Identity verification method based on biometric recognition and ladder guard robot

Abstract

The application provides an identity verification method based on biometric recognition and a landing port monitoring robot, and relates to the technical field of identity verification and monitoring. Through the identity verification method based on the landing port monitoring robot, the multi-modal biometric features of the personnel to be verified and the current landing port scene parameters are synchronously collected, a dynamic verification strategy adjustment of scene adaptation is realized by combining an adaptive identity verification model, the problem that a fixed verification mode cannot adapt to a complex landing port scene is solved, and the adaptability and accuracy of biometric recognition in different scenes such as different human flows and illuminations are improved; through real-time comparison of the biometric features after recognition processing and a preset identity information database, efficient and accurate verification of the personnel identity is realized, risk level evaluation and differentiated landing port monitoring and control are completed in combination with the identity matching result, the landing port passing efficiency is improved, the accurate prevention and control of high-risk personnel are strengthened, and the safety, accuracy and passing efficiency of the landing port scene identity verification are taken into account.

Description

Technical Field

[0001] This invention relates to the field of identity verification and monitoring technology, and in particular to an identity verification method based on biometric recognition and a ladder gate monitoring robot. Background Technology

[0002] With the increasing demand for security management in densely populated areas such as ports, industrial parks, and office buildings, the level of intelligence in identity verification and stairwell monitoring has become crucial for ensuring safe passage and improving efficiency. Currently, stairwell monitoring robots are gradually being applied to personnel identity verification and monitoring in various stairwell scenarios. Their core technology relies on biometric recognition to confirm personnel identity; common biometric features include facial, iris, and fingerprint multimodal characteristics.

[0003] Most existing authentication methods for ladder gate monitoring robots employ fixed biometric acquisition parameters and verification strategies, meaning that the acquisition rate, accuracy, angle, and other parameters of the multimodal biometric acquisition device remain unchanged. This fixed approach has significant technical drawbacks, making it difficult to balance the accuracy of authentication with access efficiency.

[0004] For example, when the entrance is in a high-traffic area, a fixed data collection rate and verification process can lead to queuing and congestion of people waiting to be verified, resulting in low passage efficiency and failing to meet the fast passage requirements of high-traffic scenarios. When the entrance is in a low-traffic area, fixed low-precision data collection parameters may result in incomplete biometric data collection and increased recognition errors, affecting the accuracy of identity verification and failing to effectively prevent potential security risks. Summary of the Invention

[0005] This invention provides an identity verification method based on biometric recognition and a ladder gate monitoring robot, which can balance the accuracy of identity verification and the efficiency of passage.

[0006] In a first aspect, the present invention provides an identity verification method based on biometric recognition and a stairwell monitoring robot. The stairwell monitoring robot includes a multimodal biometric acquisition device. The method includes: acquiring the multimodal biometrics of the person to be verified and the scene parameters of the current stairwell. The multimodal biometrics include facial features, iris features, and fingerprint features; analyzing the application scenario and dynamically adjusting the verification strategy based on the multimodal biometrics, scene parameters, and an adaptive identity verification model to obtain the biometrics after recognition processing; performing real-time comparison based on the biometrics after recognition processing and a preset identity information database to obtain the identity matching result; and conducting risk level assessment and stairwell monitoring and control based on the identity matching result.

[0007] Secondly, embodiments of the present invention provide an identity verification device based on biometric recognition and a stairwell monitoring robot. This device includes a communication module and a processing module. The communication module is used to collect multimodal biometric features of the person to be verified and scene parameters of the current stairwell. The multimodal biometric features include facial features, iris features, and fingerprint features. The processing module is used to analyze the application scenario and dynamically adjust the verification strategy based on the multimodal biometric features, scene parameters, and an adaptive identity verification model to obtain the processed biometric features. Based on the processed biometric features and a preset identity information database, a real-time comparison is performed to obtain an identity matching result. Based on the identity matching result, a risk level assessment and stairwell monitoring control are performed on the person to be verified.

[0008] Thirdly, embodiments of the present invention provide an electronic device including a memory and a processor. The memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to perform the steps of the first aspect and any possible implementation of the first aspect described above.

[0009] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, characterized in that, when executed by a processor, the computer program implements the steps of the first aspect and any possible implementation thereof described above.

[0010] This invention provides an identity verification method based on biometric recognition and a stairwell monitoring robot. This method simultaneously collects multimodal biometric features of the person to be verified along with current stairwell scene parameters. Combined with an adaptive identity verification model, it dynamically adjusts the verification strategy to adapt to different scenarios, solving the problem that fixed verification modes cannot adapt to complex stairwell scenes. This improves the adaptability and accuracy of biometric recognition under different pedestrian flow and lighting conditions. Through real-time comparison of the processed biometric features with a pre-set identity information database, efficient and accurate verification of personnel identity is achieved. Combined with the identity matching results, risk level assessment and differentiated stairwell monitoring and control are completed. This improves stairwell passage efficiency while strengthening precise control of high-risk personnel, balancing the security, accuracy, and passage efficiency of stairwell scene identity verification. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating an identity verification method based on biometric recognition and a ladder gate monitoring robot, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of an identity verification device based on biometric recognition and a ladder gate monitoring robot provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0013] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0014] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner to facilitate understanding.

[0015] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or modules is not limited to the steps or modules listed, but may optionally include other steps or modules not listed, or may optionally include other steps or modules inherent to such process, method, product, or device.

[0016] To make the objectives, technical solutions, and advantages of the present invention clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0017] like Figure 1 As shown, this invention provides an authentication method based on biometric recognition and a ladder entrance monitoring robot. The ladder entrance monitoring robot includes a multimodal biometric acquisition device, and the method includes steps S101-S104.

[0018] S101. Collect the multimodal biometrics of the personnel to be verified and the scene parameters of the current ladder entrance.

[0019] In this embodiment of the application, multimodal biometrics include facial features, iris features, and fingerprint features.

[0020] For example, facial features include global and local facial features such as the outline of the face, key point coordinates, and texture features; iris features include the unique texture features of the ring area between the pupil and sclera of the human eye, which have the characteristics of high uniqueness, strong stability, non-contact acquisition, and difficulty in forgery; fingerprint features include detailed features such as ridges, valleys, bifurcation points, and endpoints of the fingertip epidermis.

[0021] In some embodiments, scene parameters are quantitative indicators of the environment and passage status that characterize the current passage environment at the entrance and directly affect the accuracy and efficiency of identity verification. Scene parameters include pedestrian density, light intensity, passage speed, and degree of occlusion.

[0022] For example, crowd density is used to quantify the degree of crowding in the passage area at the entrance of the building. The present invention presets three threshold levels: low crowd density (<0.5 people / ㎡), medium crowd density (0.5-1.5 people / ㎡), and high crowd density (>1.5 people / ㎡).

[0023] For example, the light intensity is used to quantify the lighting conditions of the stairwell environment. The present invention presets two threshold levels: weak light (<500 lux, including nighttime, backlight, and indoor weak light environment) and strong light (≥500 lux, including daytime strong light and backlight environment).

[0024] For example, the passage rate is used to quantitatively characterize the personnel passage pressure at the stairwell and help determine the passage demand of the scenario. The normal passage rate threshold is 30-60 people / minute, and a passage rate exceeding 60 people / minute is considered a high passage pressure scenario.

[0025] For example, the degree of occlusion is used to quantify the impact of occlusion on biometric data collection. This invention presets two threshold levels: low occlusion (<0.2, effective area occlusion ratio is less than 20%) and high occlusion (≥0.2, effective area occlusion ratio is greater than 20%).

[0026] In some embodiments, the elevator entrance monitoring robot is deployed in scenarios such as border crossings, park gates, office building elevator entrances, and transportation hub ticket gates. It is an intelligent robot with the capabilities of fixed-point duty / autonomous movement, biometric data collection, scene perception, identity verification, risk warning, and on-site control. The core of the robot is equipped with multimodal biometric data collection equipment, scene perception sensors, edge computing modules, sound and light warning modules, and wireless communication modules, and can be linked with elevator entrance gates and back-end control terminals.

[0027] In some embodiments, the multimodal biometric acquisition device is integrated into the hardware module set of the ladder gate monitoring robot body, which can simultaneously acquire multiple types of unique human physiological characteristics, specifically including: a facial recognition camera (high-definition binocular camera, supporting visible light / infrared dual modes, used to acquire facial features); an iris scanner (near-infrared high-definition imaging module, used to acquire iris texture features); and a fingerprint reader (capacitive high-sensitivity fingerprint sensor, used to acquire fingerprint ridge features). The device can adjust operating parameters such as focus, exposure, acquisition rate, and scanning accuracy in real time through the robot's main control module.

[0028] For example, when the millimeter-wave radar / infrared human body sensor of the ladder gate monitoring robot detects that a person to be verified has entered the preset collection area (the effective verification range of 0.3-1.5 meters away from the robot), it automatically triggers the multimodal biometric collection device and the scene perception module to start synchronously, avoiding invalid empty collection.

