High-precision AI roll call system and method based on multi-modal biological feature fusion

Abstract

The invention relates to the technical field of identity recognition, in particular to a high-precision AI roll call system and method based on multi-modal biological feature fusion, and the system comprises a data collection module, a feature processing module, a registration template database, a fusion decision module, a feature updating module, and a business logic and application module. Compared with the prior art that identity recognition generally depends on a single biological feature such as a human face, and the recognition rate is remarkably reduced when light is poor, the face is blocked or malicious cheating occurs, the method adopts image and audio sensors for synchronous acquisition, and combines the three biological features of the human face, the voiceprint and the gait for joint decision making, so that the recognition rate is greatly improved. The method has the advantage of multi-modal information complementation, can effectively cope with complex environment interference and spoofing attacks, and remarkably improves the accuracy and robustness of a roll call system.

Description

Technical Field

[0001] This invention relates to the field of identity recognition technology, and in particular to a high-precision AI roll call system and method based on multimodal biometric fusion. Background Technology

[0002] In daily teaching, corporate attendance, and meeting sign-in scenarios, accurate and efficient identity verification and roll call recording are fundamental and crucial tasks. Traditional roll call methods, such as manual verbal roll call or card signing, are inefficient and prone to proxy signing, making them unsuitable for modern management needs. Therefore, automated roll call systems based on biometric recognition technology have emerged, aiming to achieve seamless and accurate identity authentication through individual physiological or behavioral characteristics, thereby improving management efficiency and fairness.

[0003] However, most existing automated attendance systems rely on a single biometric feature, such as facial recognition or voiceprint recognition. These systems face significant challenges in practical deployment: First, the performance of a single modality is highly susceptible to environmental factors; for example, changes in lighting can affect facial recognition, and ambient noise can severely degrade voiceprint recognition. Second, security vulnerabilities exist, such as the ability to deceive the system using facial photos or audio recordings. Furthermore, users' biometric features change slowly over time, and static template databases cannot adapt to these changes, causing system performance to degrade with prolonged use. Finally, the system architecture is often rigid, making it difficult to extend and integrate new biometric modalities to address new security challenges or scenario requirements.

[0004] To address the above problems, this invention proposes a high-precision AI roll call system and method based on multimodal biometric fusion. Summary of the Invention

[0005] To overcome the problems mentioned in the background art, this invention proposes a high-precision AI attendance system and method based on multimodal biometric fusion.

[0006] The technical solution of this invention is: a high-precision AI attendance system based on multimodal biometric fusion, comprising: The data acquisition module is used to simultaneously acquire raw data of multimodal biometrics, including image sensors and audio sensors; The feature processing module is used to process the collected biometric data of each modality separately and extract the corresponding biometric templates. A registration template database is used to securely store baseline biometric templates for authorized users; The fusion decision module receives the matching similarity scores of each modality, executes a dynamic weighted fusion algorithm, and outputs the final identity recognition result. The feature update module is used to analyze the biometric templates extracted by the feature processing module and update the baseline biometric templates in the registration template database. The business logic and application module is used to manage attendance tasks, record attendance results, generate statistical reports, and provide a human-computer interaction interface.

[0007] Preferably, the feature processing module includes: A11: Face feature extraction unit, runs the first neural network model, and is used to extract the first feature vector from the face image through the first neural network model; A12: Voiceprint feature extraction unit, which runs a second neural network model to extract a second feature vector from the voiceprint audio through the second neural network model; A13: Gait feature extraction unit, which runs a third neural network model to extract a third feature vector from gait video sequences.

[0008] Preferably, the facial feature extraction unit, when in operation, specifically includes: S11: Receive raw image data containing a human face acquired by an image sensor; S12: Preprocess the original image and locate the face region in the image using a face detection algorithm; S13: Input the preprocessed standard face image into the first neural network model for feature extraction. The first neural network model is a deep convolutional neural network, and the output of the last hidden layer of the first neural network model is defined as the first feature vector. S14: Normalize the first feature vector to obtain a first feature vector with fixed length and normalized modulus, which is used as the final face biometric template output.

