Self-updating multi-feature fusion recognition system

Abstract

The invention belongs to the technical field of safety monitoring, and particularly relates to a self-updating multi-feature fusion recognition system. According to the invention, through multi-modal data dynamic partition acquisition, a plurality of biological feature recognition technologies and random instruction recognition are subjected to weight distribution and then fused for different-dimension identity recognition of the whole system; the random security password has subjective randomness based on interdisciplinary introduction of cryptology. Each biological feature recognition module recognizes attention mechanisms associated with precision in real time to dynamically allocate feature weights, and through cross application of multi-dimensional recognition technologies, influences of external environments and accidental events are greatly reduced. The invention also provides a data post-processing technology adapted to identification results in different stages, and the result precision is further improved. And an identification log mechanism for identity matching failure is provided, and subsequent analysis, utilization and traceability of the whole system are carried out. And an effective technical support is provided for a safety monitoring application scene with high precision, confidentiality and leakage requirements.

Description

Technical Field

[0001] The present invention belongs to the technical field of security monitoring, and in particular relates to a self-updating multi-feature fusion recognition system. Background Art

[0002] In recent years, with the rapid development of computer vision, artificial intelligence, and big data technologies, feature recognition technology has made significant progress and has been widely applied in various fields. In the field of security monitoring and identification technology, biometric recognition technology has seen rapid development in recent years due to the uniqueness of individual biometrics, which has attracted great attention. Biometric recognition technology is a technology that uses the unique physiological or behavioral characteristics of an individual to authenticate an individual. Examples include facial recognition, ReID (person re-identification), and gait recognition.

[0003] Facial recognition technology uses facial features, such as the shape, size, and position of facial contours, eyes, nose, and mouth, to authenticate and identify individuals. While this technology offers advantages such as non-contact operation and rapid recognition speed, in practice, facial recognition is susceptible to factors such as distance, lighting, angle changes, and occlusion, resulting in reduced accuracy and slow recognition speed.

[0004] ReID technology is a cross-camera, cross-scene pedestrian identification technology. It analyzes pedestrian appearance characteristics (such as color, texture, and shape) to achieve continuous tracking and identification. ReID technology has broad application prospects in intelligent surveillance, smart security, and other fields. The technology has advantages such as high recognition speed and strong anti-interference ability. However, in practice, ReID recognition can be difficult due to factors such as distance and angle.

[0005] Gait recognition technology uses analysis of pedestrian gestures, gait patterns, and other characteristics to authenticate and identify pedestrians. This technology offers advantages such as non-contact, long-distance recognition, and resistance to camouflage, enabling identification without drawing the pedestrian's attention. Therefore, gait recognition offers unique advantages in security monitoring, smart access control, and other fields. However, in practical applications, gait recognition can vary due to factors such as health status and movement patterns, as well as occlusion and viewing angles, leading to inaccurate recognition results.

[0006] In summary, it can be seen that existing biometric recognition technology is affected by various factors such as distance, occlusion, lighting, angle, etc., which leads to recognition difficulties, reduced recognition accuracy, and even inaccurate recognition results. For security monitoring application scenarios with high confidentiality and leakage requirements, single biometric recognition is far from enough. Summary of the Invention

[0007] In response to the above-mentioned problems or shortcomings, and to solve the problems of recognition difficulties and inaccurate results caused by external factors in existing biometric recognition technologies, the present invention provides a self-updating multi-feature fusion recognition system, which adopts multi-mode dynamic fusion and customized updates of the prior recognition database corresponding to each recognition module to perform personnel identity recognition and anomaly detection, thereby solving the problems of low recognition accuracy and inaccurate results in existing biometric recognition technologies, and thus improving the comprehensive performance of the identity authentication system.

[0008] A self-updating multi-feature fusion recognition system includes a data acquisition module, a priori data module, a recognition module, an identity matching module and a control module.

