A curtain wall laboratory hazard source identification method, system, device and medium

Abstract

The application relates to a curtain wall laboratory hazard source identification method, system, device and medium, and relates to the technical field of curtain wall monitoring and safety warning. The method comprises the following steps: extracting a character feature based on a personnel image entering an operation area, generating a qualification key, matching the qualification key with a personnel qualification library, verifying whether the personnel have operation permission in the operation area; if yes, generating a standard key based on a safety standard template of the operation area, and generating an operation key according to real-time operation characteristics of the personnel; within the effective period of the standard key, comparing the standard key with the operation key, and if the comparison result exceeds a preset deviation range, generating alarm information of a personnel identity corresponding to an associated face feature. The application can realize dynamic binding and active warning of personnel permission, operation behavior and spatial position, and improve the accuracy and traceability of safety management.

Description

Technical Field

[0001] This application relates to the field of curtain wall monitoring and safety early warning technology, and in particular to a method, system, equipment and medium for identifying hazards in a curtain wall laboratory. Background Technology

[0002] With the continuous development of the construction industry, the importance of building curtain wall testing laboratories is becoming increasingly prominent. These laboratories often involve high-risk operations such as working at heights, hot work, and hoisting, placing extremely high demands on ensuring personnel safety and the smooth conduct of tests. Their development is not only related to the quality testing and performance evaluation of building curtain walls, but also closely linked to the safety and sustainable development of the entire construction industry. Effective safety monitoring technology can provide reliable guarantees for the research and development and production of building curtain walls, propelling the construction industry towards higher levels.

[0003] Previously, to address the safety monitoring issues in building curtain wall laboratories, existing monitoring systems typically employed cameras combined with behavior recognition algorithms. These systems relied primarily on cameras to capture video images, followed by behavior recognition algorithms to analyze and assess the behavior of personnel within the images, identifying violations such as not wearing safety helmets or seatbelts. Cameras covered various areas of the laboratory, recording personnel activities in real time, while the behavior recognition algorithms processed the captured video to identify potential safety hazards.

[0004] However, existing monitoring systems have many shortcomings. They suffer from a disconnect between identity and behavior; while they can identify violations such as not wearing seatbelts, they cannot pinpoint the specific individual responsible, making it difficult to trace the perpetrator or differentiate monitoring based on personnel qualifications. Furthermore, current monitoring strategies generally employ the same behavior recognition model across all work areas, lacking dynamic activation mechanisms for different work types, thus limiting recognition accuracy and efficiency. In specific scenarios, such as curtain wall testing laboratories where numerous objects cause interference leading to false identifications, existing systems lack targeted anti-interference mechanisms. Summary of the Invention

[0005] The purpose of this application is to provide a method for identifying hazards in a curtain wall testing laboratory, which can dynamically bind personnel permissions, work behaviors and spatial locations and provide proactive early warnings, thereby improving the accuracy and traceability of safety management.

[0006] Firstly, this application provides a method for identifying hazard sources in a curtain wall testing laboratory, which adopts the following technical solution: A method for identifying hazard sources in a curtain wall testing laboratory includes: Based on the images of personnel entering the work area, human features are extracted, and qualification keys are generated based on the human features. The qualification key is matched with the personnel qualification database to verify whether the personnel have the work permission for the work area; If the user holds the operation permission for the operation area during the operation process, a standard key is generated based on the security specification template of the operation area. Based on the real-time operational characteristics of the personnel, an operational key is generated, wherein the operational characteristics include location characteristics, action characteristics, tool characteristics, and status characteristics; During the validity period of the specification key, the specification key is compared with the job key; If the comparison result exceeds the preset deviation range, an alarm message is generated, and the alarm message is associated with the identity of the person corresponding to the facial feature.

[0007] By adopting the above technical solution, the qualification key, bound to personnel identity, provides a foundation for subsequent verification of personnel's work permissions; matching the qualification key with the personnel qualification database to verify personnel's work permissions ensures that personnel entering the work area have the corresponding qualifications, guaranteeing work safety; if personnel hold work permissions, a standardized key is generated, enabling specific monitoring of the standards for different work areas; and a work key generated based on personnel's real-time work characteristics reflects the personnel's work status in real time; comparing the standardized key with the work key during the validity period of the standardized key can promptly detect whether personnel's work behavior complies with the standards; if the comparison result exceeds the preset deviation range, an alarm message is generated, allowing for precise tracing of the responsible party for violations. Ultimately, this achieves dynamic binding and proactive early warning of personnel permissions, work behavior, and spatial location, improving the accuracy and traceability of safety management.

[0008] This application is further configured as follows: the step of extracting human features from images of personnel entering the work area and generating qualification keys based on the human features includes: Based on the results of reflection interference detection of the personnel image, it is identified and determined whether the personnel image has a virtual image region; If it is determined that the personnel image contains a virtual image area, then the virtual image area is blocked, and an instruction is sent to switch to the associated adjacent view camera to re-acquire the personnel image until a real personnel image without virtual image interference is acquired. By using a deep learning face recognition model, the face portion of the real person image is mapped to a face feature vector; The gait recognition model extracts gait features from the real person image and maps them into gait feature vectors. The gait features include gait cycle and skeleton dynamic features. The facial feature vector and gait feature vector are combined and encrypted to generate the qualification key.

[0009] By adopting the above technical solution, reflection interference detection is first performed to avoid the influence of virtual images, so as to obtain the most realistic personnel images and improve the accuracy of feature extraction. Then, deep learning and face and gait recognition models are used to extract feature vectors, which can accurately characterize personnel features. The combined encrypted qualification key can realize personnel permission binding, and finally realize the dynamic binding and proactive warning of personnel permissions, work behavior and spatial location, thereby improving the accuracy and traceability of safety management.

[0010] This application is further configured such that the step of generating a job key based on the real-time job characteristics of the personnel includes: By fusing the results of target detection, skeletal key point recognition and action classification of the personnel through a deep learning task feature model, the task features of the personnel are extracted and mapped into a task feature vector of the personnel. The real-time status characteristics of the curtain wall test panel are obtained and mapped to the curtain wall operation feature vector. The feature vectors of the human operation and the feature vectors of the curtain wall operation are combined and encrypted to generate the operation key.

[0011] By adopting the above technical solution, the results of personnel target detection, skeletal key point recognition and action classification are first integrated using a deep learning operation feature model to extract personnel operation features and map them into personnel operation feature vectors. At the same time, the real-time status features of the curtain wall test panel are obtained and mapped into curtain wall operation feature vectors. Then, the two are combined and encrypted to generate an operation key. This enables a comprehensive consideration and accurate characterization of personnel operation behavior and curtain wall status, which is conducive to dynamically binding personnel permissions, operation behavior and spatial location. This allows for timely detection of anomalies in personnel behavior and curtain wall status during operation, enabling proactive early warning and improving the accuracy and traceability of safety management.

[0012] This application is further configured as follows: the step of generating a specification key based on the security specification template of the work area if the user holds the work area's work permission during the work process includes: The spatial distance between personnel and each curtain wall test panel in the image is calculated in real time, and the curtain wall test panel with the shortest spatial distance is set as the working panel; Obtain the identification information of the work module, and retrieve the safety specification template corresponding to the work module based on the identification information of the work module; Convert the security specification template into a specification feature vector; The standardized feature vector is encrypted to generate the standardized key.

[0013] By adopting the above technical solution, the spatial distance between personnel and curtain wall test panel is calculated in real time to determine the work panel, which can accurately associate personnel with the work space location; the identification information of the work panel is obtained and the corresponding safety specification template is retrieved, which can realize the matching of work behavior with specifications; finally, the specification key is generated by encrypting the specification feature vector to ensure information security, realizing the dynamic binding and proactive early warning of the three, and improving the accuracy and traceability of safety management.