[0009] Preferably, the voiceprint feature extraction unit, when in operation, specifically includes: S21: Receive raw audio data acquired by the audio sensor; S22: Preprocess the raw audio data. First, pre-emphasize and window the audio data. Then, perform endpoint detection on each frame of audio data to remove silent segments and noise segments, and retain valid speech frames. S23: Extract acoustic feature parameters from valid speech frames to generate a speaker signature map, including Mel frequency cepstral coefficients; S24: Input the generated voiceprint feature map into the second neural network model, wherein the second neural network model is a neural network based on time-series modeling; S25: Deep feature learning and mapping of the voiceprint feature map is performed through the second neural network model, and the second feature vector is output from the fully connected layer specified by the model. S26: Normalize the second feature vector to obtain the voiceprint biometric template.

[0010] Preferably, the gait feature extraction unit includes the following components during operation: S31: Receive a gait video sequence containing pedestrian targets acquired by an image sensor; S32: Use object detection algorithms to process each frame of the video and locate pedestrian areas; S33: Based on the pedestrian region, background subtraction and contour extraction are performed to obtain the binary contour sequence of the pedestrian. The binary contour sequence is then aligned and normalized in the time dimension to construct standardized gait sequence data. S34: Input the preprocessed standardized gait sequence data into the third neural network model, wherein the third neural network model is a three-dimensional convolutional neural network; S35: The spatiotemporal features in the gait sequence are extracted and modeled hierarchically using a third neural network model to obtain the third feature vector; S36: Normalize the third feature vector to obtain the final gait biometric template.

[0011] As a preferred option, the fusion decision module, when in operation, specifically includes: S41: Compare the biometric template to be identified extracted by the feature processing module with the baseline biometric template of the corresponding user in the registration template database, and calculate the matching similarity score of each biometric modality. S42: Dynamically calculate the corresponding fusion weights for the matching similarity scores of each modality; S43: Using the calculated dynamic weights, the similarity scores of each match are weighted and fused to generate a fusion score; the weighted fusion calculation adopts a weighted summation method. S44: Compare the final fusion score with the preset recognition threshold: A. If the final fusion score is higher than the recognition threshold, the recognition is considered successful, and the corresponding user's identity identifier is output as the final identity recognition result; B. If the final fusion score is lower than the recognition threshold, it is determined as a recognition failure, and a recognition failure message is output.

[0012] Preferably, when dynamically calculating the fusion weights for the matching similarity scores of each modality, the calculation is based on the following factors: A21: Quality factor, based on the quality assessment score of the raw data of each modality collected this time. Modalities with higher data quality are assigned higher weights. A22: Confidence factor, based on the confidence of the matching quality generated by each modality feature template itself during this matching process; A23: Historical reliability factor, based on the average accuracy of each modality pre-stored in the system in historical recognition tasks.

[0013] Preferably, the feature update module includes the following components during operation: S51: An update check is triggered when one of the following conditions is met: A. Periodic trigger: Executes according to a preset time period; B. Event Trigger: Executed immediately after the fusion decision module successfully completes an identity verification; S52: Obtain biometric template data to be updated, including high-quality biometric templates extracted by the feature processing module and existing baseline biometric templates in the registration template database; S53: Perform quality and timeliness analysis on the high-quality biometric templates extracted by the feature processing module to determine whether they need to be updated; S54: Perform update decision and conflict detection, calculate the similarity between the two sets of templates obtained in step S52. When the feature similarity is higher than the first preset threshold, proceed to the next step. When the feature similarity is lower than the first preset threshold but higher than the second preset threshold, trigger a conflict alarm and pause the update. When the feature similarity is lower than the second preset threshold, determine it as a potential identity conflict, refuse to update and generate a security anomaly log. S55: Perform template update, perform weighted average fusion of the old and new templates, generate a new fused benchmark template and store it in the database. The weights are allocated based on the quality and timeliness of the template.