[0029] For example, embodiments of the present invention can use a binocular facial recognition camera to acquire visible light + infrared facial images of the person to be verified at a frame rate of 30fps, and extract global and local facial features; use an iris scanner to emit low-power near-infrared light to acquire iris images of the person to be verified's eyes, and extract iris texture details; and use a fingerprint reader to acquire fingerprint images of the person to be verified, and extract fingerprint ridges and details.

[0030] For example, embodiments of the present invention can simultaneously collect four types of parameters: People density: using a camera target detection algorithm, the number of people in the effective passage area of ​​the stairwell is counted, and the number of people per unit area is calculated by combining the area area (unit: people / ㎡); Illumination intensity: using a high-precision light sensor to collect the illuminance value of the stairwell environment (unit: lux); Passage rate: using a target tracking algorithm, the number of people passing through the stairwell within a 1-minute sliding time window is counted (unit: people / minute); Occlusion degree: using an image segmentation algorithm, the proportion of occluded pixels in the effective collection area of ​​the face / iris of the person to be verified, and the proportion of obstacle occlusion in the passage area of ​​the stairwell are calculated to obtain a quantitative occlusion degree value in the range of 0-1 (the higher the value, the more severe the occlusion).

[0031] S102. Based on multimodal biometrics, scene parameters, and an adaptive identity verification model, analyze the application scenario and dynamically adjust the verification strategy to obtain the biometrics after recognition processing.

[0032] In some embodiments, the adaptive authentication model is an end-to-end model built on a deep neural network that can dynamically adjust the authentication logic according to the ladder scene. The core includes three major units: scene recognition module, verification strategy template library, and feature fusion module, which can realize accurate determination of scene type, dynamic matching of verification strategy, and adaptive fusion of multimodal features.

[0033] In some embodiments, the verification strategy is the parameter configuration and execution logic for the entire multimodal biometric authentication process, including: the recognition weight of each biometric modality, the accuracy threshold of feature acquisition / recognition, the feature fusion algorithm, the comparison priority, and the verification process simplification / enhancement rules, etc.

[0034] As one possible implementation, step S102 can be specifically implemented as steps S1021-S1024.

[0035] S1021. Normalize the scene parameters to obtain the scene feature vector.

[0036] For example, embodiments of the present invention can use the 3σ criterion to remove abnormal data that exceed the reasonable range of values ​​for scene parameters, and use the average value of parameters at adjacent time points to fill in the missing data.

[0037] For example, in the embodiments of the present invention, the theoretical minimum and maximum values ​​of each type of scene parameter after cleaning can be predetermined, and each type of parameter can be mapped to the 0-1 interval by the formula Xnorm=(X-Xmin) / (Xmax-Xmin) to obtain the normalized single-dimensional scene parameters.

[0038] For example, embodiments of the present invention can allocate differentiated weights based on the degree of influence of various scenario parameters on identity verification: pedestrian density weight 0.35, occlusion degree weight 0.3, light intensity weight 0.2, passage speed weight 0.15, and the total weight is 1.

[0039] S1022. Denoising, alignment, and adaptive preprocessing are performed on the multimodal biometrics to obtain adaptive enhanced features.

[0040] For example, for facial feature images: bilateral filtering is used to remove illumination noise and motion blur, while preserving the edge features of facial features; for iris feature images: Gaussian filtering is used to remove spot noise from near-infrared imaging, and morphological operations are used to separate the effective iris region; for fingerprint feature images: median filtering is used to remove fingerprint stains and scratch noise, and to enhance the contrast between ridges and valleys.

[0041] For example, key point localization and standardization alignment are performed on the three types of features after denoising: Facial features: 68 core key points of the face are located, and the face is aligned to a standard frontal pose through affine transformation to unify the face image size; Iris features: the inner and outer circular boundaries of the iris are located, and the annular iris region is expanded into a standardized rectangular texture image through polar coordinate transformation to eliminate the deviation caused by eye rotation; Fingerprint features: the core points and triangular points of the fingerprint are located, and the fingerprint is aligned to a standard direction and size through rotation and scaling transformation to unify the fingerprint image specifications.

[0042] For example, embodiments of the present invention can perform quality scoring on the aligned feature images and dynamically adjust the intensity of the enhancement algorithm based on the current scene parameters: for low-brightness facial / iris images in low-light scenes, the gamma value is adaptively adjusted to improve image brightness and contrast; for facial / iris images in highly occluded scenes, a lightweight generative adversarial network (GAN) is used to repair features in the occluded areas and restore effective feature information; for blurred fingerprint images, the sharpening intensity is adaptively adjusted to enhance the clarity of ridges; simultaneously, at each step of preprocessing, the features are synchronously quality-checked to generate the recognition confidence of the adaptively enhanced features.

[0043] S1023. Input the scene feature vector and adaptive enhancement features into the adaptive identity verification model to determine the scene type and match the verification strategy to obtain the target verification strategy.

[0044] For example, step S1023 can be specifically implemented as steps A1-A4.

[0045] A1. Associate the scene feature vector with the adaptive enhancement feature, remove redundant scene information that is irrelevant to biometric recognition, and generate an associated feature set.

[0046] For example, embodiments of the present invention can establish a scene-feature association mapping matrix: using 4-dimensional scene feature vectors (human density, light intensity, traffic speed, occlusion level) as rows and the recognition confidence of 3-modal adaptive enhancement features (face, iris, fingerprint) as columns, the influence weight of each type of scene parameter on the quality of each modality feature is calculated using the mutual information method to construct the scene-feature association mapping matrix, clarifying the strong correlation between scene parameters and feature quality (e.g., light intensity is strongly correlated with facial / iris feature quality, and occlusion level is strongly correlated with facial feature quality). Then, based on the association mapping matrix, a recursive feature elimination method is used to remove redundant information in the scene feature vectors with a correlation of less than 0.1 with the quality of all modal features, and simultaneously remove invalid feature components in the adaptive enhancement features that have no linkage with the scene parameters. The filtered scene features and adaptive enhancement features are then dimensionally concatenated to generate a dimensionally unified and strongly correlated association feature set.

[0047] A2. Based on the scene recognition module in the adaptive identity verification model, classify and identify the associated feature set to determine the scene type of the current stairwell.

[0048] In some embodiments, scene types include high or low pedestrian traffic, strong or weak lighting, and high or low occlusion.

[0049] For example, in this embodiment of the invention, the associated feature set can be input into a pre-trained scene recognition module. First, deep scene-feature association features are extracted from the associated feature set through convolutional layers. Then, dimensionality reduction and compression are performed through pooling layers to remove redundant feature components. Through fully connected layers and multi-label softmax classification layers, pedestrian flow, lighting, and occlusion are independently classified, outputting the classification result and corresponding classification confidence score for each dimension. This achieves simultaneous determination of multi-dimensional scene labels (e.g., simultaneously outputting a combined scene type of "high pedestrian flow + low lighting + high occlusion"), fully adapting to the complex real-world scene of an elevator entrance. The classification results are then validated: if the classification confidence score for any dimension is ≥0.85, the classification result for that dimension is considered valid; if the confidence score is <0.85, a second collection of scene parameters and reconstruction of the associated feature set are triggered, and reclassification is performed to ensure an accuracy rate of ≥99% for scene type determination.

[0050] A3. Based on the scenario type and the verification strategy template library in the adaptive authentication model, multiple verification strategies corresponding to the scenario type are matched.

[0051] In some embodiments, the verification strategy template library is a pre-built structured database stored within the adaptive authentication model. The library is indexed by "scene type label combination", with each index corresponding to a strategy cluster. Each strategy cluster contains 2-3 sets of basic verification strategies adapted to the scene. Each strategy presets the recognition weight, verification accuracy threshold, comparison priority, and fusion algorithm rules for each modality feature.

[0052] For example, embodiments of the present invention can generate a scene tag index based on scene type, search in a verification strategy template library, and match the strategy cluster corresponding to the scene tag index. A cosine similarity algorithm is used to calculate the matching degree between each verification strategy in the strategy cluster and the scene type, and the top 3 verification strategies with the highest matching degree are selected as candidate verification strategy sets. The initial recognition weights, verification accuracy thresholds, and fusion algorithm types of the facial / iris / fingerprint features of each strategy in the candidate verification strategy set are extracted.

[0053] A4. Based on multiple verification strategies, dynamically adjust the recognition weights and verification accuracy parameters of each modality feature in the verification strategy to select the target verification strategy.

[0054] For example, step A4 can be specifically implemented as steps A41-A44.

[0055] A41. Extract the recognition weights and verification accuracy parameters of facial features, iris features, and fingerprint features from each verification strategy.

[0056] In some embodiments, the identification weight is the proportion of contribution of each modality biometric feature in the entire identity verification process. The higher the weight, the greater the impact of that modality feature on the final identity determination. The verification accuracy parameter is a quantitative parameter that controls the strictness of the identification and comparison of each modality feature, including feature similarity threshold, feature point matching number threshold, identification calculation accuracy, and acquisition frame rate threshold.