[0014] This application is further configured such that: the step of comparing the specification key with the job key during the validity period of the specification key includes: The canonical key is decrypted to obtain the canonical feature vector; The job key is decrypted to obtain the job feature vector; A transmittance level is set based on the transmittance of the curtain wall specimen in the current work area, and the confidence weight of each dimension in the work feature vector is dynamically adjusted according to the transmittance level. Based on the adjusted weights, the weighted deviation between the job feature vector and the standard feature vector is calculated. If the weighted deviation exceeds a preset threshold, the comparison result is determined to be outside the preset deviation range.

[0015] By adopting the above technical solutions, standard feature vectors can be obtained by decrypting the standard key and operation key, and operation feature vectors can be obtained by decrypting the operation key. This allows for accurate understanding of safety standards and real-time operation status. Adjusting the weight of the operation feature vector based on light transmittance can adapt to environmental changes. Calculating the weighted deviation and determining the deviation can promptly detect anomalies, achieve dynamic binding and proactive early warning, and improve the accuracy and traceability of safety management.

[0016] This application is further configured as follows: the step of generating an alarm message if the comparison result exceeds a preset deviation range, wherein the alarm message is associated with the identity of the person corresponding to the facial feature, includes: During the validity period of the specified key, the spatial distance between the personnel and the preset hazard source, safety equipment or monitoring personnel is calculated in real time, wherein the qualification key of the monitoring personnel needs to be verified by whether they hold the work permission of the work area; The spatial distance is compared with the corresponding security threshold; If the spatial distance exceeds the safety threshold, a distance warning is triggered, and the calculation frequency of the spatial distance is dynamically adjusted according to the risk level of the area where the person is located.

[0017] By adopting the above technical solutions, the spatial distance between personnel and preset hazards, safety equipment, or monitoring personnel can be calculated in real time, thereby enabling the determination of personnel's spatial location. By comparing the distance with safety thresholds, dangerous situations can be detected in a timely manner and warnings can be triggered. The calculation frequency can be adjusted according to the risk level to enhance the timeliness of warnings, realize the dynamic binding of personnel permissions, behaviors, and locations, and proactive warnings, thereby improving the accuracy and traceability of safety management.

[0018] This application is further configured such that, after the step of generating an alarm message if the comparison result exceeds a preset deviation range, and the alarm message is associated with the identity of the person corresponding to the facial feature, the application further includes: Real-time acquisition of facial features and gait features of the current person; The real-time facial features are compared with the facial features in the qualification key, and the first similarity is calculated and obtained; If the first similarity is lower than a preset first threshold, it is determined to be a suspected occlusion, triggering gait-assisted verification; The real-time gait features are compared with the gait features in the qualification key to calculate and obtain the second similarity. If the second similarity is higher than the preset second threshold, the identity verification status is maintained and the face occlusion event is recorded. If the second similarity is also lower than the preset threshold, the identity is determined to be lost and an alarm is triggered.

[0019] By adopting the above technical solutions, the latest feature information of personnel can be obtained in real time by acquiring facial and gait features. By comparing facial features and calculating similarity, facial occlusion can be detected in a timely manner. Triggering gait-assisted verification can maintain identity verification when the face is occluded. Comparing gait features can further confirm the identity, thus realizing continuous and accurate verification and traceability of personnel identity and ensuring reliable security management.

[0020] Secondly, this application provides a hazard identification system for curtain wall testing laboratories, employing the following technical solution: A hazard identification system for a curtain wall testing laboratory includes: Thirdly, this application provides an electronic device that adopts the following technical solution: Qualification Acquisition Module: Used to extract human features from images of personnel entering the work area and generate qualification keys based on the human features; Qualification verification module: used to match the qualification key with the personnel qualification database to verify whether the personnel have the work permission for the work area; Specification Acquisition Module: If the module holds the work permission for the work area during the work process, it generates a specification key based on the security specification template of the work area. The job monitoring module is used to generate job keys based on the real-time job characteristics of the personnel, wherein the job characteristics include location characteristics, action characteristics, tool characteristics, and status characteristics. Security determination module: used to compare the standard key with the job key during the validity period of the standard key; Security alarm module: If the comparison result exceeds the preset deviation range, an alarm message is generated, and the alarm message is associated with the identity of the person corresponding to the facial feature.

[0021] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for identifying hazards in a curtain wall laboratory.

[0022] Fourthly, this application provides a computer storage medium, as follows: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for identifying hazards in a curtain wall laboratory.

[0023] In summary, this application has the following beneficial technical effects: 1. This application can solve the problem of the disconnect between identity and behavior, achieve accurate identity tracing, and link violations to specific individuals, facilitating responsibility determination and safety education; 2. This application can effectively solve the problem of false image recognition caused by reflection of curtain wall glass through reflection interference detection and filtering mechanism, and ensure the accuracy of feature extraction; 3. This application can introduce a dynamic activation mechanism to generate job keys and standard keys for different job types, perform differentiated monitoring, and improve identification accuracy and efficiency; 4. This application enables closed-loop management of the entire process, covering personnel entry, operation, and departure, and records relevant data, which facilitates the generation of safety reports and backtracking optimization; 5. This application can improve the robustness of identity verification and the recognition accuracy under complex lighting conditions in the curtain wall laboratory by using multimodal feature verification and dynamic weight adjustment. Attached Figure Description

[0024] Figure 1 This is a flowchart of a hazard source identification method for a curtain wall laboratory according to one embodiment of this application.

[0025] Figure 2 This is a flowchart of a sub-step of step S1 in one embodiment of this application.

[0026] Figure 3This is a flowchart of a sub-step of step S4 in one embodiment of this application.

[0027] Figure 4 This is a flowchart of a sub-step of step S3 in one embodiment of this application.

[0028] Figure 5 This is a flowchart of a sub-step of step S5 in one embodiment of this application.

[0029] Figure 6 This is a flowchart of a sub-step of step S6 in one embodiment of this application.

[0030] Figure 7 This is a flowchart of the steps added after step S6 in one embodiment of this application.

[0031] Figure 8 This is a schematic diagram of the structure of a hazard identification system for a curtain wall laboratory according to one embodiment of this application.

[0032] Figure 9 This is a schematic block diagram of an electronic device in one embodiment of this application.

[0033] Attached reference numerals: 1. Qualification acquisition module; 2. Qualification verification module; 3. Standard acquisition module; 4. Operation monitoring module; 5. Safety judgment module; 6. Safety alarm module. Detailed Implementation

[0034] The following is in conjunction with the appendix Figure 1-9 This application will be described in further detail.

[0035] It should be noted that, in the embodiments of this invention, the data related to object information, when applied to specific products or technologies, requires the permission or consent of the object. Furthermore, the collection, use, and processing of this data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. In other words, if the embodiments of this invention involve data related to an object, it must be obtained with the object's authorization and consent, the authorization and consent of relevant departments, and in accordance with the relevant laws, regulations, and standards of the country and region. If the embodiments involve personal information, the acquisition of all personal information requires the individual's consent. If sensitive information is involved, the separate consent of the information subject is required. The embodiments also need to be implemented with the object's authorization and consent.

[0036] refer to Figure 1 A method for identifying hazards in a curtain wall testing laboratory, specifically including: S1. Extract human features from images of personnel entering the work area, and generate qualification keys based on human features.

[0037] Specifically, high-definition camera groups deployed at entrances and key passages collect real-time images of people's faces and bodies. Deep learning models are used to preprocess the images to reduce interference such as uneven lighting and motion blur. Multi-dimensional biometric features, including facial structure features, body contour features, and dynamic behavior features, are then extracted from the images. These features are digitally encoded and encrypted to generate unique and time-sensitive qualification keys.

[0038] In summary, the qualification key, as a digital credential for personnel identity, not only ensures the unique identification of personnel but also provides a reliable anchor point for all subsequent security monitoring links. This enables the system to closely associate each work action with a specific person, solving the problem of the disconnect between identity and behavior in traditional monitoring and laying a solid foundation for achieving accurate traceability and differentiated security management.

[0039] S2. Match the qualification key with the personnel qualification database to verify whether the personnel have the work permission for the work area.