[0014] Preferably, when performing quality and timeliness analysis on the high-quality biometric templates extracted by the feature processing module to determine whether an update is needed, the specific steps include: S61: Quality Analysis. Calculate the quality assessment score of the new template. Assessment factors include the template's feature clarity, completeness, and confidence level. When the quality assessment score is higher than the preset score threshold, proceed to the next step. S62: Timeliness analysis: Obtain the storage time of the old template and calculate its template age; if the template age exceeds the preset age threshold, it is determined that the old template has become outdated and is updated.

[0015] A high-precision AI roll call method based on multimodal biometric fusion includes the following steps: S71: Synchronously collects raw multimodal biometric data of the user to be identified through the data acquisition module; S72: The feature processing module processes the data of each modality and extracts the corresponding biometric templates; S73: Compare the extracted biometric template to be identified with the benchmark biometric template in the registered template database to generate a matching similarity score for each modality; S74: Through the fusion decision module, a dynamic weighted fusion algorithm is used to fuse the similarity scores of each modality and output the final identity recognition result; S75: Through business logic and application modules, complete the roll call operation based on the final identity recognition result and record the attendance result; S76: The feature update module updates the baseline biometric templates in the registration template database according to a preset strategy.

[0016] The beneficial effects of this invention are: 1. Compared with existing technologies that typically rely on a single biometric feature such as face for identity recognition, which has the disadvantage of a significant drop in recognition rate when the lighting is poor, the face is obscured, or there is malicious deception, this invention uses image and audio sensors to collect data simultaneously and integrates three biometric features, namely face, voiceprint, and gait, for joint decision-making. It has the advantage of multimodal information complementarity, can effectively cope with interference from complex environments and deception attacks, and significantly improves the accuracy and robustness of the roll call system. 2. Compared with existing technologies that mostly use fixed rules or simple weighted averages for information fusion, which are difficult to adapt to the dynamic changes in the quality of data of each modality under different scenarios and lead to rigid decision-making, this invention adopts a dynamic weighted fusion algorithm based on real-time data quality, matching confidence and historical reliability. It has the advantages of intelligently perceiving the context and adaptively adjusting the decision weights of each modality, so that the system can always prioritize the most reliable evidence, thereby ensuring the optimality of the recognition results and environmental adaptability. 3. Compared with existing technologies where biometric templates are usually fixed after registration and cannot adapt to the slow drift of biometrics caused by age and changes in body shape, resulting in the degradation of system performance over time, this invention introduces a feature update module with quality and timeliness analysis capabilities. It has the advantages of continuously evaluating newly collected templates and safely and gradually updating the baseline template, ensuring that the system can track the user's legitimate feature changes and maintain long-term recognition accuracy and stability. 4. Compared with the shortcomings of existing roll call systems, such as simple process, tight coupling between links, and difficulty in expanding new biometric modalities or optimizing specific processing units, this invention adopts a highly modular design, which clearly decouples data collection, feature extraction, fusion decision-making, template updating and business applications. It has the advantages of flexible system architecture, easy function expansion and maintenance upgrades, and provides convenience for future integration of more biometric modalities or updating of underlying algorithm models. 5. Compared with existing technologies that often focus on the identification process but lack a security maintenance mechanism for the benchmark database, which has the disadvantage of risking that templates may be incorrectly overwritten due to misidentification or malicious attacks, this invention designs a strict conflict detection and security verification process in the feature update module. It has the advantages of performing identity consistency verification before updating and pausing and alarming when an anomaly is detected, which effectively prevents the template database from being polluted and ensures the security of the system's core data and the reliability of the entire roll call result. Attached Figure Description