[0009] The data acquisition module is composed of multiple feature acquisition cameras set up in the monitoring area to complete the multimodal data collection of the target person. Multimodal data refers to the data required by the recognition module to complete the recognition of its sub-modules; the monitoring area includes the personnel passage area and the identity recognition area.

[0010] The prior data module includes a cache database and a prior database. The prior database is constructed by storing various features of relevant personnel using feature acquisition cameras and is called upon during identity recognition. The cache database is constructed by storing various features of personnel daily, collected in real time by the feature acquisition cameras. The prior database's various features are regularly updated based on the cache data. The collection of feature data from areas where personnel pass can enrich the prior database's data to improve recognition accuracy and even be used for leak detection.

[0011] The identification module consists of at least two single biometric recognition modules and a random command recognition module. This reduces the limitations of single-technology recognition (two single biometric recognition modules) and single-type recognition (biometric recognition and random security password recognition), and increases the diversity of the entire identification system in both recognition methods and types. Each single biometric recognition module captures its own feature data and outputs recognition results; the random command recognition module captures the feature data of the random security password and outputs recognition results. Each submodule supports batch recognition.

[0012] The identity matching module receives the recognition results of each submodule in the recognition module for weight allocation and fusion of multimodal features. Multimodal features refer to the recognition results of various submodules output by the recognition module. Among them, the recognition results of each single biometric recognition module involved in the recognition module are weighted according to the level of their recognition accuracy, while the recognition results of the random instruction recognition module are weighted according to a constant value. The recognition results of each submodule adopt a scoreboard mode. After the recognition results of each submodule are weighted and accumulated, they are compared with the set judgment threshold to complete the final identity matching.

[0013] The control module is used to coordinate the workflow of each module and support dynamic parameter adjustment.

[0014] In this system, the characteristic data collection and application of personnel passage areas can be used to detect unnatural characteristic data of personnel in an alert state; random security passwords are relatively less affected by the external environment and have a certain degree of subjectivity; therefore, the selection of these two strategies can be used to a certain extent for leak detection.

[0015] Furthermore, the single biometric recognition module includes a face recognition module, a gait recognition module, a ReID module, a finger vein recognition module, a fingerprint recognition module, and an iris recognition module. Each single biometric recognition module collects target person data in real time through a feature acquisition camera, compares the collected data with a priori database, completes the identity recognition of the module, and outputs the recognition result to the identity matching module.

[0016] Furthermore, the number of the single biometric identification modules is at least three, so as to cope with the situation where a single biometric identification module suddenly fails.

[0017] Furthermore, the types of identification data of the single biometric identification modules are not completely the same, so as to reduce the influence of the external environment and improve the overall identification accuracy.

[0018] Furthermore, the weight allocation adopts the attention mechanism to dynamically allocate feature weights: weight allocation is weighted according to the real-time recognition accuracy of each single biometric recognition module (the accuracy of single biometric recognition will change due to the influence of the external environment), and the higher the recognition accuracy, the higher the weight ratio; after accumulating the weighted results of each single biometric recognition module and the weighted results of the random instruction recognition module, they are compared with the preset judgment threshold to output the final recognition result, so as to enhance the robustness of the system to complex environments.

[0019] Furthermore, the random security password in the a priori database is constructed by permuting and combining the stored body language instructions and the real-time date instructions.

[0020] The real-time date instruction refers to Monday to Sunday as a sub-element, or an additional month as a sub-element, or a combination of month and audio instruction sub-element.

[0021] Body language instructions include gesture instructions and head movement instructions: gesture instructions include daily gestures and arm gestures corresponding to the left and right palms respectively; daily gestures include finger number gestures, OK gestures, fist gestures, thumbs-up gestures, love gestures, spread-out gestures and their combination gestures; head movement instructions include nodding, shaking head left and right and their combination movements.

