expected curve and the push rod reaches the target position, the action is confirmed to be completed. This generates control feedback that includes the push rod action status and execution result, completing the entire closed-loop control from identity recognition and determination to the opening and closing of the physical passage.

[0059] In summary, the biometric-based emergency passage control system for fire-resistant curtain walls according to the embodiments of this application is explained. First, it performs time-stamp synchronization and multispectral spatial coordinate registration on multi-source sensor data such as video streams, thermal imaging signals, smoke concentration, and ambient temperature to form a synchronized data packet. Then, it quantifies and evaluates smoke concentration and ambient temperature to obtain an environmental noise index characterizing the severity of the current fire environment. Based on this index, it performs adaptive defogging reconstruction and contrast enhancement on visual feature areas to extract refined feature vectors. In the modal fusion stage, the system dynamically allocates the weights of each modality based on the environmental noise index. Through feature uncertainty quantification and nonlinear similarity reconstruction using Gaussian kernel mapping, it solves the high false rejection rate and passage deadlock problems caused by nonlinear distortion of features in a fire environment. Furthermore, it introduces cross-modal consistency calibration penalty logic to address the risk of false matching recognition caused by environmental interference. Finally, it dynamically adjusts the decision threshold based on the environmental noise index and drives the fire-resistant curtain wall's electric control push rod in conjunction with the fire alarm trigger signal, ensuring reliable opening of escape passages while maintaining the tightness of access control in the building's safe areas.

Claims

1. A fire-resistant curtain wall emergency passage control system based on biometric recognition, characterized in that, include: The data synchronization and registration module is used to perform timestamp synchronization and multispectral spatial coordinate registration on the acquired sensor data sources to obtain synchronization data packets. The sensor data sources include video streams, thermal imaging signals, smoke concentration, and ambient temperature. The environmental interference factor quantification and evaluation module is used to quantify the environmental interference factors of smoke concentration and ambient temperature in the synchronization data packet to obtain the environmental noise index. The biometric robustness preprocessing and enhancement module is used to perform biometric robustness preprocessing and enhancement on visual feature regions in synchronization data packets based on the environmental noise index to obtain refined feature vectors. The modal weight dynamic allocation module is used to map the associated weight coefficients based on the environmental noise index to obtain the modal weight distribution, and to calculate the modal component similarity between the refined feature vector and the pre-stored registered template features based on the environmental noise index to obtain the original modal component set. The modal weight distribution is used to perform multi-dimensional confidence weighted fusion on the original modal component set to obtain the fusion confidence score. The adaptive decision threshold dynamic calculation module is used to dynamically calculate the decision threshold based on the preset basic safety threshold according to the environmental noise index to obtain the decision threshold. The access authorization and control feedback module is used to determine the personnel access authorization status based on the threshold exceedance judgment result of the fusion confidence score relative to the decision threshold, and to generate control feedback for driving the fireproof curtain wall electric control push rod in combination with the fire alarm trigger signal.

2. The fire-resistant curtain wall emergency passage control system based on biometric recognition according to claim 1, characterized in that, The data synchronization and registration module includes: The hardware-level interrupt capture and frame encapsulation processing unit is used to perform hardware-level interrupt capture and frame encapsulation processing on the sensor data source to obtain a buffer stream array. The clock reference time offset correction unit is used to perform clock reference time offset correction on the buffer stream array based on the master clock to obtain a time alignment matrix; The multispectral spatial coordinate registration unit is used to perform multispectral spatial coordinate registration of the visible light pixel coordinates and thermal imaging pixel coordinates in the time alignment matrix using a pre-calibrated affine transformation matrix to obtain spatiotemporal registration data. The splicing and encapsulation unit is used to splice and encapsulate the image pixel matrix and physical values ​​in the spatiotemporal registration data to generate synchronization data packets.