[0017] Figure 1 The diagram shown is a schematic representation of the high-precision AI attendance system based on multimodal biometric fusion according to the present invention. Figure 2 The diagram shown is a flowchart of the high-precision AI roll call method based on multimodal biometric fusion of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Please see Figure 1 - Figure 2 The present invention provides an embodiment of a high-precision AI roll call system based on multimodal biometric fusion, comprising: The data acquisition module is used to simultaneously acquire raw data of multimodal biometrics, including image sensors and audio sensors; The feature processing module is used to process the collected biometric data of each modality separately and extract the corresponding biometric templates. A registration template database is used to securely store baseline biometric templates for authorized users; The fusion decision module receives the matching similarity scores of each modality, executes a dynamic weighted fusion algorithm, and outputs the final identity recognition result. The feature update module is used to analyze the biometric templates extracted by the feature processing module and update the baseline biometric templates in the registration template database. The business logic and application module is used to manage attendance tasks, record attendance results, generate statistical reports, and provide a human-computer interaction interface.

[0020] In this embodiment, the present invention constructs a complete system architecture integrating data acquisition, feature processing, fusion decision-making, template updating, and business application, realizing the collaborative operation of multimodal biometrics in attendance scenarios. By simultaneously collecting multimodal data such as face, voiceprint, and gait, and performing comprehensive judgment through the fusion decision-making module, the system effectively overcomes the shortcomings of single biometrics being susceptible to environmental interference and spoofing attacks. This significantly improves the accuracy, robustness, and anti-counterfeiting capabilities of identity recognition, providing a system-level solution for achieving high-precision, contactless, and automated AI attendance taking.

[0021] Preferably, the feature processing module includes: A11: Face feature extraction unit, runs the first neural network model, and is used to extract the first feature vector from the face image through the first neural network model; A12: Voiceprint feature extraction unit, which runs a second neural network model to extract a second feature vector from the voiceprint audio through the second neural network model; A13: Gait feature extraction unit, which runs a third neural network model to extract a third feature vector from gait video sequences.

[0022] In this embodiment, the present invention deploys dedicated neural network models in the three feature processing units for face, voiceprint, and gait, enabling deep feature learning tailored to the characteristics of different modalities. This design fully leverages the advantages of deep learning in extracting highly discriminative and nonlinear features, allowing the extracted first, second, and third feature vectors to more accurately and fundamentally represent an individual's biological characteristics. This lays a solid feature foundation for subsequent high-precision fusion recognition and avoids the limitations of traditional manual feature representation.

[0023] Preferably, the facial feature extraction unit, when in operation, specifically includes: S11: Receive raw image data containing a human face acquired by an image sensor; S12: Preprocess the original image and locate the face region in the image using a face detection algorithm; S13: Input the preprocessed standard face image into the first neural network model for feature extraction. The first neural network model is a deep convolutional neural network, and the output of the last hidden layer of the first neural network model is defined as the first feature vector. S14: Normalize the first feature vector to obtain a first feature vector with fixed length and normalized modulus, which is used as the final face biometric template output.

[0024] In this embodiment, the present invention utilizes a standardized face preprocessing procedure (detection and alignment) combined with a deep convolutional neural network for feature extraction and normalization. This effectively eliminates interference from factors such as lighting, pose, and occlusion, extracting face feature templates with strong discriminative power and uniform scale. This process ensures that even face images acquired under non-ideal conditions can generate high-quality, comparable feature vectors, directly improving the reliability and stability of the face recognition process.

[0025] Preferably, the voiceprint feature extraction unit, when in operation, specifically includes: S21: Receive raw audio data acquired by the audio sensor; S22: Preprocess the raw audio data. First, pre-emphasize and window the audio data. Then, perform endpoint detection on each frame of audio data to remove silent segments and noise segments, and retain valid speech frames. S23: Extract acoustic feature parameters from valid speech frames to generate a speaker signature map, including Mel frequency cepstral coefficients; S24: Input the generated voiceprint feature map into the second neural network model, wherein the second neural network model is a neural network based on time-series modeling; S25: Deep feature learning and mapping of the voiceprint feature map is performed through the second neural network model, and the second feature vector is output from the fully connected layer specified by the model. S26: Normalize the second feature vector to obtain the voiceprint biometric template.