[0022] Furthermore, the update cycle of the a priori database is customized by the control module according to the differences of each sub-module: the face recognition module manually starts data update when the face of the person changes significantly due to external reasons; the gait recognition module manually starts data update according to the occasional health status of the person; the ReID module is updated in real time through the feature collection camera of the personnel passage area; the fingerprint recognition module manually starts data update when the fingerprint of the person is damaged due to external reasons; the random security password of the random instruction recognition module is updated by introducing a chaotic encryption algorithm to save computing resources.

[0023] Or, update in real time to improve recognition accuracy.

[0024] Furthermore, the update cycle of the a priori database is customized according to the differences of each submodule through the control module: The face recognition module manually initiates data updates when a person's face changes significantly due to external factors; the gait recognition module manually initiates data updates based on a person's occasional health condition; the ReID module updates in real time using feature acquisition cameras in areas where people pass through; the fingerprint recognition module manually initiates data updates when a person's fingerprint is damaged due to external factors; and the random security password of the random instruction recognition module is updated using a chaotic encryption algorithm. Or, real-time update: dynamically optimize the prior database based on the latest recognition results; if the recognition is successful, the system will incorporate the current feature data into the person's feature prior database, and update, fuse or replace historical feature information in a timely manner based on timeliness and weight distribution strategies to improve matching accuracy and system adaptability.

[0025] Furthermore, the prior data module is also provided with an identification log; when identity matching fails, the system will trigger an exception handling mechanism, mark and record relevant information in the identification log for subsequent analysis and tracing; the log content includes: timestamp, camera location, matching score, identity matching status and feature data summary; the log is used for system performance evaluation, anomaly detection, troubleshooting records in the event of a leak, and adjustment of the system's later optimization strategy; the optimization strategy uses manual intervention or automatic learning strategy to optimize the recognition effect.

[0026] The present invention utilizes dynamic partitioning of multimodal data to combine at least two single biometric recognition modules and random instruction recognition for different dimensions of identity recognition across the entire system. The recognition results from each single biometric recognition module are then weighted based on accuracy and then combined with random instruction recognition to achieve final identity matching. The introduction of a random security password, based on cryptography and subjectivity, can enhance security and even be used to identify internal information leaks. The system further provides an attention mechanism that dynamically assigns feature weights based on the real-time recognition accuracy of each single biometric recognition module. Furthermore, the system utilizes single biometric recognition modules that recognize different types of data, limiting the number of single biometric recognition modules and reducing the impact of external environments and incidental events. Data post-processing technology is also provided to adapt recognition results at different stages, further improving accuracy. Furthermore, a logging mechanism is provided for failed identity matches, enabling subsequent analysis, utilization, and traceability across the entire system.

[0027] In summary, the present invention provides a self-updating multi-feature fusion recognition system, which improves the accuracy and robustness of existing feature recognition systems and is suitable for security monitoring, especially for application scenarios involving security and leak detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a system framework diagram of an embodiment; Figure 2 is an identification flow chart of an embodiment; Figure 3 A self-updating data logic diagram of an embodiment; Figure 4 Schematic diagram of the personnel passage area effect of the embodiment. DETAILED DESCRIPTION

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0030] In this example, in a security surveillance scenario, the system uses feature acquisition cameras to collect facial images, gait videos, pedestrian images from multiple cameras, and feature data (images and passwords) of a target person. It then extracts features, performs weighted scoring, and integrates these features to ultimately complete identity recognition. The system supports real-time processing and dynamic updates, effectively addressing recognition needs in complex environments.

[0031] The system of this embodiment includes: a data acquisition module, a priori data module, a recognition module, an identity matching module and a control module. Among them, the single biometric recognition module of the recognition module includes a face recognition module, a gait recognition module and a ReID module.