[0040] Specifically, the cloud platform receives qualification keys uploaded by edge nodes in real time, decrypts them to restore personnel characteristic information, and compares it with a pre-established personnel qualification database. This database stores information such as the qualification certificate type, validity period, and authorized work area for all registered personnel. Finally, the system automatically retrieves the corresponding qualification requirements based on the personnel's current location to quickly determine whether the personnel have the legal authority to enter that area or perform specific tasks.

[0041] The above achieves continuous perception and digital representation of personnel's work behavior. The work key is updated every second, and the interaction details between personnel and the work environment are fully recorded, providing the system with real-time and comprehensive situational awareness capabilities at the work site, ensuring that any minute behavioral changes can be captured in a timely manner and incorporated into the monitoring system.

[0042] The above-mentioned work permission verification process can not only prevent unauthorized personnel from entering high-risk areas, but also dynamically identify expired qualifications or changes in permissions. It reduces the safety hazards caused by unqualified personnel from the access control stage, ensuring that only authorized and qualified personnel can enter the corresponding work area, thereby improving the safety management level of the work site.

[0043] S3. If you have the work permission for the work area during the work process, generate a standard key based on the safety standard template of the work area.

[0044] Specifically, after personnel pass qualification verification, the system automatically retrieves the corresponding safety specification template library based on the identifier of their current work area. The template library predefines the standard operating requirements for each area, including permitted action types, required safety equipment, tool usage specifications, and safe distances from hazards. The system converts these specifications into structured digital feature vectors and, combined with the area identifier and personnel characteristics, generates an encrypted specification key, which is then distributed to the edge computing nodes in that area.

[0045] In summary, the standardized key can transform static security procedures into dynamically executable monitoring benchmarks, providing a reference for subsequent real-time behavior comparisons. It also enables differentiated customization of monitoring strategies between different areas, allowing dedicated safety monitoring models to be matched for high-altitude work areas, hot work areas, hoisting areas, etc.

[0046] S4. Generate a job key based on the real-time job characteristics of personnel. The job characteristics include location characteristics, action characteristics, tool characteristics, and status characteristics.

[0047] Specifically, a network of cameras covering the entire work area continuously captures real-time images of personnel. A multi-task deep learning model is used to simultaneously perform target detection, skeletal key point extraction, and action recognition. From this, multi-dimensional information such as the personnel's spatial coordinates, current action category, type of handheld tool, and status of safety equipment are extracted. These dynamic features are then fused and encrypted to generate a real-time work key.

[0048] The above achieves continuous perception and digital representation of personnel's work behavior, and the work key is updated every second, fully recording the interaction details between personnel and the work environment. This provides the system with real-time and comprehensive situational awareness capabilities at the work site, ensuring that any minute behavioral changes can be captured in a timely manner and incorporated into the monitoring system.

[0049] S5. During the validity period of the standard key, compare the standard key with the job key.

[0050] Specifically, during the validity period of the standard key, edge nodes decrypt the standard key and the operation key in real time to obtain the standard feature vector and the real-time operation feature vector, and calculate the comprehensive deviation degree through a weighted similarity algorithm. When multiple violations exist simultaneously, the system sorts the deviation degree calculation results according to preset alarm priority rules, prioritizing the violation type with higher warning level. If the warning levels are the same, they are responded to in chronological order to ensure clear and orderly on-site information broadcasting.

[0051] In summary, the real-time comparison mechanism enables the system to continuously monitor the operation process, promptly detect and locate safety hazards, provide precise triggering conditions for proactive early warning, and effectively avoid the lag in post-event tracing.

[0052] S6. If the comparison result exceeds the preset deviation range, an alarm message is generated, and the alarm message is associated with the identity of the person corresponding to the facial features.

[0053] Specifically, when the weighted deviation between the real-time operation feature vector and the standard feature vector exceeds a preset threshold, the system immediately generates graded alarm information and provides targeted reminders through on-site voice broadcasting equipment, such as "Please wear a safety helmet when entering the curtain wall test area," "Please immediately fasten your five-point safety belt when working at height," and "Please immediately clear any flammable materials present in the hot work area." The system also triggers audible and visual alarms based on the severity of the violation and pushes alarm details to the administrator terminal, including the identity of the violator, the type of violation, the specific location, and the time.

[0054] For recurring violations, the system will periodically issue reminders at preset intervals until the violation is corrected. Administrators can directly intervene on-site via a remote voice system. For scenarios such as light flickering misinterpreting as sparks or hoisting assistants accidentally entering the wrong area, the system performs multi-source collaborative verification before triggering an alarm. This involves retrieving personnel identification and tool recognition results. If the person lacks welding qualifications and is not holding a welding torch, it is determined to be flickering interference. Alternatively, an alarm may be triggered based on the exemption location marked by the assistant. The system only triggers an alarm after confirmation, effectively reducing the false alarm rate. False alarm examples marked by the administrator will be added to the training set for subsequent model optimization.

[0055] The above achieves a closed loop from the occurrence of a violation to the alarm response, accurately assigning safety management responsibilities to specific individuals, improving the efficiency and traceability of safety supervision, and providing strong technical support for accident prevention and liability determination.

[0056] refer to Figure 2 Furthermore, in one embodiment, step S1 is further refined into the following sub-steps: S10. Based on the results of reflection interference detection of personnel images, identify and determine whether the personnel images form virtual image regions.

[0057] Specifically, the system first acquires real-time images of personnel using high-definition cameras deployed at the entrance of the curtain wall testing laboratory and in various work areas. The system then preprocesses the raw images, including noise reduction and brightness equalization to eliminate the effects of uneven lighting. Subsequently, a multimodal reflection interference detection algorithm is used to perform in-depth analysis of the images, specifically including: First, symmetry detection is performed based on the mirror surface's geometric features. By calculating the centroid coordinates and edge point distribution of candidate regions in the image, it is analyzed whether the region has a mirror structure symmetrical about a certain axis. If a clear axis of symmetry exists and the image content on both sides of the symmetry is highly similar, it is determined to be a virtual image region suspected to be formed by glass reflection.

[0058] Then, virtual image identification is performed based on optical flow field analysis. The motion vector field of pixels is calculated through multiple consecutive frames of images. Real people exhibit a coherent motion trajectory that conforms to physical laws in consecutive frames, while virtual images formed by glass reflections are inverted, and their optical flow direction is opposite to the real motion direction or has an abnormal angle. The system distinguishes between real and virtual images by analyzing the consistency of the optical flow vectors.

[0059] Next, based on depth information, the system confirms the presence of virtual images by calculating the depth value of candidate regions using the spatial relationship jointly calibrated by multiple cameras. If the detected image depth is significantly greater than the real background or does not match the distance to the cameras, it is further confirmed as a virtual image. Combining the above detection results, the system quantifies and scores each suspected region in the image. When the score exceeds a preset threshold, the image of the person is determined to contain a virtual image region.

[0060] Based on the above detection results, the system quantifies and scores each suspected area in the image. When the score exceeds a preset threshold, it is determined that the image of the person contains a virtual image area.

[0061] The above-mentioned system accurately identifies the virtual image interference caused by reflections from curtain wall glass, ensuring that the system extracts features based solely on real human images. This avoids misidentification caused by virtual images, provides a clean image foundation for the accurate extraction of facial and gait features, and enhances the system's adaptability and reliability in complex curtain wall environments.

[0062] S11. If it is determined that the image of a person contains a virtual image area, then the virtual image area is blocked, and an instruction is sent to switch to the camera with the associated adjacent viewpoint to re-acquire the image of the person until a real image of the person without virtual image interference is acquired.

[0063] Specifically, based on the system's pre-defined camera network topology, when a ghost image is detected in an image captured by a camera, the system immediately locates the current camera and triggers adjacent cameras to collaboratively acquire the image. Simultaneously, it automatically masks image areas containing ghost images from the original camera, preventing them from participating in subsequent processing. The system uses multi-camera joint calibration technology to adjust the focal length and exposure parameters of adjacent cameras in real time, ensuring that the recaptured image is clear and free of reflection interference. This is continuously verified until a real person image meeting quality requirements is obtained. This is a feature acquisition method that rapidly acquires sufficiently clear images with easily identifiable facial features within a short time.