[0057] A42. Combining the recognition confidence of adaptive enhancement features, the recognition weights and verification accuracy parameters of facial features, iris features, and fingerprint features in each verification strategy are normalized and calibrated to obtain normalized verification parameters.

[0058] In some embodiments, the recognition confidence of the adaptive enhancement feature is obtained by quality verification after adaptive preprocessing during the adaptive enhancement feature generation process.

[0059] For example, embodiments of the present invention can combine the single-modal recognition confidence scores of face, iris, and fingerprint, as well as the comprehensive recognition confidence score, to clarify the reliability level of each modality feature. For the parameter matrix of each candidate strategy, using the single-modal recognition confidence score as the core, a parameter calibration coefficient for each modality is calculated: for modalities with recognition confidence scores higher than the mean, the calibration coefficient is >1, used to increase weights and tighten the accuracy threshold; for modalities with recognition confidence scores lower than the mean, the calibration coefficient is <1, used to reduce weights and moderately relax the accuracy threshold, avoiding interference from low-quality features in identity determination. The recognition weights and accuracy parameters of each strategy are corrected row by row using the calibration coefficients. For example, if the iris feature calibration coefficient is 1.2, then the initial weight of the iris feature in this strategy is multiplied by 1.2, and the similarity threshold is multiplied by 1.05 (tightening the threshold); if the face feature calibration coefficient is 0.7, then the initial weight is multiplied by 0.7, and the similarity threshold is multiplied by 0.95 (moderately relaxing the threshold). All corrected parameters are normalized and mapped to the 0-1 interval. Simultaneously, the weight class parameters are also normalized. The final result is standardized, normalized validation parameters.

[0060] A43. Based on the normalized verification parameters and the core requirements of the current scenario type of the ladder gate, prioritize them and calculate the adaptation score of each verification strategy.

[0061] In some embodiments, core requirements include prioritizing efficiency in high-traffic scenarios, prioritizing iris recognition in low-light scenarios, and prioritizing fingerprint recognition in high-obstruction scenarios.

[0062] For example, embodiments of the present invention can determine the corresponding core requirements and priority weight matrix based on scenario type. This invention predefines the weight rules for four core scenarios: High-traffic scenarios: the core requirement is efficiency first, with priority weights of "traffic efficiency weight 0.6, recognition accuracy weight 0.4", and modal priority is face > iris > fingerprint. Low-light scenarios: the core requirement is iris recognition first, with priority weights of "recognition accuracy weight 0.7, traffic efficiency weight 0.3", and modal priority is iris > fingerprint > face. High-occlusion scenarios: the core requirement is fingerprint recognition first, with priority weights of "recognition accuracy weight 0.7, traffic efficiency weight 0.3", and modal priority is fingerprint > iris > face. Low-traffic, low-occlusion scenarios: the core requirement is accuracy first, with priority weights of "recognition accuracy weight 0.7, traffic efficiency weight 0.3", and modal priority is iris = face = fingerprint.

[0063] Subsequently, in this embodiment of the invention, the normalized verification parameters of each strategy can be substituted into the corresponding priority weight matrix, and the comprehensive fit score can be calculated using a weighted summation formula. The core calculation formula is: Fit Score = (Σ(Single Modal Parameter Calibration Value × Modal Priority Weight) × Accuracy Weight + Efficiency Parameter × Efficiency Weight) × 100. Based on the calculated fit scores, all candidate verification strategies are sorted from high to low.

[0064] A44. Select the verification strategy with the highest fit score as the target verification strategy.

[0065] S1024. Based on the target verification strategy, perform cross-modal fusion operations on multimodal biometrics to obtain the biometrics after recognition processing.

[0066] For example, step S1024 can be specifically implemented as steps B1-B3.

[0067] B1. Based on the target verification strategy, determine the fusion weights and fusion algorithms for three modalities: face, iris, and fingerprint.

[0068] For example, embodiments of the present invention can extract the recognition weights, core scene requirements, and accuracy / efficiency priorities of the three modalities (face, iris, and fingerprint) preset in the target verification strategy based on the target verification strategy. The fusion weights correspond 1:1 with the recognition weights in the target verification strategy. For example, in a low-light scene, if the iris recognition weight is 0.5, the fingerprint weight is 0.3, and the face weight is 0.2, then the fusion weights are 0.5, 0.3, and 0.2, allowing the high-priority, high-reliability modalities to occupy a core position in the fusion process. Based on the scene's accuracy / efficiency priorities, the appropriate fusion algorithm is dynamically selected to perfectly match the scene's requirements: High-traffic-flow efficiency-first scenarios: a weighted summation fusion algorithm is selected, which has minimal computational load and the fastest speed, achieving millisecond-level fusion and ensuring passage efficiency; Medium-traffic-flow-balanced scenarios: a multimodal feature splicing fusion algorithm is selected, balancing fusion accuracy and computational speed; Low-light / high-occlusion / low-traffic-flow accuracy-first scenarios: a deep fusion algorithm based on an attention mechanism is selected, which learns the complementary relationships between modalities through neural networks, achieving the highest fusion accuracy and maximizing the compensation for the feature defects of a single modality.

[0069] B2. An attention mechanism is used to effectively locate and extract key features of each modality's biometrics, thereby obtaining the key features of each modality.

[0070] In some embodiments, the attention mechanism can automatically enhance effective features that contribute highly to identity recognition and suppress interfering features. It is divided into two categories: channel attention and spatial attention. This case uses a lightweight SE-Net channel attention + spatial attention combination module, adapted to the edge computing power of the robot. Effective region: Among the various biometric modalities, the core region containing unique identity information and contributing the most to recognition. Key features: After enhancement by the attention mechanism and extraction of the effective region, highly recognizable and reliable features are obtained.

[0071] For example, embodiments of the present invention can adaptively enhance features for each modality by inputting a lightweight channel attention + spatial attention combination module. The channel attention module automatically assigns weights to the channel dimensions of each modality feature: high weights are assigned to feature channels containing unique identity information for enhancement; low weights are assigned to channels containing noise, background, occlusion, and other interference information for suppression, thus improving the signal-to-noise ratio of the features from the channel dimension. The spatial attention module accurately locates the effective recognition areas of each modality feature: facial features: locating the core areas of the eyes, nose, and mouth, excluding invalid areas such as cheeks, forehead, and hair that are easily affected by occlusion; iris features: locating the core effective area of ​​the iris ring texture, excluding invalid areas such as the pupil, sclera, eyelids, and light spots; fingerprint features: locating the core area with clear fingerprint ridges, excluding invalid areas with blurred edges, stains, or damage. Based on the located effective areas, high-discrimination features enhanced by channel attention are extracted.

[0072] B3. Based on the key features of each modality, as well as the fusion weights and fusion algorithms, feature mapping and fusion are performed to obtain the biometric features after recognition processing.

[0073] For example, in embodiments of the present invention, key features of each modality can be input into a feature mapping network, and feature transformation can be performed through a fully connected layer to obtain 512-dimensional feature vectors for face, iris, and fingerprint single modalities with consistent dimensions and uniform distribution, thereby achieving feature alignment across different modalities. Based on fusion weights and fusion algorithms, cross-modal fusion operations are performed on the three single-modal feature vectors, such as weighted summation fusion: fused feature vector = face feature vector × face fusion weight + iris feature vector × iris fusion weight + fingerprint feature vector × fingerprint fusion weight; or feature concatenation fusion: fused feature vector = weighted by the three single-modal feature vectors according to the fusion weights, concatenated in dimensions, and then compressed to 512 dimensions through a dimensionality reduction layer.

[0074] S103. Based on the biometric features after recognition processing and the preset identity information database, a real-time comparison is performed to obtain the identity matching result.

[0075] In some embodiments, the preset identity information database is a pre-built encrypted database that stores the identity information of legally authorized personnel. Each piece of personnel information corresponds to the storage of facial, iris, and fingerprint feature vectors encrypted with national cryptographic algorithms, as well as basic personnel identity information, historical access records, and risk level labels. The database adopts a deployment architecture of "local edge server + cloud backup", which supports distributed fast retrieval and encrypted comparison, balancing comparison efficiency and data security.

[0076] In some embodiments, the identity matching result is a standardized result generated by feature comparison, representing the identity verification status of the person to be verified, including: whether the identity matching is successful, the identity information of the matched person, the comprehensive matching score, the similarity score of each modality feature, the matching confidence score, and other core data.

[0077] As one possible implementation, step S103 can be specifically implemented as steps S1031-S1037.

[0078] S1031. Extract feature vectors from the biometric features after identification and processing, and encrypt them using an asymmetric encryption algorithm to obtain encrypted feature vectors.

[0079] For example, in the edge computing security zone of the ladder gate monitoring robot, the SM2 public key corresponding to the preset identity information database can be called to fully encrypt the feature vector extracted from the biometric features after identification processing. The encryption process does not generate any local storage or external transmission of plaintext features, thus eliminating the risk of plaintext leakage. At the same time, a digital digest of the feature vector is generated through the SM3 hash algorithm as an integrity verification code.