[0032] The random security password of the random command recognition module is randomly generated based on the following sub-elements: standing finger (gesture): index finger on Monday, index finger + middle finger on Tuesday, index finger + middle finger + ring finger on Wednesday, index finger + middle finger + ring finger + thumb on Thursday, thumb on Friday, thumb + index finger on Saturday, thumb + index finger + middle finger on Sunday; plus a voice command (two words), one command for each shift.

[0033] The specific workflow is: Step 1: Multimodal data collection and personnel information binding; The initial data collection of personnel and the corresponding personal information of personnel are completed by feature collection cameras deployed in the identity recognition area. The corresponding areas allowed to enter are divided according to the entry permission level of the personnel, and then a priori database is constructed.

[0034] Step 2: The prior database is constructed; Based on the verified identity, when a person passes through the feature collection camera in the monitoring area, the system simultaneously calls the feature collection camera to collect multimodal feature information of the person, including facial features (from a bird's-eye view), clothing appearance, and gait characteristics, to build a complete individual identification data set. The system will be officially launched until the data information required for all three single biometric recognition modules covered by the recognition module is collected.

[0035] Step 3: Feature extraction and recognition of each single biometric feature, ad hoc feature recognition; When a person appears in the identity recognition area and remains there for more than a preset threshold (such as 3 seconds), the system automatically triggers the feature extraction module of the single biometric recognition module to start identifying the identity; and compares the real-time data collected based on the time spent in the identity recognition area with the personal feature information in the prior database, and then sends each single biometric recognition result to the identity matching module.

[0036] Step 4: Identity matching; The identity matching module assigns weights to the multimodal features and calculates the similarity score based on the recognition results sent by each sub-recognition module of the recognition module, and then integrates and compares it with the judgment threshold; if the score exceeds the set threshold, the identity match is determined to be successful.

[0037] For passage to certain sensitive areas with a higher security level, the passerby is required to show a random security password before the identity matching module grants access; if the passerby fails to show the random security password, he / she will be directly denied access and the password will be recorded in the log for later inspection.

[0038] This embodiment takes into account the fact that the three single biometric recognition modules are affected by the external environment (such as light), as well as the subjectivity introduced by the random security password (generated by a random combination of sub-elements provided by the management personnel, and with a certain a priori knowledge scope). On the one hand, it enhances the recognition accuracy and adaptability, and on the other hand, it has a strong effect in certain scenarios of identifying theft of secrets (such as commercial or other secrets).

[0039] In the prior art, although many identity authentication systems use relatively accurate recognition algorithms, the deep learning models are usually large and the computing resources required for the inference process are high, resulting in a slow system response time, which makes it difficult to meet the application scenarios with high concurrency and high real-time requirements. In order to solve this problem, the embodiments of the present invention use optimized miniaturized deep learning models, such as FaceNet for face recognition, OSNet_x0_75 for ReID recognition, and GaitBase for gait recognition. These models not only have high recognition accuracy, but also ensure low computational complexity, so that the system can achieve low-latency real-time reasoning while ensuring high accuracy. By combining the efficient data interaction mechanism of the FastAPI framework, the present invention ensures the efficiency of the identity authentication process, can respond to changes in different environments in real time, and meet the real-time requirements in video streaming scenarios.

[0040] In changing environments, many systems only work effectively at fixed angles or under ideal conditions, struggling to cope with dynamic movements of people, occlusions, or changes in camera position. For example, many facial recognition systems only recognize people from specific angles; if a person turns or the angle changes, recognition accuracy decreases. Furthermore, gait recognition technology requires high accuracy in understanding a person's dynamic behavior and motion, and its performance in complex environments is often affected by occlusion and blur.

[0041] This embodiment addresses this problem by employing YOLOv11 target detection and tracking technology, combined with a dedicated camera and a general-purpose camera. YOLOv11 not only accurately captures the dynamic movements of people in complex environments, but also uses the Segmentation (Segmentation) model to perform mask segmentation on people, providing clear and accurate input for gait recognition. This technology enhances the system's robustness, enabling it to maintain high recognition accuracy despite motion, occlusion, and environmental changes, significantly improving the system's adaptability.