[0026] In this embodiment, the present invention performs a series of refined preprocessing steps on the original audio, including pre-emphasis, framing, and endpoint detection, and extracts MFCC features. Finally, it utilizes a temporal neural network for modeling, which effectively filters environmental noise and focuses on effective speech segments, thereby extracting high-quality voiceprint features that are low in dependence on text content but sensitive to speaker identity. This method enhances the adaptability of voiceprint recognition in far-field and noisy environments and improves the usability of the roll call system in purely audio or visually limited scenarios.

[0027] Preferably, the gait feature extraction unit includes the following components during operation: S31: Receive a gait video sequence containing pedestrian targets acquired by an image sensor; S32: Use object detection algorithms to process each frame of the video and locate pedestrian areas; S33: Based on the pedestrian region, background subtraction and contour extraction are performed to obtain the binary contour sequence of the pedestrian. The binary contour sequence is then aligned and normalized in the time dimension to construct standardized gait sequence data. S34: Input the preprocessed standardized gait sequence data into the third neural network model, wherein the third neural network model is a three-dimensional convolutional neural network; S35: The spatiotemporal features in the gait sequence are extracted and modeled hierarchically using a third neural network model to obtain the third feature vector; S36: Normalize the third feature vector to obtain the final gait biometric template.

[0028] In this embodiment, the present invention processes gait video sequences, extracts the binary contours of pedestrians, constructs spatiotemporal sequence data, and utilizes a three-dimensional convolutional neural network to capture the dynamic spatiotemporal patterns of gait. This method enables the system to recognize people based on their walking posture, possessing the advantages of being non-contact, long-distance, and difficult to masquerade. It effectively supplements the recognition capabilities in scenarios where faces are occluded or clear faces cannot be captured, enhancing the system's applicability in complex real-world environments.

[0029] As a preferred option, the fusion decision module, when in operation, specifically includes: S41: Compare the biometric template to be identified extracted by the feature processing module with the baseline biometric template of the corresponding user in the registration template database, and calculate the matching similarity score of each biometric modality. S42: Dynamically calculate the corresponding fusion weights for the matching similarity scores of each modality; S43: Using the calculated dynamic weights, the similarity scores of each match are weighted and fused to generate a fusion score; the weighted fusion calculation adopts a weighted summation method. S44: Compare the final fusion score with the preset recognition threshold: A. If the final fusion score is higher than the recognition threshold, the recognition is considered successful, and the corresponding user's identity identifier is output as the final identity recognition result; B. If the final fusion score is lower than the recognition threshold, it is determined as a recognition failure, and a recognition failure message is output.

[0030] In this embodiment, the present invention utilizes a strategy of first calculating independent matching scores for each modality and then performing dynamic weighted fusion, rather than a simple "yes or no" primary fusion, thus fully leveraging the complementary information of each modality. The weighted summation fusion algorithm is computationally efficient and highly interpretable. By comparing the final fusion score with a unified threshold, it achieves reliable identity determination based on joint decision-making from multiple sources, significantly reducing the probability of false identification and false rejection, thereby ensuring the fairness and accuracy of the roll call results.

[0031] Preferably, when dynamically calculating the fusion weights for the matching similarity scores of each modality, the calculation is based on the following factors: A21: Quality factor, based on the quality assessment score of the raw data of each modality collected this time. Modalities with higher data quality are assigned higher weights. A22: Confidence factor, based on the confidence of the matching quality generated by each modality feature template itself during this matching process; A23: Historical reliability factor, based on the average accuracy of each modality pre-stored in the system in historical recognition tasks.