[0080] S1032. Based on the encrypted feature vector, a distributed hierarchical retrieval algorithm is used to select a set of candidate identity information with a similarity greater than the similarity threshold from the preset identity information database.

[0081] In some embodiments, the distributed hierarchical retrieval algorithm is a two-stage retrieval algorithm of "coarse screening-fine screening" designed for massive identity databases, solving the pain point of extremely low efficiency in traversing the entire database. The database adopts a distributed deployment of "edge nodes + core cloud". The first layer uses comprehensive feature vectors to perform millisecond-level coarse screening at the edge nodes, and the second layer performs secondary convergence at the core nodes, balancing retrieval efficiency and recall.

[0082] For example, embodiments of the present invention can dynamically adjust the retrieval mode and similarity threshold based on the current scenario type: High-traffic efficiency priority scenario: Enable edge node priority retrieval mode, only search the database of high-frequency passersby, and set the coarse similarity threshold to 0.6 to minimize retrieval time; Low-traffic / high-security accuracy priority scenario: Enable full database retrieval mode, and set the coarse similarity threshold to 0.7 to ensure that no legitimate personnel are missed.

[0083] For example, the first layer of distributed coarse screening involves sending the encrypted comprehensive feature vector to a distributed database. Using homomorphic encryption comparison technology, the cosine similarity between the feature to be verified and the features in the database is calculated without decrypting the ciphertext. This quickly filters out personnel information with similarity greater than a dynamic threshold, generating an initial candidate set. The second layer of fine screening convergence involves sending the initial candidate set to the core cloud node if the number of returned initial candidate sets exceeds 30. This is combined with single-modal feature vectors for a second coarse screening, eliminating invalid information with single-modal similarity below 0.5 to ensure the efficiency of subsequent accurate comparisons. Finally, the candidate set validity is verified by performing permission and validity checks on the converged candidate identity information set, automatically eliminating invalid personnel information and ultimately generating a valid candidate identity information set.

[0084] S1033. For each piece of information in the candidate identity information set, perform hierarchical and precise comparison of facial features, iris features, and fingerprint features in sequence, and calculate the similarity score of each modality feature.

[0085] For example, embodiments of the present invention can prioritize modalities in the target verification strategy and strictly determine the comparison order according to the core requirements of the scenario: in low-light scenarios, the priority is iris > fingerprint > face; in high-occlusion scenarios, it is fingerprint > iris > face; and in high-traffic scenarios, it is face > iris > fingerprint. Hierarchical modal comparison: according to the priority order, for each piece of information in the candidate identity information set, homomorphic comparison of single-modal encrypted features is performed sequentially, and the cosine similarity score of the corresponding modality is calculated: after completing the first priority modal comparison, if the modal similarity score of a candidate piece of information is <0.5, it is directly removed from the candidate set; for the remaining valid candidate information, the comparison of the second and third priority modalities is continued, and the standardized similarity score of each modality is calculated sequentially. Similarity score calibration: The similarity score is dynamically calibrated by combining the confidence levels of each modality feature: high-reliability modalities with confidence levels above the mean are identified, with a calibration coefficient > 1 to amplify their impact on the results; low-quality modalities with confidence levels below the mean are identified, with a calibration coefficient < 1 to reduce their interference with the matching results. Finally, similarity scores for the face, iris, and fingerprint modalities are generated for each candidate piece of information.

[0086] S1034. Based on the fusion weights of each modal feature in the target verification strategy, the similarity scores of each modal feature are weighted and fused to generate a comprehensive matching score for each piece of information in the candidate identity information set.

[0087] In some embodiments, the comprehensive matching score ranges from 0 to 1. The closer the score is to 1, the higher the degree of identity matching between the person to be verified and the candidate, and the stronger the reliability of the matching result.

[0088] For example, for each piece of information in the candidate identity information set, a standardized weighted summation formula is used to calculate the comprehensive matching score: Comprehensive Matching Score = Facial Similarity Score × Facial Fusion Weight + Iris Similarity Score × Iris Fusion Weight + Fingerprint Similarity Score × Fingerprint Fusion Weight. The calculated comprehensive matching scores are then validated to remove invalid data with abnormal scores or contradictory calculation logic, ensuring that each piece of candidate information corresponds to a unique and valid comprehensive matching score.

[0089] S1035. Compare the comprehensive matching score of each piece of information with the dynamically adjusted matching threshold to determine whether the identity matching is successful.

[0090] In some embodiments, the dynamically adjusted matching threshold is the lowest comprehensive matching score used to determine whether identity matching is successful, which is dynamically adjusted based on the security level and scenario type of the current stairwell scenario.

[0091] For example, embodiments of the present invention can dynamically adjust the identity matching threshold based on the current security level (daily / holiday / major security), scenario type, and target verification strategy of the current entrance scenario: for major security / high-security port entry and exit scenarios, the matching threshold is set to 0.85-0.9 to minimize the risk of misjudgment; for office building / park morning peak high-traffic scenarios, the matching threshold is set to 0.8 to balance security and passage efficiency; for low-traffic high-security scenarios, the matching threshold is set to 0.88 to ensure verification accuracy.

[0092] For example, in this embodiment of the invention, the comprehensive matching score can be compared with the dynamic matching threshold one by one: if the comprehensive matching score of any information in the candidate set is greater than or equal to the matching threshold, the identity matching is directly determined to be unsuccessful; if the comprehensive matching score of one or more information in the candidate set is greater than or equal to the matching threshold, the identity matching is determined to be successful, and the information with the highest comprehensive matching score is selected as the final matched identity information of the person to be verified. The uniqueness of the highest score information that has been matched is verified. Based on the final judgment result, an encrypted standardized identity matching result data packet is generated, including: basic judgment conclusion: identity matching successful / identity matching unsuccessful; when matching is successful: encrypted identity information of the matched person, comprehensive matching score, similarity score of each modality, matching confidence, scene type label, access permission information; when matching is unsuccessful: reason for matching failure, highest comprehensive matching score, abnormal feature marker.

[0093] S1036. If the identity matching is successful, the information with the highest comprehensive matching score among the information that has successfully matched the identity will be used as the matching information for the person to be verified.

[0094] S1037. Based on the matching information of the person to be verified, the comprehensive matching score, and the similarity scores of each modal feature corresponding to the matching information, generate the identity matching result.

[0095] S104. Based on the identity matching results, conduct risk level assessments and entrance monitoring and control for the personnel to be verified.

[0096] In some embodiments, risk level assessment is based on identity matching results, combined with multi-dimensional risk-related data, to quantitatively assess and classify the access security risks of the personnel to be verified. In this case, the risk level is divided into three levels: low risk, medium risk, and high risk, each corresponding to different control strategies.

[0097] In some embodiments, the stairwell monitoring and control is a series of operations performed by the stairwell monitoring robot based on the risk level assessment results, including stairwell access control, on-site monitoring, and risk warning. It is the core execution link to ensure the safety of stairwell access and can be linked with stairwell gates, back-end control terminals, and on-site personnel.

[0098] As one possible implementation, step S104 can be specifically implemented as steps S1041-S1045.

[0099] S1041. Based on the identity matching results, as well as the preset historical risk list and the historical passage records corresponding to the matching information, construct a multi-dimensional risk assessment factor system.

[0100] In some embodiments, the multi-dimensional risk assessment factor system is a set of quantifiable risk assessment indicators specifically constructed for the entrance monitoring scenario. This invention covers four core dimensions: identity matching quality, historical risk attributes, compliance of passage behavior, and real-time scenario adaptability, to identify explicit and implicit security risks.

[0101] In some embodiments, the preset historical risk list is a structured risk personnel database stored in the background management system with encryption, which is divided into a red list (high-risk wanted / controlled personnel, prohibited personnel), a yellow list (medium-risk personnel with violation records, temporary control personnel), and a white list (low-risk compliant personnel).

[0102] In some embodiments, the historical access record is a full-data encrypted archive of access records corresponding to the identity, including data such as historical access frequency, access time period, access gate, historical verification results, violation records, and permission validity period.

[0103] For example, embodiments of the present invention can simultaneously retrieve four core data sources: the first category is identity matching results, including comprehensive matching score, similarity score of each modality, matching confidence, whether the matching is successful, and feature anomaly markers; the second category is a preset historical risk list, querying whether the person to be verified is on the list, the corresponding risk level, and control requirements; the third category is historical access records corresponding to the matched identity, including access frequency in the past 3 months, time period compliance, historical verification anomaly records, and permission status; the fourth category is real-time scene parameters (people density, scene security level) and real-time behavioral characteristics of the person to be verified (loitering, deliberately obscuring the face, tailgating, unauthorized forced entry, and other abnormal behaviors).

[0104] For example, embodiments of the present invention can decompose the retrieved data source into the smallest quantifiable evaluation unit to construct a complete factor system: Dimension 1: Identity matching quality factor (core basic factor), including four sub-factors: comprehensive matching score, matching confidence, proportion of single-modal feature anomalies, and whether identity matching fails; Dimension 2: Historical risk attribute factor (core risk factor), including four sub-factors: whether on the red / yellow risk list, number of historical violations, historical risk level, and whether permissions are valid; Dimension 3: Passage behavior compliance factor (implicit risk factor), including four sub-factors: historical passage frequency, whether passage time is compliant, number of historical verification anomalies, and whether there is tailgating / forced entry behavior; Dimension 4: Real-time scenario adaptation factor (dynamic adjustment factor), including three sub-factors: current scenario security level, crowd density, and real-time abnormal behavior.