[0042] Furthermore, to prevent misidentification, this embodiment incorporates a dynamic feature data update mechanism: when a person passes through an entrance, the system automatically identifies and collects their facial, ReID, and gait features. If the similarity with the data in the database falls below 50%, the system automatically updates the feature information, ensuring the database is always up-to-date. This avoids the tedious manual updating process and improves the timeliness and accuracy of the data.

[0043] 1) Multiple biometric fusion: This embodiment combines three biometric features: facial recognition, ReID, and gait recognition for identity verification. By integrating these three features, the system can provide high-precision identity verification in various application environments, significantly improving its adaptability to factors such as complex scenes, occlusion, and changing lighting. This effectively improves the accuracy and security of verification, avoiding the limitations of a single recognition technology.

[0044] 2) Use YOLOv11 for target detection and person tracking: This example uses YOLOv11 as the core technology for object detection and tracking, combined with YOLO's Segment model to segment person masks, providing high-quality input for gait recognition. This ensures accurate tracking of people under different camera conditions while being highly robust to occlusion and other issues, addressing the technical bottleneck of traditional cameras that prevent them from effectively tracking people.

[0045] 3) Selecting AI models that prioritize real-time performance and low latency: In the face recognition, ReID recognition, and gait recognition modules, all AI models use smaller versions (such as FaceNet, osnet_x0_75, and GaitBase), ensuring real-time video stream processing and reducing latency. This innovation makes the system feasible for efficient identity authentication and large-scale deployment scenarios, meeting the requirements of real-time verification.

[0046] At this level, the present invention also provides another adjustable prior database update strategy: The control module customizes the various submodules within the recognition module: The face recognition module manually initiates data updates when a person's face changes significantly due to external factors; the gait recognition module manually initiates data updates based on a person's occasional health condition; the ReID module updates in real time using feature-collecting cameras in areas where people pass through; the fingerprint recognition module manually initiates data updates when a person's fingerprint is damaged due to external factors; and the random instruction recognition module uses a chaotic encryption algorithm to update the random security password. This reduces processing delays caused by excessive data flow while also ensuring the timeliness of the prior database.

[0047] 4) Layered recognition strategy for cameras collecting features in the personnel passage area and identity recognition area in the monitoring area: The cache database is constructed by storing various daily feature data of personnel collected in real time by feature collection cameras, and regularly updates various feature data of personnel in the prior database based on the cache data; Dedicated cameras for collecting signatures in identity recognition areas and a larger number of cameras for collecting signatures in traffic areas work together to identify individuals and collect and update signature data, ensuring accurate identity authentication at the entrance and continued matching within the monitored area. This layered recognition strategy effectively addresses the issues of sporadic and timely continuous verification of identity verification.

[0048] 5) Intelligent database update and feature matching mechanism: The system features an intelligent feature data update mechanism. When the database lacks corresponding feature data for a person, it automatically updates the data from the cached database. When similarity is low, the system updates the database with newly collected features. This mechanism ensures that the system's prior database always maintains the latest feature information, improving data accuracy over the long term.

[0049] 6) Efficient matching and fusion based on inference server: The AI ​​model inference server matches facial, ReID, and gait features, using a fusion algorithm combining weighted and threshold settings to select the best match from multiple recognition results. This technology ensures that the comprehensive judgment of different features can produce the most accurate results in complex environments.

[0050] Furthermore, this embodiment provides an optimization strategy for post-processing recognition result data (both the recognition results of individual biometric recognition modules and the fused recognition results) to improve recognition accuracy and reduce false positives, thereby enhancing the robustness and real-time performance of the system. Post-processing includes time series analysis, trajectory tracking, contextual information integration, deduplication and merging, and confidence thresholding.

[0051] After the recognition results are generated, they are further processed to improve the accuracy and robustness of the final results or to meet specific application requirements.