[0032] In this embodiment, the present invention introduces quality factors, confidence factors, and historical reliability factors for dynamic weight allocation, enabling the fusion decision to possess context-awareness and self-optimization capabilities. This mechanism can adaptively adjust the contribution of each modality: reducing the weight of a modality when its data quality is poor or the confidence of the current match is low, and vice versa, while referring to historical performance to trust modalities with long-term reliability. This intelligent weight allocation strategy greatly improves the system's adaptability and robustness in complex and ever-changing real-world environments.

[0033] Preferably, the feature update module includes the following components during operation: S51: An update check is triggered when one of the following conditions is met: A. Periodic trigger: Executes according to a preset time period; B. Event Trigger: Executed immediately after the fusion decision module successfully completes an identity verification; S52: Obtain biometric template data to be updated, including high-quality biometric templates extracted by the feature processing module and existing baseline biometric templates in the registration template database; S53: Perform quality and timeliness analysis on the high-quality biometric templates extracted by the feature processing module to determine whether they need to be updated; S54: Perform update decision and conflict detection, calculate the similarity between the two sets of templates obtained in step S52. When the feature similarity is higher than the first preset threshold, proceed to the next step. When the feature similarity is lower than the first preset threshold but higher than the second preset threshold, trigger a conflict alarm and pause the update. When the feature similarity is lower than the second preset threshold, determine it as a potential identity conflict, refuse to update and generate a security anomaly log. S55: Perform template update, perform weighted average fusion of the old and new templates, generate a new fused benchmark template and store it in the database. The weights are allocated based on the quality and timeliness of the template.

[0034] In this embodiment, the present invention achieves automated, intelligent, and secure management of baseline biometric templates by setting up both periodic and event-triggered update mechanisms and introducing a strict conflict detection process. This solution can effectively track legitimate, slow changes in user biometrics over time (such as weight gain or loss, voice changes), prevent template aging through incremental updates, and improve long-term system performance. Simultaneously, the conflict detection mechanism effectively prevents template contamination caused by identification errors or malicious injection, ensuring the security and integrity of the database.

[0035] Preferably, when performing quality and timeliness analysis on the high-quality biometric templates extracted by the feature processing module to determine whether an update is needed, the specific steps include: S61: Quality Analysis. Calculate the quality assessment score of the new template. Assessment factors include the template's feature clarity, completeness, and confidence level. When the quality assessment score is higher than the preset score threshold, proceed to the next step. S62: Timeliness analysis: Obtain the storage time of the old template and calculate its template age; if the template age exceeds the preset age threshold, it is determined that the old template has become outdated and is updated.

[0036] In this embodiment, the present invention quantifies the update conditions into specific quality score thresholds and template age thresholds, making update decisions based on evidence and objectively reliable. Quality analysis ensures that only high-performance new templates are used for updates, avoiding the introduction of low-quality data that leads to performance degradation; timeliness analysis proactively identifies and replaces old templates that have become invalid over time. This dual-factor analysis mechanism guarantees the scientific nature and effectiveness of the template update process and maintains the long-term stability of the system's recognition accuracy.

[0037] A high-precision AI roll call method based on multimodal biometric fusion includes the following steps: S71: Synchronously collects raw multimodal biometric data of the user to be identified through the data acquisition module; S72: The feature processing module processes the data of each modality and extracts the corresponding biometric templates; S73: Compare the extracted biometric template to be identified with the benchmark biometric template in the registered template database to generate a matching similarity score for each modality; S74: Through the fusion decision module, a dynamic weighted fusion algorithm is used to fuse the similarity scores of each modality and output the final identity recognition result; S75: Through business logic and application modules, complete the roll call operation based on the final identity recognition result and record the attendance result; S76: The feature update module updates the baseline biometric templates in the registration template database according to a preset strategy.