[0064] Considering the potential for large, continuous glass reflective surfaces in the curtain wall testing chamber, which could lead to all associated cameras capturing virtual images, a multi-frame temporal analysis and virtual image probability assessment module will be activated to dynamically compare consecutive frames. By calculating the virtual image confidence score of candidate regions in the image sequence, if the virtual image confidence score of a region exceeds a threshold for three consecutive frames, it is determined that the current region has persistent reflection interference. At this point, the system no longer passively waits for images without virtual images but activates a virtual image suppression processing module based on an image restoration algorithm to perform real-time image enhancement and virtual image removal. This module utilizes a generative adversarial network to perform semantic repair on regions containing virtual images. A trained model replaces the virtual image regions with background completion content, thereby obtaining a near-realistic virtual image for feature extraction without relying on camera switching.

[0065] Furthermore, if switching cameras results in incomplete facial features or degraded image quality, the system will activate a gait-based priority verification mode. When key points in the facial region are missing or the resolution is insufficient due to changes in viewing angle, the system automatically increases the weight of gait features in identity verification. Simultaneously, it activates the multi-camera fusion reconstruction module to reconstruct the 3D pose of the same person from different perspectives captured by adjacent cameras, thus completing the missing facial information. If the reconstructed facial features still fail to meet the recognition requirements, the system will temporarily use gait features as the primary basis for identity verification and record the current status as "facial information pending completion." When the person moves to a better viewing angle, the system will automatically acquire a new facial image for feature updates.

[0066] Through the aforementioned multi-level fault-tolerance mechanism, the system ensures that real or equivalent real images of people can be obtained for feature extraction under any extreme conditions, solving the problem of blind spots caused by reflections from curtain wall glass and guaranteeing the continuity and accuracy of identity verification.

[0067] S12. Using a deep learning face recognition model, the face portion of a real person image is mapped to a face feature vector.

[0068] Specifically, this process in a curtain wall testing environment needs to fully consider special factors such as glass reflection, lighting changes, and personnel wearing safety equipment. During implementation, the system first uses a face detector based on a convolutional neural network to locate the face region in real personnel images after virtual image filtering. It employs a multi-scale sliding window and feature pyramid structure to adapt to faces of different sizes in the image. Furthermore, to address the common issue in curtain wall testing environments where personnel wear safety helmets, masks, or face shields that partially obscure the face, the system incorporates a large number of occlusion samples during training to enhance the model's robustness, ensuring accurate facial region definition even when key areas such as the eyes and forehead are visible.

[0069] After locating the face region, the system further performs facial key point alignment. Using a regression network, it accurately locates five to dozens of facial key points, such as the center of the eyes, the tip of the nose, and the corners of the mouth. Based on these key points, it calculates affine transformation parameters to correct the original face image to a standard pose, reducing geometric distortion caused by shooting angle, head rotation, or tilt. Simultaneously, the system performs illumination normalization processing on the corrected face region, employing histogram equalization or adaptive illumination compensation algorithms to suppress localized overexposure or shadow areas caused by reflections from the curtain wall glass, ensuring consistent brightness distribution in the input image.

[0070] The normalized face image is then input into a pre-trained deep face recognition model. This model is typically built on metric learning principles, using loss functions such as ArcFace or CosFace for training. This ensures that facial features of people with the same identity are close together in the feature space, while facial features of people with different identities are far apart. The deep face recognition model consists of multiple convolutional layers, pooling layers, and fully connected layers. It abstracts and extracts low-level edge texture features, mid-level facial structure features, and high-level semantic features of the face layer by layer. Finally, the last fully connected layer of the network outputs a fixed-dimensional deep feature vector, typically 128, 256, or 512 dimensions. Each dimension of this vector represents an abstract representation of the face in a specific implicit dimension. It does not have an intuitive physical meaning, but the whole vector constitutes a unique identity identifier.

[0071] In conclusion, the final generated facial feature vector serves as a core component of the qualification key, providing a precise and unique identity anchor for subsequent personnel authorization verification and behavior tracing.

[0072] S13. Using a gait recognition model, gait features are extracted from real person images and mapped to gait feature vectors. Gait features include gait cycle and skeleton dynamic features.

[0073] Specifically, in the curtain wall laboratory environment, to address situations where personnel may wear masks, safety helmets, or protective face shields, resulting in partial facial occlusion, a gait-based identity supplementation verification mechanism is simultaneously activated. During implementation, the system first extracts multiple consecutive frames of full-body images from a sequence of real personnel images that have undergone virtual image filtering, ensuring the temporal continuity and spatial consistency of the input data. Subsequently, a deep learning-based skeletal keypoint detection network, such as a high-resolution network or a convolutional pose machine, is used to locate the joints of the human body in each frame. Typically, major joints such as the head, shoulders, elbows, wrists, hips, knees, and ankles are detected, and the two-dimensional image coordinates and confidence scores of each joint are output.

[0074] After obtaining the joint coordinates of each frame, the system constructs a spatiotemporal skeleton sequence, linking and tracking the joint trajectories of the same person over continuous time. Based on this spatiotemporal sequence, the system further extracts multidimensional gait features: first, gait periodic features, which calculate the walking cycle duration, the ratio of swing phase to support phase duration, and cadence by analyzing the periodic changes in leg joint angles; second, skeleton dynamic features, including the motion trajectory curves of each joint, the variation of joint angles over time, and the displacement trajectory of the body's center of mass, using these dynamic parameters to characterize an individual's unique walking habits and posture patterns. To ensure the robustness of the features, the system performs smoothing filtering on the extracted original trajectory to eliminate jitter caused by image noise or joint detection errors, and simultaneously eliminates scale differences caused by changes in the distance between the person and the camera through normalization processing.

[0075] To address potential body occlusion issues in the curtain wall testing laboratory, such as lower body obstruction during operation at the workbench, the system employs a local gait feature extraction strategy. It prioritizes dynamically constructing feature vectors based on visible upper body joints, inferring identity information using the relative movements of the shoulders, elbows, and head. Furthermore, the system incorporates an attention mechanism, dynamically adjusting the contribution weight of each joint in the final feature vector based on its confidence score. Joints with lower confidence scores are assigned smaller weights to prevent unreliable information from interfering with the recognition results.

[0076] Finally, the system encodes and fuses the aforementioned multi-dimensional gait features, mapping them through a fully connected layer into a fixed-dimensional gait feature vector, typically 128 or 256 dimensions. This vector is then concatenated with the facial feature vector in subsequent feature layer steps to form a complementary sequence of personal features.

[0077] In summary, the gait feature vector is robust to occlusion, changes in lighting, and changes in viewing angle. It can stably represent the identity of a person without relying on complete facial information, providing a reliable redundancy backup for continuous identity verification in face occlusion scenarios, and significantly enhancing the system's adaptability and reliability in complex operating environments.

[0078] S14. Combine the facial feature vector and gait feature vector and encrypt them to generate an qualification key.

[0079] Specifically, the facial feature vector and gait feature vector are concatenated at the feature layer to generate a human feature sequence containing dual biometric information. The sequence is then encrypted and a timestamp and a random number are added to generate an qualification key.

[0080] After generating the qualification key, the system automatically retrieves the pre-set safety prompt voice template for the area based on the personnel's current work area identifier. The system then delivers a complete safety reminder through the on-site voice broadcasting equipment, covering topics such as wearing safety helmets, working at heights regulations, hot work requirements, hoisting avoidance rules, and material placement requirements. This ensures that personnel entering the area are aware of the area's safety regulations immediately, strengthens their safety awareness, and lays a cognitive foundation for subsequent work monitoring.