[0105] S1042. Based on a multi-dimensional risk assessment factor system, the analytic hierarchy process is used to quantify and assign weights to each risk assessment factor, and the risk assessment score of the person to be verified is obtained through risk assessment calculation.

[0106] In some embodiments, the Analytic Hierarchy Process (AHP) is a multi-dimensional weighted decision-making algorithm optimized for entrance control scenarios. It decomposes risk assessment into a three-level hierarchical structure of "target layer - criterion layer - solution layer". By combining qualitative and quantitative methods, it scientifically allocates the weights of each factor, thus solving the shortcomings of existing technologies that have fixed weights and cannot adapt to changes in scenarios.

[0107] In some embodiments, the risk assessment score is a final score that is calculated by weighted summation and quantitatively represents the overall safety risk of the personnel to be verified. The higher the score, the higher the risk level.

[0108] For example, for the entrance control scenario, a dedicated AHP hierarchical structure is built: Target layer: Risk assessment score of the personnel to be verified; Criteria layer: Four major risk factors (identity matching quality, historical risk attributes, compliance of passage behavior, and real-time scenario adaptability); Solution layer: 15 specific sub-factors corresponding to each dimension.

[0109] For example, constructing judgment matrices and calculating weights: experts in port security, park management, and identity verification are invited to construct pairwise judgment matrices based on the importance of the four dimensions of the criteria layer, and the initial weights of the four dimensions are calculated through matrix operations.

[0110] For example, dynamic weight adaptation and adjustment: The initial weights are dynamically adjusted based on the target verification strategy and the core requirements of the current scenario. For major security / high-security port scenarios: Increase the weight of historical risk attribute factors (40%) and identity matching quality factors (35%) to strengthen security control priorities; For office building / park morning rush hour high-traffic scenarios: Increase the weight of identity matching quality factors (45%) and decrease the weight of historical passage behavior factors (10%) to prioritize improving passage efficiency while ensuring basic security; For low-light / high-obscurity complex scenarios: Increase the weight of identity matching quality factors (40%) and real-time scenario adaptation factors (25%) to reduce the interference of low-quality features on risk assessment.

[0111] For example, for each sub-factor, a standardized value of 0-100 is assigned according to the preset quantification rules. Then, the final risk assessment score is obtained by combining the dimension weight and the sub-factor weight through weighted summation. The formula is: Risk assessment score = Σ(Dimension weight × Σ(Sub-factor weight × Sub-factor quantification score)).

[0112] S1043. Based on the current safety level of the stairwell scenario, dynamically adjust the risk level classification threshold.

[0113] For example, embodiments of the present invention can automatically determine the security level of the current scenario by combining the type of entrance (port / industrial park / office building), preset security plan, time node, and population density: Level 1 (Major Security Level): scenarios with the highest security requirements such as port entry and exit, major event sites, and confidential industrial parks; Level 2 (Holiday Level): scenarios with large passenger flow in transportation hubs and commercial parks during statutory holidays and weekends; Level 3 (Daily Level): routine scenarios in office buildings and industrial parks during off-peak hours on weekdays; Level 4 (Peak Traffic Level): commuting scenarios in office buildings and industrial parks during morning and evening peak hours on weekdays, with the core requirement being to improve traffic efficiency.

[0114] For example, the present invention can dynamically adjust the thresholds for classifying low, medium, and high risks based on security levels, with the following adjustment rules: Level 1 (Major Security Level): Low risk 0-30 points, medium risk 30-70 points, high risk ≥70 points, tightening the high-risk threshold to minimize security risks; Level 2 (Holiday Level): Low risk 0-35 points, medium risk 35-75 points, high risk ≥75 points, balancing security control with the needs of large passenger flow; Level 3 (Daily Level): Low risk 0-40 points, medium risk 40-80 points, high risk ≥80 points, standard thresholds for regular scenarios; Level 4 (Peak Traffic Level): Low risk 0-45 points, medium risk 45-85 points, high risk ≥85 points, relaxing the low-risk threshold to reduce unnecessary review and interception, and alleviate peak congestion.

[0115] S1044. Based on the risk assessment score of the person to be verified and the risk level classification threshold, determine the risk level of the person to be verified.

[0116] In some embodiments, the risk levels include high risk, medium risk, and low risk.

[0117] For example, in this embodiment of the invention, the risk assessment score of the person to be verified can be compared with a dynamically adjusted threshold to complete the classification: score < low risk threshold: classified as low risk level; low risk threshold ≤ score < high risk threshold: classified as medium risk level; score ≥ high risk threshold: classified as high risk level.

[0118] For example, individuals whose identity fails to match will be directly upgraded to a medium-risk level, regardless of their score; high-risk individuals on the red list and those exhibiting serious abnormal behaviors such as forced entry / tailgating will be directly upgraded to a high-risk level; medium-risk individuals on the yellow list will be classified as medium-risk at the very least and will not be downgraded to low-risk.

[0119] For example, only for peak traffic scenarios, for individuals with initial medium-risk or no-risk lists and whose identity matching scores are close to the threshold, a downgrade verification is performed. After verifying their historical compliance with traffic regulations, they can be downgraded to a low-risk level, thereby maximizing traffic efficiency during peak hours.

[0120] S1045. Based on the risk level of the personnel to be verified, implement differentiated monitoring and control.

[0121] Low-risk individuals are automatically allowed to pass and their passage records are retained. Medium-risk individuals trigger a second multimodal verification and manual review. High-risk individuals trigger sound and light warnings, channel isolation, and information is simultaneously pushed to the back-end control terminal.

[0122] For example, the management and control of low-risk personnel is implemented as follows: the elevator entrance monitoring robot sends an automatic release command to the turnstile, the turnstile opens and releases the person, and the whole process is done without human intervention. The release time for a single person is ≤200ms, maximizing the efficiency of passage. At the same time, a full encrypted record of all personnel passage is stored, including identity information, passage time, elevator entrance number, risk level, and matching score, and is updated to the historical passage record database. In peak passage scenarios, the turnstile fast passage mode is activated simultaneously to shorten the time for the turnstile to open and close, further improve the passage speed, and alleviate queuing congestion.

[0123] For example, the management and control of personnel at medium risk levels is implemented as follows: The elevator entrance monitoring robot immediately triggers a voice prompt to guide personnel to the side verification channel, without occupying the main passageway and avoiding affecting the efficiency of the main passageway; a secondary multimodal verification is initiated, increasing the recognition weight of the high-reliability modal of iris and fingerprint, tightening the matching threshold, and completing the secondary identity verification; personnel who pass the secondary verification are automatically released and the verification record is retained; personnel who fail the secondary verification are immediately triggered to trigger the manual review process, and the personnel's identity information, matching results, risk factors, and real-time images are simultaneously pushed to the back-end management terminal for remote / on-site review by security personnel. If the review is passed, the personnel are released; if the review is failed, the management is upgraded to high risk control.

[0124] For example, the management and control of high-risk personnel is implemented as follows: The elevator entrance monitoring robot immediately activates on-site audible and visual warnings, and simultaneously sends a locking command to the elevator entrance gate to achieve temporary isolation of the passage. The robot moves synchronously to the passage entrance to form a physical barrier, prohibiting personnel from passing through; within 100ms, the robot transmits the personnel's identity information, risk level, cause of risk, real-time video, and location information; and initiates full-process key monitoring, with the robot tracking the personnel's location and behavior in real time, recording the entire process to retain tamper-proof video evidence, and prohibiting personnel from entering the controlled area until on-site security personnel arrive to handle the situation.

[0125] This invention provides an identity verification method based on biometric recognition and a stairwell monitoring robot. The method simultaneously collects multimodal biometric features of the person to be verified along with current stairwell scene parameters. Combined with an adaptive identity verification model, it dynamically adjusts the verification strategy to adapt to different scenarios, solving the problem that fixed verification modes cannot adapt to complex stairwell scenes. This improves the adaptability and accuracy of biometric recognition under different pedestrian flow and lighting conditions. Through real-time comparison of the processed biometric features with a pre-set identity information database, efficient and accurate verification of personnel identity is achieved. The identity matching results are used to complete risk level assessment and differentiated stairwell monitoring and control. This improves stairwell passage efficiency while strengthening precise control of high-risk personnel, balancing the security, accuracy, and passage efficiency of stairwell scene identity verification.

[0126] Optionally, the identity verification method based on biometric recognition and ladder monitoring robot provided in this embodiment of the invention further includes steps S201-S205 before step S102.

[0127] S201. Collect scene parameter samples, multimodal biometric samples, and corresponding scene verification requirement samples for multiple ladder entrances under different known scenarios. In some embodiments, the known scenario is a standard entrance scenario with pre-defined scenario types and quantified scenario parameters, including core operating conditions such as high / low pedestrian flow, strong / weak lighting, and high / low occlusion.