[0052] 1). Time series analysis: using temporal continuity to improve recognition accuracy.

[0053] For example, if a person is identified as a different person in consecutive video frames, this error can be corrected through time series analysis. Method: Use Kalman Filter to smooth the recognition results.

[0054] 2). Trajectory tracking: Track a person’s movement trajectory to ensure the continuity and consistency of the recognition results.

[0055] Methods: The DeepSORT target tracking algorithm is used to track the identified individuals, and the recognition results are corrected during the tracking process.

[0056] 3) Contextual Information Integration: Leverage other information from the scene to assist in recognition. For example, if it's known that only certain people can enter certain areas, this contextual information can be used to correct the recognition results. Method: Post-process the recognition results by combining scene knowledge, rules, and prior information.

[0057] 4) Deduplication and merging: This addresses duplicate or misidentification scenarios. For example, if the same person is identified multiple times as different individuals, deduplication and merging can correct this. Method: Use a similarity metric to identify and merge duplicate recognition results.

[0058] 5) Confidence Threshold: Filter out recognition results with low confidence levels to reduce false positives. Method: Set a confidence threshold and only accept a recognition result if its confidence level is higher than the threshold.

[0059] It can be seen from the above embodiments that the present invention uses dynamic partitioning and collection of multimodal data, assigns weights to multiple biometric recognition technologies and random instruction recognition, and then integrates them for different dimensional identity recognition of the entire system; random security passwords are based on the interdisciplinary introduction of cryptography, have subjective randomness, and are particularly suitable for application scenarios involving security and leak detection. The attention mechanism dynamically assigns feature weights associated with the real-time recognition accuracy of each biometric recognition module, and the cross-application of multi-dimensional recognition technologies greatly reduces the impact of external environment and accidental events. It also provides a data post-processing technology that adapts the recognition results at different stages to further improve the accuracy of the results. It also provides an identification log mechanism for identity matching failures for subsequent analysis, utilization and traceability of the entire system. It provides effective technical support for security monitoring application scenarios with high precision, confidentiality and high leakage requirements.

Claims

1. A self-updating multi-feature fusion recognition system, characterized by: It includes data acquisition module, prior data module, recognition module, identity matching module and control module; The data acquisition module is composed of multiple feature acquisition cameras set up in the monitoring area to complete the multimodal data collection of the target person. Multimodal data refers to the data required by the recognition module to complete the recognition of its submodules; the monitoring area includes the personnel passage area and the identity recognition area; The priori data module includes a cache database and a priori database; The priori database is constructed by collecting and inputting various feature data of relevant personnel through feature collection cameras, and is called when performing identity recognition; The cache database is constructed by storing various daily feature data of personnel collected in real time by feature collection cameras, and regularly updates various feature data of personnel in the prior database based on the cache data; The recognition module is composed of at least two single biometric recognition modules and a random instruction recognition module; each single biometric recognition module completes its own feature data capture and recognition result output; the random instruction recognition module completes the feature data capture and recognition result output of the random security password; each submodule supports batch recognition; The identity matching module receives the recognition results of each submodule in the identification module for weight assignment and fusion of multimodal features. Multimodal features refer to the recognition results of various submodules output by the identification module. The recognition results of each single biometric recognition module involved in the identification module are weighted according to their recognition accuracy, while the recognition results of the random instruction recognition module are weighted according to a constant value. The recognition results of each submodule are weighted and accumulated using a scoreboard model. After the weighted and accumulated scores are fused, they are compared with a set judgment threshold to complete the final identity matching. The control module is used to coordinate the workflow of each module and support dynamic parameter adjustment.

2. The self-updating multi-feature fusion recognition system according to claim 1, characterized in that: The single biometric recognition module includes a face recognition module, a gait recognition module, a ReID module, a finger vein recognition module, a fingerprint recognition module and an iris recognition module; Each single biometric recognition module collects target person data in real time through a feature collection camera, compares the collected data with the prior database, completes the identity recognition of this module, and outputs the recognition results to the identity matching module.