[0081] In summary, the complementary fusion of two types of biometric features is achieved through feature layer splicing. Facial features provide detailed facial identity information, while gait features serve as a reliable backup in occluded scenarios. The combination of the two significantly improves the robustness of identity recognition. Encryption ensures the security of feature data during transmission and storage, and the generated qualification key serves as a unique digital identity credential, providing a precise anchor point for subsequent permission verification and behavior tracing.

[0082] In addition, refer to Figure 3 Furthermore, in one embodiment, step S4 is refined into the following sub-steps: S40. By integrating the results of target detection, skeletal key point recognition and action classification of personnel through a deep learning task feature model, the task features of the person are extracted and mapped into a task feature vector.

[0083] Specifically, firstly, based on the real-time video streams captured by high-definition cameras deployed in various work areas of the curtain wall laboratory, a unified work feature extraction model is constructed using a multi-task deep learning network framework. This model shares the underlying convolutional feature extraction layer and integrates multi-scale information through a feature pyramid structure to address scale variations caused by different distances from personnel.

[0084] The upper layer comprises three branches: an object detection branch identifies the location of personnel and the type of handheld tools, outputting the bounding box coordinates of the personnel and the confidence scores of tools such as welding torches and wrenches; a skeletal keypoint recognition branch uses a high-resolution network to regress the two-dimensional coordinates of the main joints of the human body and calculates the joint angles and body orientation; and an action classification branch introduces a temporal convolutional network to model the joint point sequences of multiple consecutive frames, identifying action categories such as walking, climbing, welding, and commanding, as well as their confidence scores. Finally, the system fuses the tool encoding output from object detection, the joint angle vector output from skeletal keypoint recognition, and the action probability distribution output from action classification in a feature layer, mapping them through a fully connected layer to a fixed-dimensional human task feature vector.

[0085] The above describes the personnel's current position, actions, tool usage, and posture information, providing rich and accurate digital evidence for subsequent comparison with the standard template, enabling the system to fully perceive several details of the work behavior.

[0086] S41. Obtain the real-time status characteristics of the curtain wall test panel and map them into a curtain wall operation feature vector.

[0087] Specifically, for the glass curtain wall specimens unique to the curtain wall testing laboratory, real-time images of the curtain wall panels are acquired using the same set of cameras or dedicated monitoring cameras, and dynamic features reflecting the stability of the curtain wall structure are extracted. In this embodiment, the system first uses an edge detection algorithm to extract the boundary contours of the curtain wall panels and locates and tracks the displacement of the panel edges in consecutive frames; simultaneously, it calculates the motion vectors of each pixel on the panel surface based on optical flow field analysis, and statistically analyzes the vibration frequency and amplitude; furthermore, a deep learning-based deformation detection model is introduced to identify geometric distortions in local areas of the panel and outputs a deformation index. Finally, these state parameters are normalized and input into the encoding network, mapped into a curtain wall operation feature vector reflecting the current health status of the curtain wall.

[0088] In summary, by using feature vectors, curtain wall anomalies caused by operational impacts, temperature changes, or unstable installation can be captured in real time, transforming the originally static curtain wall components into dynamically perceptible monitoring objects. This allows the system to not only focus on personnel behavior but also perceive the curtain wall's own response during the interaction between people and the curtain wall, providing a new data dimension for a comprehensive assessment of operational safety.

[0089] S42. Combine the feature vectors of the human operation and the feature vectors of the curtain wall operation and encrypt them to generate an operation key.

[0090] Specifically, the feature vectors of personnel operations and curtain wall operations are concatenated at the feature layer to generate a high-dimensional comprehensive feature sequence that integrates personnel behavior and curtain wall status. This sequence is then encrypted and appended with a timestamp, region identifier, and curtain wall panel number to generate a unique and time-sensitive operation key. Finally, this operation key is distributed to the corresponding edge nodes as a benchmark for real-time monitoring, enabling subsequent standard comparisons to simultaneously consider personnel operations and curtain wall responses, achieving human-machine-environment collaborative monitoring.

[0091] In summary, the operation key can deeply bind dynamic personnel behavior with dynamic curtain wall status, enabling subsequent standard comparisons to simultaneously consider two dimensions: whether the personnel are operating in accordance with regulations and whether the curtain wall is responding abnormally. Once the comprehensive characteristics deviate from the preset safety template, the system can accurately locate whether it is a personnel violation or a curtain wall instability, thereby achieving a higher level of human-machine-environment collaborative monitoring and providing richer information support for proactive early warning and safety traceability.

[0092] In addition, refer to Figure 4 Furthermore, in one embodiment, step S3 is refined into the following sub-steps: S30. Real-time calculation of the spatial distance between personnel and each curtain wall test panel in the image, and setting the curtain wall test panel with the shortest spatial distance as the working panel.

[0093] Specifically, the system first analyzes real-time video images using instance segmentation or object detection models to identify the outline and position of each curtain wall test panel in the image and extract its edge features. Based on this, the system establishes a global spatial coordinate system using multi-camera joint calibration technology. Combining stereo vision principles or monocular depth estimation algorithms, it calculates the 3D spatial coordinates of key points for each panel, such as the four corner points. Simultaneously, it obtains the 3D coordinates of the person's body center point or feet using skeletal keypoint recognition technology. Then, it calculates the Euclidean distance between the person and each panel's key points, taking into account the working surface height and the person's operating posture for weighted correction. Finally, the panel with the smallest weighted distance value is determined as the current person's working panel, and its identification information is temporarily associated with the person's identity.

[0094] In summary, by calculating spatial distance, precise matching between personnel and specific curtain wall operation objects is achieved, ensuring that subsequent safety regulations can accurately correspond to the actual operation panels, avoiding mismatch of regulations caused by overly coarse area division, and providing precise spatial anchor points for differentiated monitoring.

[0095] S31. Obtain the identification information of the work module, and retrieve the corresponding safety specification template based on the identification information of the work module.

[0096] Specifically, based on the identified work segment identifier, the exclusive safety specification template for that segment is quickly retrieved from the cloud database or local cache. Each curtain wall test segment is assigned a unique identifier during installation and is bound and stored with metadata such as design parameters, construction requirements, and safety standards.

[0097] When retrieving a template, the template is precisely matched in the preset template library based on the section identifier to obtain the specification content containing the specific operation requirements of that section, such as the location of the safety belt attachment point for high-altitude operations, the scope of flammable material clearance around hot work operations, and prohibited areas for hoisting operations.

[0098] The above achieves dynamic adaptation of security standards, enabling different sectors within the same area to implement differentiated monitoring strategies based on their own characteristics. This enhances the pertinence and effectiveness of standard implementation and ensures that the monitoring content is highly consistent with the actual risk situation of each sector.

[0099] S32. Convert the safety specification template into a specification feature vector.

[0100] Specifically, the text descriptions or structured rules in the safety specification templates are parsed to extract key constraints such as permitted action types, required safety equipment, tool usage requirements, and distance thresholds. These constraints are then encoded as numerical or discrete features. For example, "wearing a face mask" is encoded as a mandatory requirement in the state feature dimension, and "distance from a fire extinguisher less than 3 meters" is encoded as a threshold constraint in the position feature dimension. Subsequently, a feature mapping network transforms these rules into fixed-dimensional specification feature vectors. These specification feature vectors share the same dimension and semantic space as the subsequently extracted operational feature vectors in real time.

[0101] The above approach digitizes and vectorizes abstract safety rules, providing a calculable mathematical benchmark for subsequent real-time comparisons, reducing ambiguity in rule interpretation, and enabling the system to assess the compliance of operational behaviors with a unified metric.

[0102] S33. Encrypt the standard feature vector to generate a standard key.

[0103] Specifically, the same encryption algorithm as the qualification key and work key is used, combined with contextual information such as the current timestamp, work segment identifier, and region identifier, to encrypt the standardized feature vector, forming encrypted data with uniqueness and timeliness. This standardized key, together with the personnel qualification key and work key, constitutes a three-level key system, which is encrypted and transmitted between the edge node and the cloud and bound to the edge node hardware to ensure that the key is not stolen or tampered with during use. After the standardized key is generated, it is issued to the edge node of the corresponding region as the monitoring benchmark for the personnel's current work and is automatically destroyed after the work is completed or the personnel leave.