[0128] In some embodiments, the scene parameter samples are standardized samples of four core parameters—human flow density, light intensity, passage speed, and occlusion degree—pre-collected and labeled with ground truth values ​​for a known scene.

[0129] In some embodiments, the multimodal biometric samples are three types of feature samples—facial, iris, and fingerprint—with ground truth labels for personal identity in known scenarios, covering actual traffic conditions such as different ages, genders, wearing masks / glasses, fingerprint wear, and facial occlusion.

[0130] In some embodiments, the verification requirement samples are pre-calibrated core verification requirements that meet the scenario control requirements under known scenarios, and are divided into three categories: efficiency priority, accuracy priority, and security priority.

[0131] S202. Construct model training samples based on scene parameter samples and multimodal biometric samples under different known scenarios.

[0132] S203. Based on the model training samples and the pre-built neural network architecture, the scene recognition module is trained.

[0133] For example, embodiments of the present invention can address the edge computing power limitations and multi-dimensional scene classification requirements of stairwell scenarios by building a lightweight multi-label classification neural network architecture optimized based on MobileNetV3. The core consists of three parts: a feature extraction layer composed of depthwise separable convolutional blocks, which extracts deep semantic information of scene-feature association features, significantly reducing the amount of computation; a multi-label classification layer composed of three parallel fully connected branches, corresponding to the classification outputs of pedestrian flow, illumination, and occlusion dimensions, respectively, to achieve simultaneous determination of multi-dimensional scenes; and an output layer that outputs the classification results and classification confidence of each dimension through a sigmoid activation function.

[0134] For example, the pre-trained weight loading and training parameter settings are as follows: A transfer learning strategy is adopted, and the feature extraction layer weights pre-trained on a large-scale public scene classification dataset are loaded as the initial weights of the model to avoid the overfitting problem of training the model from scratch; for multi-label classification tasks, the binary cross-entropy loss function (BCEWithLogitsLoss) is used as the model's loss function, and the L2 regularization coefficient is set to further suppress model overfitting; the optimizer adopts the Adam optimizer, with the initial learning rate set to 0.001, and the learning rate decays by 10% every 10 training epochs.

[0135] For example, model iterative training and early stopping mechanism: The packaged model training samples are input into the built neural network in batches for iterative training; during the training process, after each epoch, the classification accuracy of the model is verified with validation set samples. If the validation set accuracy does not improve for 5 consecutive epochs, the early stopping mechanism is triggered to terminate the training in advance to avoid model overfitting.

[0136] For example, model performance testing and optimization: After training is terminated, the model is subjected to industrial-grade performance testing using completely independent test set samples. The core test indicators include: scene classification accuracy, single-sample inference time, and classification confidence calibration error. The model is required to have a scene classification accuracy of ≥99% and a single-sample inference time at the robot edge of ≤50ms to meet the efficiency requirements of real-time identity verification. If the test results are not up to standard, the model is retrained by expanding the training samples, adjusting the model hyperparameters, and optimizing the network structure until the performance requirements are met.

[0137] S204. Based on verification requirement samples under different known scenarios, design corresponding multimodal feature recognition weights and verification accuracy parameters, and construct a verification strategy template library.

[0138] For example, specific strategy design rules are formulated for each known scenario type and corresponding verification requirement sample: High-traffic efficiency priority scenario: Prioritize compressing the time consumed in a single round of verification, increase the weight of facial features, and appropriately reduce the accuracy requirements of non-core modalities to solve the problem of queuing and congestion in high-traffic scenarios; Low-light / high-occlusion accuracy priority scenario: Prioritize ensuring recognition accuracy, increase the weight of iris / fingerprint features that are less affected by the environment, tighten the accuracy threshold, and solve the problem of large recognition errors in complex scenarios; Major security priority scenario: Full-modal high-precision verification, tighten all matching thresholds, improve the risk control level, and meet the needs of high-security scenarios; Daily balance scenario: Balance accuracy and efficiency, balance the weight of each modality, and adapt to regular passage needs.

[0139] For example, for each known scenario, a corresponding standardized verification strategy template is designed according to the mapping rules. Each template contains four types of core parameters to achieve full-link linkage between the verification strategy and the acquisition device: weight parameters: recognition weight, fusion weight, and comparison priority of facial, iris, and fingerprint features; accuracy parameters: similarity threshold of each modality feature, minimum number of feature point matches, and identity matching threshold; efficiency parameters: frame rate of the acquisition device, database retrieval range, and verification process steps; and device parameters: configuration parameters of multimodal acquisition devices such as camera focal length, exposure, and sensor sensitivity.

[0140] For example, for each strategy template, on-site testing is conducted in the corresponding scenario to collect core indicators such as identity verification accuracy, single-round verification time, personnel passage efficiency, false positive rate, and false negative rate. If the measured indicators do not meet the scenario requirements, the template parameters are iteratively optimized. For example, if the single-round verification time exceeds the standard in a high-traffic scenario, the comparison process is optimized and the retrieval efficiency is improved until the template fully adapts to the accuracy, efficiency, and security requirements of the scenario.

[0141] S205. Integrate the scene recognition module with the verification strategy template library, and incorporate feature association and parameter calibration algorithms to obtain an adaptive identity verification model.

[0142] Thus, this invention can complete the construction of an end-to-end adaptive identity verification model by building a dedicated dataset for the entire scenario of the stairwell, training a lightweight scenario recognition module, and building a scenario-adaptive verification strategy template library. This addresses the shortcomings of existing technologies in terms of the disconnect between scenarios and verification strategies from the bottom up, and provides core algorithm support for dynamic adaptive verification.

[0143] Optionally, the identity verification method based on biometric recognition and ladder monitoring robot provided in this embodiment of the invention further includes steps S301-S306.

[0144] S301. Real-time collection of scene parameters, biometric data, identity matching results, and risk control feedback data during the monitoring process at each stairwell.

[0145] For example, scene parameters include pedestrian density, light intensity, traffic speed, degree of occlusion, and scene type and classification confidence output by the scene recognition module.

[0146] For example, biometric data collection includes adaptive enhancement features, recognition confidence of each modality feature, operating parameters of the collection device, collection success rate, and feature quality score.

[0147] For example, the identity matching result includes a comprehensive matching score, similarity scores for each modality, whether the match is successful, the matching confidence level, and the annotation of the reason for the matching failure.

[0148] For example, risk control feedback data includes risk level assessment results, control action execution status, final handling results of security personnel, manual annotation of misjudgments / missed judgments, and traffic efficiency statistics (average time per person, queue length, and gate throughput rate).

[0149] S302. Based on the scene parameters, biometric data collection, identity matching results and risk control feedback data during the monitoring process at each ladder entrance, generate a model iteration dataset.

[0150] S303. Filter the model iteration dataset, remove abnormal and invalid data, and merge it with the model training dataset to obtain the merged dataset.

[0151] S304. Based on the fused dataset, the parameters of the classification algorithm and verification strategy template library of the scene recognition module are fine-tuned, and the weight allocation logic of each modality feature is optimized to obtain the iteratively updated adaptive identity verification model.

[0152] For example, the incremental fine-tuning training scheme is set as follows: the weights of the basic feature extraction layer of the adaptive authentication model are frozen, and only the top multi-label classification layer, parameter calibration algorithm module, and weight allocation module of the scene recognition module are fine-tuned to avoid the loss of the model's basic capabilities; the optimizer adopts the AdamW optimizer with a very small initial learning rate (1e-5), and only the model parameters are slightly adjusted to ensure that the model learns new scene features while retaining its original capabilities; the loss function is set separately for different optimization objectives: the scene recognition module adopts the optimized binary cross-entropy loss function, and the weight allocation module adopts the mean squared error loss function, focusing on reducing the loss value of difficult samples and strengthening the model's adaptability to complex scenes.

[0153] For example, the scene recognition module is optimized in a targeted manner: based on the scene recognition optimization subset of the fusion dataset, the classification layer of the scene recognition module is fine-tuned, with a focus on optimizing the classification accuracy of complex scenes such as high pedestrian traffic, high occlusion, low light, and backlight.

[0154] For example, the optimization of verification strategy template library parameters: Based on the traffic efficiency and recognition accuracy data of the verification strategy optimization subset, the strategy parameters corresponding to each scenario in the template library are optimized, and the recognition weight, accuracy threshold and device parameter configuration of each modality are adjusted; for example, for high traffic scenarios, the time consumption of single-round verification is reduced by ≥15% after optimization, while maintaining the recognition accuracy; for low light scenarios, the false judgment rate is reduced by ≥30% after optimization.

[0155] For example, the weight allocation logic is optimized by optimizing the misjudged samples and matching result data of the weight allocation optimization subset, improving the weight allocation algorithm of each modality feature, increasing the weight ratio of high reliability features, and reducing the interference of low quality features on the matching results.