[0038] In this embodiment, the present invention connects the functions of the various modules of the above system into a coherent method flow, fully illustrating the end-to-end automated roll call operation from multimodal data acquisition, feature extraction, fusion recognition to result recording and template updating. This method surpasses traditional manual roll call or single biometric roll call methods, significantly improving the efficiency, accuracy, anti-spoofing ability, and long-term adaptive performance of roll call work, providing an advanced technical path for identity verification needs in education, security, attendance, and other fields.

[0039] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A high-precision AI attendance system based on multimodal biometric fusion, characterized in that: include: The data acquisition module is used to simultaneously acquire raw data of multimodal biometrics, including image sensors and audio sensors; The feature processing module is used to process the collected biometric data of each modality separately and extract the corresponding biometric templates. A registration template database is used to securely store baseline biometric templates for authorized users; The fusion decision module receives the matching similarity scores of each modality, executes a dynamic weighted fusion algorithm, and outputs the final identity recognition result. The feature update module is used to analyze the biometric templates extracted by the feature processing module and update the baseline biometric templates in the registration template database. The business logic and application module is used to manage attendance tasks, record attendance results, generate statistical reports, and provide a human-computer interaction interface.

2. The high-precision AI attendance system based on multimodal biometric fusion according to claim 1, characterized in that: The feature processing module includes: A11: Face feature extraction unit, runs the first neural network model, and is used to extract the first feature vector from the face image through the first neural network model; A12: Voiceprint feature extraction unit, which runs a second neural network model to extract a second feature vector from the voiceprint audio through the second neural network model; A13: Gait feature extraction unit, which runs a third neural network model to extract a third feature vector from gait video sequences.

3. The high-precision AI attendance system based on multimodal biometric fusion according to claim 2, characterized in that: The facial feature extraction unit, when in operation, specifically includes: S11: Receive raw image data containing a human face acquired by an image sensor; S12: Preprocess the original image and locate the face region in the image using a face detection algorithm; S13: Input the preprocessed standard face image into the first neural network model for feature extraction. The first neural network model is a deep convolutional neural network, and the output of the last hidden layer of the first neural network model is defined as the first feature vector. S14: Normalize the first feature vector to obtain a first feature vector with fixed length and normalized modulus, which is used as the final face biometric template output.

4. The high-precision AI attendance system based on multimodal biometric fusion according to claim 3, characterized in that: The voiceprint feature extraction unit, when in operation, specifically includes: S21: Receive raw audio data acquired by the audio sensor; S22: Preprocess the raw audio data. First, pre-emphasize and window the audio data. Then, perform endpoint detection on each frame of audio data to remove silent segments and noise segments, and retain valid speech frames. S23: Extract acoustic feature parameters from valid speech frames to generate a speaker signature map, including Mel frequency cepstral coefficients; S24: Input the generated voiceprint feature map into the second neural network model, wherein the second neural network model is a neural network based on time-series modeling; S25: Deep feature learning and mapping of the voiceprint feature map is performed through the second neural network model, and the second feature vector is output from the fully connected layer specified by the model. S26: Normalize the second feature vector to obtain the voiceprint biometric template.

5. The high-precision AI attendance system based on multimodal biometric fusion according to claim 4, characterized in that: The gait feature extraction unit, when in operation, specifically includes: S31: Receive a gait video sequence containing pedestrian targets acquired by an image sensor; S32: Use object detection algorithms to process each frame of the video and locate pedestrian areas; S33: Based on the pedestrian region, background subtraction and contour extraction are performed to obtain the binary contour sequence of pedestrians. The binary contour sequence is then aligned and normalized in the time dimension to construct standardized gait sequence data. S34: Input the preprocessed standardized gait sequence data into the third neural network model, wherein the third neural network model is a three-dimensional convolutional neural network; S35: The spatiotemporal features in the gait sequence are extracted and modeled hierarchically using a third neural network model to obtain the third feature vector; S36: Normalize the third feature vector to obtain the final gait biometric template.