[0104] The above measures ensure the security of the standard template during transmission and storage, preventing security rules from being maliciously tampered with or leaked. At the same time, the timeliness of the key and the hardware binding mechanism ensure the real-time performance and reliability of the monitoring strategy, providing encrypted protection for the secure operation of the entire system.

[0105] In addition, refer to Figure 5 Furthermore, in one embodiment, step S5 is refined into the following sub-steps: S50. Decrypt the standard key to obtain the standard feature vector.

[0106] Specifically, after receiving the encryption specification key from the cloud, the edge node first verifies the integrity and timeliness of the key. Then, it uses the decryption algorithm corresponding to the encryption algorithm of the key, combined with a decryption key derived from the edge node's unique hardware identifier, to decrypt the ciphertext and restore the original specification feature vector. This specification feature vector contains the standard operation feature thresholds and feature combinations set for the specific operation type in the current work module, such as permitted action categories, required safety equipment, tool usage specifications, and digital constraints such as safe distances from hazardous sources.

[0107] S51. Decrypt the job key to obtain the job feature vector.

[0108] Specifically, edge nodes receive operation keys encrypted and generated by themselves or other nodes in real time, decrypt the operation keys using the same decryption mechanism, and extract feature vectors reflecting the current real-time operation status of personnel. This operation feature vector is formed by fusing multimodal information such as target detection, skeletal keypoint recognition, and action classification, and includes dimensions such as position, action, tool, status, and curtain wall panel status.

[0109] S52. Set the transmittance level based on the transmittance of the curtain wall specimen in the current work area, and dynamically adjust the confidence weight of each dimension in the work feature vector according to the transmittance level.

[0110] Specifically, the system first analyzes real-time images captured by cameras, calculating indicators such as average brightness, brightness variance, and edge sharpness to comprehensively evaluate the light transmittance level of the curtain wall specimen, typically categorized into three levels: low transmittance, medium transmittance, and high transmittance. According to preset mapping rules, when light transmittance is high, visual features such as motion recognition and skeletal keypoint extraction may be affected by background interference. Therefore, the system appropriately reduces the confidence weight of motion and posture features while increasing the weight of features less affected by lighting, such as tool recognition and safety equipment detection. Conversely, when light transmittance is low, the system adjusts accordingly.

[0111] The above enables the system to adapt to the transparency of the curtain wall glass, dynamically optimize the credibility allocation of the feature vector, effectively suppress the interference of changes in ambient light on the judgment of work behavior, and improve the stability and accuracy of the recognition results.

[0112] S53. Based on the adjusted weights, calculate the weighted deviation between the obtained job feature vector and the standard feature vector.

[0113] Specifically, the weighted real-time job feature vector is compared with the standardized feature vector dimension by dimension, and the overall deviation between the two is calculated using metrics such as weighted Euclidean distance or weighted cosine similarity. The difference in each dimension is multiplied by the confidence weight corresponding to that dimension and then summed to obtain the final deviation value.

[0114] In summary, by comprehensively considering the credibility of each dimension of characteristics in the current environment, the deviation can more accurately reflect the actual gap between the operation behavior and the standard, avoid misjudgment caused by interference from a single dimension, and provide a scientific basis for subsequent alarm decisions.

[0115] S54. If the weighted deviation exceeds the preset threshold, the comparison result is determined to be outside the preset deviation range.

[0116] Specifically, the calculated weighted deviation is compared with a pre-set safety threshold, which is set differently for different job types and risk levels. When the deviation exceeds the threshold, the system immediately determines that the current job behavior does not comply with safety regulations and triggers an alarm process.

[0117] The above enables real-time anomaly detection of work behavior, ensuring that violations are identified as soon as they occur, providing accurate trigger signals for on-site intervention and post-event traceability, and effectively improving the response speed and reliability of safety management.

[0118] In addition, refer to Figure 6 Furthermore, in one embodiment, step S6 is refined into the following sub-steps: S60. During the validity period of the standard key, calculate in real time the spatial distance between the personnel and the preset hazard source, safety equipment or monitoring personnel, wherein the qualification key of the monitoring personnel needs to be verified by whether they hold the work permission of the work area.

[0119] Specifically, based on multi-camera joint calibration and spatial positioning technology, a global three-dimensional coordinate system is established. The three-dimensional coordinates of key points of the human body are obtained in real time through target detection and skeletal key point recognition. At the same time, the boundary coordinates of dangerous sources such as edges, hot work points, and areas under suspended objects are retrieved from the preset database, as well as the precise locations of safety equipment such as fire extinguishers and safety belt attachment points, and the real-time location of the monitoring personnel.

[0120] For monitoring personnel, the system first verifies whether their qualification key has passed the work area permission verification. Only after confirming their monitoring qualifications are they included in the distance calculation objects. Then, the Euclidean distance between the personnel and each object is calculated, and the height of the work surface is taken into account for correction to obtain accurate spatial distance values.

[0121] The above achieves real-time distance quantification between personnel and key spatial objects, providing an accurate data foundation for subsequent security assessments. At the same time, qualification verification ensures the legality of the guardian's identity, avoiding security vulnerabilities caused by unqualified personnel acting as guardians.

[0122] S61. Compare the spatial distance with the corresponding safety threshold.

[0123] Specifically, the distance thresholds are set as follows: 1 meter for distance to an adjacent edge, 5 meters for distance to a hot work point, 3 meters for distance to a fire extinguisher, and 10 meters for distance to monitoring personnel. These thresholds can be dynamically adjusted based on the type of work, such as working at heights, hot work, or hoisting. The system compares the calculated spatial distances one by one with the corresponding thresholds in real time to determine if they exceed the safe limits. When multiple distances exist simultaneously, a parallel comparison mechanism is used to ensure that all dimensions are monitored in a timely manner.

[0124] In summary, by comparing multiple thresholds in parallel, it is possible to quickly and accurately identify whether personnel are in dangerous areas, whether safety equipment is accessible, and whether monitoring personnel have left their posts. This provides clear and quantifiable triggering conditions for proactive early warning, avoiding the ambiguity of subjective judgment.

[0125] S62. If the spatial distance exceeds the safety threshold, a distance warning will be triggered, and the calculation frequency of spatial distance will be dynamically adjusted according to the risk level of the area where the personnel are located.

[0126] Specifically, once any spatial distance is detected to exceed the corresponding safety threshold, a graded early warning message is immediately generated, and real-time reminders are provided through on-site voice broadcasts and sound and light alarms, while the warning details are pushed to the administrator terminal.

[0127] Based on this, the system automatically adjusts the distance calculation frequency according to the risk level of the area where the personnel are currently located. For example, when personnel are in high-risk areas such as high-altitude edges, hot work areas, or hoisting areas, the calculation frequency is increased from once every 5 seconds to once per second; when personnel are in ordinary work areas or far from the hazard source, the frequency is restored to a lower frequency.

[0128] The above not only enables real-time early warning of distance anomalies, but also optimizes edge node computing resources while ensuring monitoring density in high-risk scenarios through a risk-adaptive frequency adjustment mechanism, effectively balancing system response speed and computing power consumption, and improving the timeliness of early warnings and system operating efficiency.

[0129] In addition, refer to Figure 7 Furthermore, in one embodiment, after step S6, steps S60, S61, and S62 are added: S63. Real-time acquisition of the current person's facial features and gait features.

[0130] Specifically, high-definition cameras continuously capture images of personnel covering the work area. A face recognition model running on front-end edge computing nodes extracts real-time facial feature vectors from the video frames. This model employs a lightweight convolutional neural network structure, ensuring both accuracy and real-time performance. Simultaneously, a gait recognition model extracts real-time gait feature vectors from multiple consecutive frames of full-body images. This model utilizes a temporal convolutional network to analyze the trajectory of human joint points, extracting dynamic features such as gait cycle, stride length, and joint angle changes. These two feature extraction processes are performed in parallel and independently, ensuring that the latest biometric data is available at all times.