[0156] S305. Conduct offline testing and online trial runs on the iteratively updated adaptive authentication model to verify the model's recognition accuracy and adaptability, and determine whether the iteratively updated adaptive authentication model is abnormal.

[0157] S306. If there are no abnormalities during operation, replace the original model with the iteratively updated adaptive authentication model, and use the iteratively updated adaptive authentication model for ladder gate authentication.

[0158] Thus, this invention achieves continuous iterative optimization of the adaptive identity verification model through a closed-loop process of full-link operation data collection, incremental learning fine-tuning, multi-dimensional verification, and gray-scale deployment. This solves the pain points of existing solutions, such as fixed models and long-term performance degradation, and ensures the accuracy and efficiency of gate identity verification in the long term.

[0159] Optionally, the identity verification method based on biometric recognition and ladder monitoring robot provided in this embodiment of the invention further includes steps S401-S404.

[0160] S401. Real-time acquisition of scene type and pedestrian density data at the current elevator entrance.

[0161] S402. When a scene is identified as a high-traffic area and the crowd density exceeds a high threshold, increase the facial feature acquisition rate to reduce the acquisition accuracy of non-core modalities, and adjust the camera focal length to adapt to multi-person acquisition in order to shorten the acquisition time of single-person biometric features.

[0162] For example, after triggering the adjustment command for high-traffic scenes, the acquisition frame rate of the facial recognition camera is increased from the usual 30fps to 60fps, and a lightweight face detection model is enabled, compressing the face detection inference time from 100ms to less than 30ms, thereby achieving simultaneous detection, tracking and feature acquisition of multiple faces.

[0163] For example, embodiments of the present invention can reduce the acquisition accuracy of non-core iris features from the conventional 2048×2048 resolution to 1024×1024 resolution, turn off the fine texture scanning mode of iris, and compress the time for a single round of iris acquisition from 200ms to less than 80ms; at the same time, it locks the basic acquisition accuracy of core facial and fingerprint features, taking into account both passage efficiency and security bottom line.

[0164] For example, in this embodiment of the invention, the facial camera can be switched from single-person face focusing mode to wide-angle multi-person tracking mode, and the lens focal length can be adjusted to a 3.5mm wide-angle mode, with the focusing range covering a queuing area of ​​0.3-2 meters. The facial features of the first 3 people to be verified in the queue can be pre-collected simultaneously. When the person walks to the gate verification position, only the fingerprint features need to be collected to complete the full collection, reducing the time for single-person full collection from the usual 300ms to less than 150ms, greatly improving the passage efficiency.

[0165] For example, embodiments of the present invention can statistically analyze the average collection time per person, queue length, and gate passage rate in real time. If the queue length continues to exceed a preset threshold, the iris collection process is further shut down, and only the face + fingerprint dual-modal collection is retained to minimize the collection time. If the passage efficiency returns to normal, the basic collection parameters are gradually restored to achieve dynamic closed-loop adjustment and avoid parameter stagnation.

[0166] S403. When the scene is determined to be a medium-sized pedestrian flow scenario and the pedestrian flow density is within the medium threshold range, maintain the core modality acquisition accuracy, adjust the angle and position of the acquisition device to reduce the queuing time of the people to be verified, and enable parallel acquisition through dual acquisition channels.

[0167] For example, after triggering the adjustment command in a crowded scene, the high-precision acquisition parameters of the face, iris and fingerprint modalities are locked, and the conventional acquisition resolution of 2048×2048 and full texture scanning mode are maintained to ensure that the core accuracy of identity verification is not less than 99.5% and achieve bidirectional balanced adaptation.

[0168] For example, embodiments of the present invention can detect key points of the human body through a facial camera, and identify the height, facial position, and standing distance of the person to be verified in real time. The robot's electric gimbal can adjust the pitch angle (adjustment range -30° to +30°) and horizontal rotation angle of the acquisition device in real time. The movable chassis can fine-tune the front and back position of the device so that the acquisition device is always in the optimal acquisition position with the person's face and fingers.

[0169] For example, in this embodiment of the invention, the multimodal acquisition device can be divided into two independent parallel acquisition threads: the first channel is a face + iris vision acquisition channel, and the second channel is a fingerprint acquisition channel. After a person stands up, the two channels start acquisition simultaneously, which greatly reduces the waiting time for the person to stand.

[0170] For example, in this embodiment of the invention, the number of people in the verification queue and the waiting time can be monitored in real time by a camera. When the number of people in the queue exceeds 3, the pre-collection mode is activated, and the facial features of the second person in the queue are pre-collected. When the person walks to the verification position, only the iris and fingerprint features need to be collected, which further reduces the queuing time and avoids the queue from continuing to grow into a state of high congestion.

[0171] S404. When the scene is determined to be a low-traffic scene and the traffic density is below the low threshold, improve the acquisition accuracy of each modal feature and adjust the device exposure and sensor sensitivity to adapt to the current lighting scene.

[0172] For example, after triggering the adjustment command for low-traffic scenarios, the acquisition parameters of the three modalities are fully upgraded: the facial camera resolution is increased from 1080P to 4K, and infrared + visible light dual-modal fusion acquisition is enabled; the iris scanner resolution is increased from 2048×2048 to 4096×4096, and the ring fine texture full scan mode is enabled; the fingerprint sensor is increased from 508DPI to 1000DPI, and the three-level ridge detail scan mode is enabled to capture more subtle unique features.

[0173] For example, embodiments of the present invention can collect real-time data on the light intensity and angle of the current stairwell using a light sensor, and dynamically adjust the exposure parameters of the acquisition device by combining the brightness histograms of the facial and iris images: in low-light scenarios, the exposure time is extended, the gain is moderately increased, and infrared supplementary lighting without red exposure is enabled to avoid the face being too dark and the iris texture being unclear; in strong backlight scenarios, HDR wide dynamic range mode is enabled, the aperture size is adjusted, the strong light overexposure is suppressed, the details of the dark areas of the face are restored, and the recognition accuracy is improved.

[0174] For example, embodiments of the present invention can identify the dryness or wetness of the finger of the person to be verified and the degree of fingerprint wear through real-time feedback from the fingerprint sensor, and dynamically adjust the sensor’s sensitivity, scanning frequency and driving voltage: for dry fingers and worn fingerprints, increase the sensor’s sensitivity and scanning frequency; for wet fingers, decrease the sensitivity and enable the surface stain filtering algorithm.

[0175] For example, after each feature acquisition is completed, the quality of the feature image is scored in real time from multiple dimensions. If the quality score is lower than the preset high-precision threshold, a second adaptive adjustment acquisition is immediately triggered, automatically optimizing parameters such as exposure, sensitivity, and focal length, and re-acquiring high-quality features to improve the security of identity verification in low-traffic scenarios.

[0176] Thus, this invention can dynamically adapt the parameters of multimodal acquisition devices driven by scene type and pedestrian density, solve the pain point of poor adaptability of fixed parameters from the source of acquisition, balance the acquisition accuracy and passage efficiency in different scenarios, form a closed loop with backend adaptive verification, and improve the full-scenario adaptability of stairwell identity verification.

[0177] 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 the present invention.

[0178] Figure 2 This diagram illustrates the structure of an authentication device for a ladder-entry monitoring robot based on biometric recognition, according to an embodiment of the present invention. The authentication device 500 includes a communication module 501 and a processing module 502.

[0179] The communication module 501 is used to collect the multimodal biometrics of the person to be verified and the scene parameters of the current ladder entrance. The multimodal biometrics include facial features, iris features and fingerprint features.

[0180] The processing module 502 is used to analyze the application scenario and dynamically adjust the verification strategy based on multimodal biometrics, scene parameters, and an adaptive identity verification model to obtain the biometrics after recognition processing; to perform real-time comparison based on the biometrics after recognition processing and a preset identity information database to obtain the identity matching result; and to conduct risk level assessment and access control management for the personnel to be verified based on the identity matching result.

[0181] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 600 includes: a processor 601, a memory 602, and a computer program 603 stored in the memory 602 and executable on the processor 601. When the processor 601 executes the computer program 603, it implements the steps in the above-described method embodiments. Alternatively, when the processor 601 executes the computer program 603, it implements the functions of each module / unit in the above-described device embodiments.

[0182] For example, computer program 603 may be divided into one or more modules / units, one or more of which are stored in memory 602 and executed by processor 601 to complete the present invention. One or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 603 in electronic device 600.

[0183] The processor 601 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0184] The memory 602 can be an internal storage unit of the electronic device 600, such as a hard disk or RAM of the electronic device 600. The memory 602 can also be an external storage device of the electronic device 600, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or FlashCard equipped on the electronic device 600. Furthermore, the memory 602 can include both internal and external storage units of the electronic device 600. The memory 602 is used to store computer programs and other programs and data required by the terminal. The memory 602 can also be used to temporarily store data that has been output or will be output.