6. The high-precision AI attendance system based on multimodal biometric fusion according to claim 5, characterized in that: When the fusion decision-making module is working, it specifically includes: S41: Compare the biometric template to be identified extracted by the feature processing module with the baseline biometric template of the corresponding user in the registration template database, and calculate the matching similarity score of each biometric modality. S42: Dynamically calculate the corresponding fusion weights for the matching similarity scores of each modality; S43: Using the calculated dynamic weights, the similarity scores of each match are weighted and fused to generate a fusion score; the weighted fusion calculation adopts a weighted summation method. S44: Compare the final fusion score with the preset recognition threshold: A. If the final fusion score is higher than the recognition threshold, the recognition is considered successful, and the corresponding user's identity identifier is output as the final identity recognition result; B. If the final fusion score is lower than the recognition threshold, it is determined as a recognition failure, and a recognition failure message is output.

7. The high-precision AI attendance system based on multimodal biometric fusion according to claim 6, characterized in that: When dynamically calculating the fusion weights for the matching similarity scores of each modality, the calculation is based on the following factors: A21: Quality factor, based on the quality assessment score of the raw data of each modality collected this time. Modalities with higher data quality are assigned higher weights. A22: Confidence factor, based on the confidence of the matching quality generated by each modality feature template itself during this matching process; A23: Historical reliability factor, based on the average accuracy of each modality pre-stored in the system in historical recognition tasks.

8. The high-precision AI attendance system based on multimodal biometric fusion according to claim 7, characterized in that: The feature update module, when working, includes: S51: An update check is triggered when one of the following conditions is met: A. Periodic trigger: Executes according to a preset time period; B. Event Trigger: Executed immediately after the fusion decision module successfully completes an identity verification; S52: Obtain biometric template data to be updated, including high-quality biometric templates extracted by the feature processing module and existing baseline biometric templates in the registration template database; S53: Perform quality and timeliness analysis on the high-quality biometric templates extracted by the feature processing module to determine whether they need to be updated; S54: Perform update decision and conflict detection, calculate the similarity between the two sets of templates obtained in step S52, if the feature similarity is higher than the first preset threshold, proceed to the next step, if the feature similarity is lower than the first preset threshold but higher than the second preset threshold, trigger a conflict alarm and pause the update, if the feature similarity is lower than the second preset threshold, determine it as a potential identity conflict, refuse to update and generate a security anomaly log. S55: Perform template update, perform weighted average fusion of the old and new templates, generate a new fused benchmark template and store it in the database. The weights are allocated based on the quality and timeliness of the template.

9. The high-precision AI attendance system based on multimodal biometric fusion according to claim 8, characterized in that: When performing quality and timeliness analysis on the high-quality biometric templates extracted by the feature processing module to determine whether updates are needed, the specific steps include: S61: Quality Analysis. Calculate the quality assessment score of the new template. Assessment factors include the template's feature clarity, completeness, and confidence level. When the quality assessment score is higher than the preset score threshold, proceed to the next step. S62: Timeliness analysis: Obtain the storage time of the old template and calculate its template age; if the template age exceeds the preset age threshold, it is determined that the old template has become outdated and is updated.

10. A high-precision AI roll call method based on multimodal biometric fusion, characterized in that: Includes the following steps: S71: Synchronously collects raw multimodal biometric data of the user to be identified through the data acquisition module; S72: The feature processing module processes the data of each modality and extracts the corresponding biometric templates; S73: Compare the extracted biometric template to be identified with the benchmark biometric template in the registered template database to generate a matching similarity score for each modality; S74: Through the fusion decision module, a dynamic weighted fusion algorithm is used to fuse the similarity scores of each modality and output the final identity recognition result; S75: Through business logic and application modules, complete the roll call operation based on the final identity recognition result and record the attendance result; S76: The feature update module updates the baseline biometric templates in the registration template database according to a preset strategy.

Images