[0131] The above provides real-time, multi-dimensional feature input for continuous identity verification, enabling the system to track personnel identity throughout the process, avoiding the risk of being replaced by relying solely on one-time verification at the entry point, and laying a reliable data foundation for subsequent comparisons.

[0132] S64. Compare the real-time facial features with the facial features in the qualification key, and calculate the first similarity.

[0133] Specifically, the edge node retrieves the qualification key corresponding to the individual, decrypts it to restore the facial feature vector stored during initial registration, and then uses a cosine similarity algorithm to calculate the similarity score between the real-time facial features and the initial features. This score reflects the degree of matching between the current face and the original face. During the comparison process, the system automatically normalizes the feature vector based on lighting conditions and camera angle to reduce interference from environmental factors.

[0134] The above quantitative method quickly determines whether the current person's face matches the initial identity, providing accurate numerical basis for identifying abnormal situations such as occlusion and person switching, and ensuring the objectivity and accuracy of identity verification.

[0135] S65. If the first similarity is lower than the preset first threshold, it is determined to be a suspected occlusion and gait-assisted verification is triggered.

[0136] Specifically, the system pre-sets a similarity threshold, which is calibrated based on the typical similarity distribution under normal conditions of wearing a helmet, mask, etc. in real-world scenarios. When the similarity between the real-time facial features and the initial facial features is lower than this threshold, it indicates that the face may be obscured by a helmet, mask, or other objects, or that the shooting angle is too off-center, resulting in incomplete feature extraction. In this case, the system does not immediately determine that the identity is lost, but instead activates the gait-assisted verification process, shifting the verification focus to gait features.

[0137] S66. Compare the real-time gait features with the gait features in the qualification key, and calculate the second similarity.

[0138] Specifically, the system decrypts the gait feature vector stored during initial registration from the qualification key. This vector contains dynamic information such as the individual's unique gait cycle and joint angle change curve. Subsequently, a dynamic time warping algorithm is used to compare the real-time gait feature sequence with the initial feature sequence and calculate the similarity score between the two. This algorithm can effectively deal with the time axis scaling problem caused by changes in walking speed, making the comparison results more accurate.

[0139] In summary, by utilizing the characteristic that gait features are not affected by facial occlusion, gait features can be used as an effective supplement to facial verification, ensuring that identity verification can continue even when the face is unavailable, thus providing a reliable redundancy mechanism for continuous identity verification.

[0140] S67. If the second similarity is higher than the preset second threshold, the identity verification is maintained and the face occlusion event is recorded. If the second similarity is also lower than the preset threshold, the identity is determined to be lost and an alarm is triggered.

[0141] The system presets a gait similarity threshold. When the second similarity is higher than this threshold, it is determined that the current person matches the initial identity, and the face verification failed only due to facial occlusion. In this case, the system maintains the original identity verification status to ensure uninterrupted operation monitoring. At the same time, the face occlusion event is recorded in the log for subsequent analysis of occlusion patterns or to remind personnel to adjust equipment. If the second similarity is also lower than the threshold, it indicates that the current person is neither a face match nor a gait match, and the identity is suspected to have been replaced or lost. The system immediately triggers an alarm and pushes it to the administrator terminal.

[0142] The above achieves collaborative verification of face and gait modes, improves the reliability and anti-interference ability of continuous identity verification, ensures accurate traceability of personnel identity in complex working environments, and provides a solid guarantee for the continuity and traceability of safety management.

[0143] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0144] This application also provides a hazard identification system for a curtain wall laboratory, which corresponds one-to-one with the hazard identification method for a curtain wall laboratory in the embodiments.

[0145] refer to Figure 8 A hazard identification system for a curtain wall testing laboratory includes: 1. Qualification acquisition module; 2. Qualification verification module; 3. Standard acquisition module; 4. Operation monitoring module; 5. Safety judgment module; and 6. Safety alarm module. Detailed descriptions of each functional module are as follows: Qualification Acquisition Module 1: Used to extract human features from images of personnel entering the work area and generate qualification keys based on these features.

[0146] Qualification Verification Module 2: This module is used to match qualification keys with the personnel qualification database to verify whether a person has the necessary work permissions for the work area.

[0147] Specification Acquisition Module 3: If the user holds the work permissions for the work area during the work process, it generates a specification key based on the security specification template of the work area.

[0148] Job monitoring module 4: Used to generate job keys based on the real-time job characteristics of personnel. Job characteristics include location characteristics, action characteristics, tool characteristics and status characteristics.

[0149] Security Judgment Module 5: Used to compare the standard key with the job key during the validity period of the standard key.

[0150] Security alarm module 6: If the comparison result exceeds the preset deviation range, an alarm message will be generated, and the alarm message will be associated with the identity of the person corresponding to the facial features.

[0151] The system comprises several modules: Module 1 (Qualification Acquisition) extracts facial features from personnel images and generates a qualification key for easy identification; Module 2 (Qualification Verification) matches the qualification key against a personnel qualification database to verify work permissions; Module 3 (Standardization Acquisition) generates a standardization key based on a work area safety standard template to ensure compliance; Module 4 (Work Monitoring) generates a work key based on real-time personnel work characteristics for monitoring; Module 5 (Safety Judgment) compares the standardization key with the work key during the standardization key's validity period to determine compliance; and Module 6 (Safety Alarm) generates an alarm message associated with the personnel's identity when the comparison result exceeds a preset deviation range, providing timely feedback on violations. Through the combined action of these modules, precise monitoring and early warning of personnel work in the curtain wall laboratory can be achieved.

[0152] Specific limitations regarding the hazard identification system for curtain wall testing laboratories can be found in the context of the limitations on hazard identification methods for curtain wall testing laboratories, and will not be repeated here. Each module in the aforementioned hazard identification system for curtain wall testing laboratories can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in an electronic device, or stored in the memory of the electronic device as software, so that the processor can call and execute the operations corresponding to each module. In one embodiment, an electronic device is provided, which is a user terminal. (Reference) Figure 9The electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores test data tables. The network interface allows communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for identifying hazard sources in a curtain wall testing laboratory.

[0153] In one embodiment, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: S1. Extract human features from images of personnel entering the work area, and generate qualification keys based on human features.

[0154] S2. Match the qualification key with the personnel qualification database to verify whether the personnel have the work permission for the work area.

[0155] S3. If you have the work permission for the work area during the work process, generate a standard key based on the safety standard template of the work area.

[0156] S4. Generate a job key based on the real-time job characteristics of personnel. The job characteristics include location characteristics, action characteristics, tool characteristics, and status characteristics.

[0157] S5. During the validity period of the standard key, compare the standard key with the job key.

[0158] S6. If the comparison result exceeds the preset deviation range, an alarm message is generated, and the alarm message is associated with the identity of the person corresponding to the facial features.

[0159] In one embodiment, the sub-steps of step S1 refinement include: S10. Based on the results of reflection interference detection of personnel images, identify and determine whether the personnel images form virtual image regions.

[0160] S11. If it is determined that the image of a person contains a virtual image area, then the virtual image area is blocked, and an instruction is sent to switch to the camera with the associated adjacent viewpoint to re-acquire the image of the person until a real image of the person without virtual image interference is acquired.

[0161] S12. Using a deep learning face recognition model, the face portion of a real person image is mapped to a face feature vector.

[0162] S13. Using a gait recognition model, gait features are extracted from real person images and mapped to gait feature vectors. Gait features include gait cycle and skeleton dynamic features.

[0163] S14. Combine the facial feature vector and gait feature vector and encrypt them to generate an qualification key.

[0164] In one embodiment, the sub-steps of step S4 refinement include: S40. By integrating the results of target detection, skeletal key point recognition and action classification of personnel through a deep learning task feature model, the task features of the person are extracted and mapped into a task feature vector.

[0165] S41. Obtain the real-time status characteristics of the curtain wall test panel and map them into a curtain wall operation feature vector.