[0185] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An authentication method based on biometric recognition and a ladder gate monitoring robot, characterized in that, The ladder entrance monitoring robot includes a multimodal biometric acquisition device, and the method includes: Collect multimodal biometrics of the person to be verified and scene parameters of the current ladder entrance. The multimodal biometrics include facial features, iris features, and fingerprint features. Based on the multimodal biometrics, the scene parameters, and the adaptive authentication model, the application scenario is analyzed and the authentication strategy is dynamically adjusted to obtain the biometrics after recognition processing. Based on the biometric features after identification processing, and a preset identity information database, a real-time comparison is performed to obtain the identity matching result; Based on the identity matching results, risk level assessment and entrance monitoring and control are carried out on the personnel to be verified.

2. The authentication method based on biometric recognition and ladder gate monitoring robot according to claim 1, characterized in that, The scene parameters include pedestrian density, light intensity, traffic speed, and degree of occlusion; The process involves analyzing the application scenario and dynamically adjusting the verification strategy based on the multimodal biometrics, the scene parameters, and the adaptive authentication model to obtain the processed biometrics, including: The scene parameters are normalized to obtain the scene feature vector; The multimodal biometric features are denoised, aligned, and adaptively preprocessed to obtain adaptively enhanced features; The scene feature vector and the adaptive enhancement feature are input into the adaptive identity verification model to determine the scene type and match the verification strategy to obtain the target verification strategy. Based on the target verification strategy, cross-modal fusion operations are performed on the multimodal biometrics to obtain the biometrics after recognition processing.

3. The identity verification method based on biometric recognition and ladder gate monitoring robot according to claim 2, characterized in that, The step of inputting the scene feature vector and the adaptive enhancement feature into the adaptive identity verification model to determine the scene type and match the verification strategy to obtain the target verification strategy includes: The scene feature vector is associated with the adaptive enhancement feature, and redundant scene information that is irrelevant to biometric recognition is removed to generate an associated feature set. Based on the scene recognition module in the adaptive identity verification model, the associated feature set is classified and identified to determine the scene type of the current stairwell. The scene type includes high or low pedestrian flow, strong or weak light, and high or low occlusion. Based on the scenario type and the verification strategy template library in the adaptive authentication model, multiple verification strategies corresponding to the scenario type are matched and obtained. Based on the multiple verification strategies, the recognition weights and verification accuracy parameters of each modality feature in the verification strategy are dynamically adjusted to select the target verification strategy.

4. The authentication method based on biometric recognition and ladder gate monitoring robot according to claim 3, characterized in that, The process of dynamically adjusting the recognition weights and verification accuracy parameters of each modality feature in the verification strategies based on the multiple verification strategies, and then selecting the target verification strategy, includes: Extract the recognition weights and verification accuracy parameters of facial features, iris features, and fingerprint features from each verification strategy; Based on the recognition confidence of the adaptive enhancement features, the recognition weights and verification accuracy parameters of facial features, iris features, and fingerprint features in each verification strategy are normalized and calibrated to obtain normalized verification parameters; the recognition confidence of the adaptive enhancement features is obtained by quality verification after adaptive preprocessing during the adaptive enhancement feature generation process. Based on the normalized verification parameters and the core requirements of the current entrance scenario, priority is sorted and the adaptation score of each verification strategy is calculated. The core requirements include prioritizing efficiency in high-traffic scenarios, prioritizing iris recognition in low-light scenarios, and prioritizing fingerprint recognition in high-obstruction scenarios. Select the verification strategy with the highest fit score as the target verification strategy.

5. The authentication method based on biometric recognition and ladder gate monitoring robot according to claim 2, characterized in that, The step of performing cross-modal fusion operations on the multimodal biometrics based on the target verification strategy to obtain the biometrics after recognition processing includes: Based on the target verification strategy, the fusion weights and fusion algorithms for three modalities of facial, iris, and fingerprint features are determined; An attention mechanism is used to effectively locate and extract key features of biometric features of each modality, thereby obtaining the key features of each modality. Based on the key features of each modality, as well as the fusion weights and fusion algorithm, feature mapping and fusion are performed to obtain the biometric features after recognition processing.

6. The authentication method based on biometric recognition and ladder gate monitoring robot according to claim 1, characterized in that, The process of performing real-time comparison based on the biometric features after identification processing and a preset identity information database to obtain an identity matching result includes: The biometric features after identification processing are subjected to feature vector extraction, and then encrypted using an asymmetric encryption algorithm to obtain the encrypted feature vector. Based on the encrypted feature vector, a distributed hierarchical retrieval algorithm is used to filter out a set of candidate identity information with a similarity greater than a similarity threshold from a preset identity information database. For each piece of information in the candidate identity information set, a hierarchical and precise comparison of facial features, iris features, and fingerprint features is performed in sequence, and the similarity score of each modality feature is calculated. Based on the fusion weights of each modal feature in the target verification strategy, the similarity scores of each modal feature are weighted and fused to generate a comprehensive matching score for each piece of information in the candidate identity information set. The overall matching score of each piece of information is compared with the dynamically adjusted matching threshold to determine whether the identity matching is successful. If the identity is matched, the information with the highest overall matching score among the information that has been matched will be used as the matching information for the person to be verified. The identity matching result is generated based on the matching information of the person to be verified, the comprehensive matching score, and the similarity score of each modal feature corresponding to the matching information.

7. The authentication method based on biometric recognition and ladder gate monitoring robot according to claim 1, characterized in that, The step of conducting risk level assessment and entrance monitoring and control for the person to be verified based on the identity matching result includes: Based on the identity matching results, as well as the preset historical risk list and the historical access records corresponding to the matching information, a multi-dimensional risk assessment factor system is constructed. Based on the multi-dimensional risk assessment factor system, the analytic hierarchy process is used to quantify and assign weights to each risk assessment factor, and the risk assessment score of the person to be verified is obtained through risk assessment calculation. Based on the current safety level of the stairwell scenario, dynamically adjust the risk level classification threshold; Based on the risk assessment score of the person to be verified and the risk level classification threshold, the risk level of the person to be verified is determined; the risk level includes high risk level, medium risk level and low risk level. Based on the risk level of the person to be verified, differentiated monitoring and control are implemented; low-risk personnel are automatically allowed to pass and their passage records are retained; medium-risk personnel trigger a second multimodal verification and manual review linkage; high-risk personnel activate sound and light warnings, channel isolation, and simultaneously push information to the background control terminal.

8. The authentication method based on biometric recognition and ladder gate monitoring robot according to claim 1, characterized in that, The process of analyzing application scenarios and dynamically adjusting verification strategies based on the multimodal biometrics, the scene parameters, and the adaptive authentication model to obtain the processed biometrics also includes: Collect scene parameter samples, multimodal biometric samples, and corresponding scenario verification requirement samples for multiple ladder entrances under different known scenarios; Based on the scene parameter samples and multimodal biometric samples under the different known scenarios, model training samples are constructed; The scene recognition module is trained based on the model training samples and the pre-built neural network architecture. Based on the verification requirement samples under the different known scenarios, we design corresponding multimodal feature recognition weights and verification accuracy parameters, and construct a verification strategy template library. The scene recognition module is integrated with the verification strategy template library, and feature association and parameter calibration algorithms are incorporated to obtain the adaptive identity verification model.

9. The authentication method based on biometric recognition and ladder gate monitoring robot according to claim 1, characterized in that, The method further includes: Real-time collection of scene parameters, biometric data, identity matching results, and risk control feedback data during the monitoring process at each stairwell; Based on the scene parameters, biometric data collection, identity matching results, and risk control feedback data during the monitoring process at each ladder entrance, a model iteration dataset is generated. The model iteration dataset is filtered to remove abnormal and invalid data, and then fused with the model training dataset to obtain the fused dataset. Based on the fused dataset, the parameters of the classification algorithm and verification strategy template library of the scene recognition module are fine-tuned, and the weight allocation logic of each modality feature is optimized to obtain the iteratively updated adaptive identity verification model. Offline testing and online trial operation were conducted on the iteratively updated adaptive authentication model to verify the model's recognition accuracy and adaptability, and to determine whether the iteratively updated adaptive authentication model was abnormal. If there are no abnormalities during operation, the original model will be replaced by the iteratively updated adaptive authentication model, and the ladder gate authentication will be performed using the iteratively updated adaptive authentication model.

10. The authentication method based on biometric recognition and ladder gate monitoring robot according to claim 1, characterized in that, The method further includes: Real-time acquisition of scene type and pedestrian density data at the current elevator entrance; When a scene is identified as a high-traffic area and the crowd density exceeds a high threshold, the facial feature acquisition rate is increased to reduce the acquisition accuracy of non-core modalities, and the camera focal length is adjusted to adapt to multi-person acquisition in order to shorten the acquisition time of single-person biometric features. When the scene is determined to be a medium-sized pedestrian flow scenario and the pedestrian flow density is within the medium threshold range, the core modality acquisition accuracy is maintained, the angle and position of the acquisition device are adjusted to reduce the queuing time of the people to be verified, and dual acquisition channels are enabled for parallel acquisition. When the scene is determined to be a low-traffic scene and the traffic density is below the low threshold, the accuracy of feature acquisition for each modality is improved, and the device exposure and sensor sensitivity are adjusted to adapt to the current lighting scene.

Images