3. The fire-resistant curtain wall emergency passage control system based on biometric recognition according to claim 1, characterized in that, The environmental disturbance factor quantitative assessment module includes: The parsing and extraction processing unit is used to parse and extract the synchronization data packet based on a preset bit offset to obtain the original environmental sampling value composed of smoke concentration value and ambient temperature value. An environmental anomaly normalization unit is used to normalize the original environmental sampling values ​​based on preset limit values ​​to obtain normalized interference features that include normalized smoke factors and normalized temperature factors. The environmental noise index weighted synthesis unit is used to perform environmental noise index weighted synthesis on the normalized interference characteristics to obtain the environmental noise index.

4. The fire-resistant curtain wall emergency passage control system based on biometric recognition according to claim 1, characterized in that, The biometric robustness preprocessing and enhancement module includes: An environment-adaptive dehazing reconstruction unit is used to perform environment-adaptive dehazing reconstruction on video frames in a synchronization data packet based on the environmental noise index to obtain a repaired intermediate data packet. An adaptive histogram equalization processing unit is used to dynamically adjust the shearing limit threshold in the repair intermediate data packet based on the environmental noise index, and to perform contrast-limited adaptive histogram equalization processing on the thermal imaging signal to obtain an enhanced multimodal image group. The feature mapping and attention-weighted pooling processing unit is used to perform feature mapping and attention-weighted pooling processing on the enhanced multimodal image group using a pre-trained convolutional neural network to obtain refined feature vectors.

5. The fire-resistant curtain wall emergency passage control system based on biometric recognition according to claim 1, characterized in that, The modal weight dynamic allocation module includes: The associated weight coefficient mapping unit is used to perform associated weight coefficient mapping in a preset weight matrix based on the environmental noise index to obtain a modal weight distribution that includes visual modal weights and non-visual modal weights; The modal component similarity calculation unit is used to perform modal component similarity calculation on the refined feature vector and the pre-stored registered template features based on the environmental noise index to obtain the original modal component set; The multidimensional confidence weighted fusion unit is used to perform multidimensional confidence weighted fusion of the original modal component set using modal weight distribution to obtain the fusion confidence score.

6. The fire-resistant curtain wall emergency passage control system based on biometric recognition according to claim 1, characterized in that, The adaptive decision threshold dynamic calculation module includes: The correction mapping unit is used to correct and map the environmental noise index and the basic safety threshold based on the escape urgency coefficient using an exponential decay function to obtain the dynamic decay threshold. The boundary anchoring processing unit is used to perform boundary anchoring processing on the dynamic decay threshold and the preset minimum safety lower limit threshold using the maximum value screening function to obtain the decision threshold.

7. The fire-resistant curtain wall emergency passage control system based on biometric recognition according to claim 1, characterized in that, The access authorization and control feedback module includes: The difference polarity determination processing unit is used to perform difference polarity determination processing on the fusion confidence score and decision threshold using a sign function to obtain the authorization logic state; The logic OR operation processing unit is used to perform a logic OR operation on the authorization logic state and the fire alarm trigger signal to obtain the final access decision; The closed-loop monitoring and processing unit is used to generate a pulse width modulation signal based on the final access decision and to perform closed-loop monitoring and processing using the push rod stroke current to generate control feedback.

8. The fire-resistant curtain wall emergency passage control system based on biometric recognition according to claim 5, characterized in that, The modal component similarity calculation processing unit includes: The exponential weight gain processing subunit is used to perform exponential weight gain processing on the local entropy values ​​of the feature space components of the refined feature vector based on the environmental noise index to obtain the feature uncertainty matrix. The nonlinear mapping alignment processing subunit is used to perform nonlinear mapping alignment processing on the refined feature vectors using the feature uncertainty matrix as the kernel radius adjustment parameter of the Gaussian kernel function to obtain the preliminary calibration similarity. The smoothing correction processing subunit is used to perform smoothing correction processing on the preliminary calibration similarity based on cross-modal consistency penalty logic, based on the difference between the visual modal scores and non-visual modal scores in the preliminary calibration similarity, to obtain the original modal component set.

Images