[0166] S42. Combine the feature vectors of the human operation and the feature vectors of the curtain wall operation and encrypt them to generate an operation key.

[0167] In one embodiment, the sub-steps of step S3 refinement include: S30. Real-time calculation of the spatial distance between personnel and each curtain wall test panel in the image, and setting the curtain wall test panel with the shortest spatial distance as the working panel.

[0168] S31. Obtain the identification information of the work module, and retrieve the corresponding safety specification template based on the identification information of the work module.

[0169] S32. Convert the safety specification template into a specification feature vector.

[0170] S33. Encrypt the standard feature vector to generate a standard key.

[0171] In one embodiment, the sub-steps of step S5 refinement include: S50. Decrypt the standard key to obtain the standard feature vector.

[0172] S51. Decrypt the job key to obtain the job feature vector.

[0173] S52. Set the transmittance level based on the transmittance of the curtain wall specimen in the current work area, and dynamically adjust the confidence weight of each dimension in the work feature vector according to the transmittance level.

[0174] S53. Based on the adjusted weights, calculate the weighted deviation between the obtained job feature vector and the standard feature vector.

[0175] S54. If the weighted deviation exceeds the preset threshold, the comparison result is determined to be outside the preset deviation range.

[0176] In one embodiment, the sub-steps of step S6 are further refined as follows: S60. During the validity period of the standard key, calculate in real time the spatial distance between the personnel and the preset hazard source, safety equipment or monitoring personnel, wherein the qualification key of the monitoring personnel needs to be verified by whether they hold the work permission of the work area.

[0177] S61. Compare the spatial distance with the corresponding safety threshold.

[0178] S62. If the spatial distance exceeds the safety threshold, a distance warning will be triggered, and the calculation frequency of spatial distance will be dynamically adjusted according to the risk level of the area where the personnel are located.

[0179] In one embodiment, the additional steps following step S6 include: S63. Real-time acquisition of the current person's facial features and gait features.

[0180] S64. Compare the real-time facial features with the facial features in the qualification key, and calculate the first similarity.

[0181] S65. If the first similarity is lower than the preset first threshold, it is determined to be a suspected occlusion and gait-assisted verification is triggered.

[0182] S66. Compare the real-time gait features with the gait features in the qualification key, and calculate the second similarity.

[0183] S67. If the second similarity is higher than the preset second threshold, the identity verification is maintained and the face occlusion event is recorded. If the second similarity is also lower than the preset threshold, the identity is determined to be lost and an alarm is triggered.

[0184] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0185] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

Claims

1. A method for identifying hazard sources in a curtain wall testing laboratory, characterized in that, include: Based on the images of personnel entering the work area, human features are extracted, and qualification keys are generated based on the human features. The qualification key is matched with the personnel qualification database to verify whether the personnel have the work permission for the work area; If the user holds the operation permission for the operation area during the operation process, a standard key is generated based on the security specification template of the operation area. Based on the real-time operational characteristics of the personnel, an operational key is generated, wherein the operational characteristics include location characteristics, action characteristics, tool characteristics, and status characteristics; During the validity period of the specification key, the specification key is compared with the job key; If the comparison result exceeds the preset deviation range, an alarm message is generated, and the alarm message is associated with the identity of the person corresponding to the facial feature.

2. The method according to claim 1, characterized in that, The step of extracting human features from images of personnel entering the work area and generating qualification keys based on these human features includes: Based on the results of reflection interference detection of the personnel image, it is identified and determined whether the personnel image has a virtual image region; If it is determined that the personnel image contains a virtual image area, then the virtual image area is blocked, and an instruction is sent to switch to the associated adjacent view camera to re-acquire the personnel image until a real personnel image without virtual image interference is acquired. By using a deep learning face recognition model, the face portion of the real person image is mapped to a face feature vector; The gait recognition model extracts gait features from the real person image and maps them into gait feature vectors. The gait features include gait cycle and skeleton dynamic features. The facial feature vector and gait feature vector are combined and encrypted to generate the qualification key.

3. The method according to claim 2, characterized in that, The step of generating a job key based on the real-time job characteristics of the personnel includes: By fusing the results of target detection, skeletal key point recognition and action classification of the personnel through a deep learning task feature model, the task features of the personnel are extracted and mapped into a task feature vector of the personnel. The real-time status characteristics of the curtain wall test panel are obtained and mapped to the curtain wall operation feature vector. The feature vectors of the human operation and the feature vectors of the curtain wall operation are combined and encrypted to generate the operation key.

4. The method according to claim 3, characterized in that, The step of generating a specification key based on the security specification template of the work area if the holder has work permissions for the work area during the work process includes: The spatial distance between personnel and each curtain wall test panel in the image is calculated in real time, and the curtain wall test panel with the shortest spatial distance is set as the working panel; Obtain the identification information of the work module, and retrieve the safety specification template corresponding to the work module based on the identification information of the work module; Convert the security specification template into a specification feature vector; The standardized feature vector is encrypted to generate the standardized key.

5. The method according to claim 4, characterized in that, The step of comparing the standard key with the job key during the validity period of the standard key includes: The canonical key is decrypted to obtain the canonical feature vector; The job key is decrypted to obtain the job feature vector; A transmittance level is set based on the transmittance of the curtain wall specimen in the current work area, and the confidence weight of each dimension in the work feature vector is dynamically adjusted according to the transmittance level. Based on the adjusted weights, the weighted deviation between the job feature vector and the standard feature vector is calculated. If the weighted deviation exceeds a preset threshold, the comparison result is determined to be outside the preset deviation range.

6. The method according to claim 1, characterized in that, The step of generating an alarm message if the comparison result exceeds a preset deviation range, and associating the alarm message with the identity of the person corresponding to the facial feature, includes: During the validity period of the specified key, the spatial distance between the personnel and the preset hazard source, safety equipment or monitoring personnel is calculated in real time, wherein the qualification key of the monitoring personnel needs to be verified by whether they hold the work permission of the work area; The spatial distance is compared with the corresponding security threshold; If the spatial distance exceeds the safety threshold, a distance warning is triggered, and the calculation frequency of the spatial distance is dynamically adjusted according to the risk level of the area where the person is located.

7. The method according to claim 4, characterized in that, After the step of generating an alarm message if the comparison result exceeds a preset deviation range, and associating the alarm message with the identity of the person corresponding to the facial feature, the method further includes: Real-time acquisition of facial features and gait features of the current person; The real-time facial features are compared with the facial features in the qualification key to calculate and obtain the first similarity. If the first similarity is lower than a preset first threshold, it is determined to be a suspected occlusion, triggering gait-assisted verification; The real-time gait features are compared with the gait features in the qualification key to calculate and obtain the second similarity. If the second similarity is higher than the preset second threshold, the identity verification status is maintained and the face occlusion event is recorded. If the second similarity is also lower than the preset threshold, the identity is determined to be lost and an alarm is triggered.

8. A hazard identification system for a curtain wall testing laboratory, characterized in that, include: Qualification acquisition module (1): used to extract human features based on images of personnel entering the work area, and generate qualification keys based on the human features; Qualification verification module (2): used to match the qualification key with the personnel qualification database to verify whether the personnel have the work permission for the work area; Specification acquisition module (3): If it holds the operation permission of the operation area during the operation process, it generates a specification key based on the security specification template of the operation area; The job monitoring module (4) is used to generate a job key based on the real-time job characteristics of the personnel, wherein the job characteristics include location characteristics, action characteristics, tool characteristics and status characteristics; Security judgment module (5): used to compare the standard key with the job key during the validity period of the standard key; Security alarm module (6): If the comparison result exceeds the preset deviation range, an alarm message is generated, and the alarm message is associated with the identity of the person corresponding to the facial feature.

9. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as any one of the curtain wall laboratory hazard identification methods as claimed in claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program capable of being loaded by a processor and executing any of the curtain wall laboratory hazard identification methods as claimed in claims 1 to 7.

Images