3. The self-updating multi-feature fusion recognition system according to claim 1, characterized in that: There are at least three types of single biometric identification modules.

4. The self-updating multi-feature fusion recognition system according to claim 1, characterized in that: The identification data types of the single biometric identification modules are not completely the same.

5. The self-updating multi-feature fusion recognition system according to claim 1, characterized in that: The weight allocation uses the attention mechanism to dynamically allocate feature weights: The weighting is performed according to the real-time recognition accuracy of each single biometric recognition module. The higher the recognition accuracy, the higher the weight ratio. The weighted results of each single biometric recognition module and the weighted results of the random instruction recognition module are accumulated and compared with the preset judgment threshold to output the final recognition result.

6. The self-updating multi-feature fusion recognition system according to claim 1, characterized in that: The random security password in the a priori database is constructed by permuting and combining the stored body language instructions and the real-time date instructions; Real-time date instructions refer to Monday to Sunday as sub-elements, or additionally add the month as a sub-element, or a combination of the month and audio instructions; Body language instructions include hand gesture instructions and head movement instructions: Gesture commands include daily gestures and arm gestures corresponding to the left and right palms respectively; daily gestures include finger number gestures, OK gestures, fist gestures, thumbs-up gestures, heart gestures, spread-out gestures and their combination gestures; Head movement commands include nodding, shaking head left and right, and their combination.

7. The self-updating multi-feature fusion recognition system according to claim 1, characterized in that: Update of the a priori database: The control module is customized according to the differences of each sub-module: the face recognition module manually starts data update when the person's face changes significantly due to external reasons; the gait recognition module manually starts data update based on the person's occasional health status; the ReID module is updated in real time through the feature collection camera of the personnel passage area; The fingerprint recognition module manually starts data update when the fingerprint is damaged due to external reasons; the random security password of the random instruction recognition module is updated by introducing a chaotic encryption algorithm; Or, real-time update: dynamically optimize the prior database based on the latest recognition results; if the recognition is successful, the system will incorporate the current feature data into the person's feature prior database, and update, integrate or replace historical feature information in a timely manner based on timeliness and weight distribution strategies.

8. The self-updating multi-feature fusion recognition system according to claim 1, characterized in that: The prior data module is also provided with an identification log; When identity matching fails, the system will trigger the exception handling mechanism, mark and record relevant information in the identification log for subsequent analysis and tracing; Log content includes: timestamp, camera location, match score, identity match status, and feature data summary; This log is used for system performance evaluation, anomaly detection, troubleshooting records during leaks, and adjustment of system optimization strategies in the future. The optimization strategy uses manual intervention or automatic learning strategies to optimize recognition effects.

9. The self-updating multi-feature fusion recognition system according to claim 1, characterized in that: The recognition results of the single biometric recognition module and the fused recognition results also adopt a post-processing optimization strategy, and the post-processing includes time series analysis, trajectory tracking, context information integration, deduplication and merging, and confidence thresholding; 1) Time series analysis: using temporal continuity to improve recognition accuracy; Method: Use Kalman filter to smooth the recognition results; 2) Trajectory tracking: Tracking the movement trajectory of people to ensure the continuity and consistency of recognition results; Methods: The DeepSORT target tracking algorithm is used to track the identified individuals and the recognition results are corrected during the tracking process. 3) Contextual information integration: using other information in the scene to assist recognition; Method: Combining scene knowledge, rules and prior information, the recognition results are post-processed; 4) Deduplication and merging: handling duplicate or misidentification situations; Methods: Similarity metrics are used to identify and merge duplicate recognition results; 5) Confidence threshold: filter out recognition results with too low confidence; Method: Set a confidence threshold and accept the result only when the confidence of the recognition result is higher than the threshold.

Images