Intelligent campus access control system based on artificial intelligence

Abstract

The invention discloses an intelligent campus access control system based on artificial intelligence, and the system comprises an intelligent sensing layer which is used for deploying an AI camera with a skeleton key point extraction function, an integrated RFID / NFC card reader, a two-dimensional code scanning module and an environment sensor, and extracting human skeleton key points in real time by adopting a posture estimation algorithm, multi-mode identity information and environment metadata are synchronously collected; the behavior understanding layer is used for constructing a space-time diagram model based on a space-time diagram neural network engine, taking skeleton key points as nodes, taking joint connection as space edges, connecting continuous frame space diagrams in series according to a time sequence to form a space-time diagram, learning a space relationship and a time evolution rule among the nodes by adopting a graph convolutional network, and completing behavior recognition, trailing detection and abnormal behavior distinguishing; the graph analysis layer is used for constructing a campus behavior knowledge graph based on the metadata tag library, and performing sub-graph mode mining by applying a graph neural network and a community discovery algorithm; and the decision and execution layer is used for configuring a grading early warning strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart campus, and particularly relates to a smart campus access control system based on artificial intelligence. BACKGROUND

[0002] The traditional campus access control system has the problems of low personnel passing efficiency, frequent security management vulnerabilities, lagging data statistics, and high management cost, etc. With the rapid development and deep application of artificial intelligence technology, it provides technical support for building an efficient, safe and intelligent campus access control system. The purpose is to realize the automatic and accurate identification of campus personnel access, real-time dynamic monitoring, abnormal behavior warning and multi-dimensional data visualization analysis through the integration of advanced technologies such as face recognition, behavior analysis, intelligent warning and data fusion, so as to improve the campus passing efficiency, reduce the management cost, strengthen the security protection ability, and provide scientific decision-making basis for campus managers. It not only reflects the overall upgrade of the campus security protection system, effectively preventing the intrusion of outsiders and the abnormal departure of students through high-precision identity verification and intelligent behavior monitoring, but also promotes the intelligent transformation of the campus management mode. Through data-driven fine management and service optimization, it realizes the change of management thinking from "passive response" to "active prevention", and improves the campus life experience and satisfaction of teachers and students, and finally builds a "safe, efficient, intelligent and humanized" modern smart campus ecology, and provides strong technical support and practical example for the modernization of campus governance in the era of educational information 2.0.

[0003] In the prior art, existing tail detection is mostly based on fixed thresholds such as too close distance between two people and time difference. However, in real scenes, normal behaviors such as dense passing of students during peak hours and walking side by side with friends are easily misjudged as tailing. The real malicious tailing may deliberately maintain a distance and use visual blind area, resulting in missed judgment. In addition, if only based on time threshold, students who are waiting at the door will be misjudged, therefore, the present application provides a smart campus access control system based on artificial intelligence. SUMMARY

[0004] The present application aims to solve the problems existing in the prior art and provides a smart campus access control system based on artificial intelligence.

[0005] In order to achieve the above purpose, the present application adopts the following technical scheme: A smart campus access control system based on artificial intelligence, comprising: Intelligent perception layer: Deploy AI cameras with bone key point extraction, integrate RFID / NFC card readers, two-dimensional code scanning modules and environmental sensors, use pose estimation algorithm to extract human bone key points in real time, and synchronously collect multi-modal identity information and environmental metadata. The original data collected by the edge computing node is compressed and desensitized in real time, and then transmitted to the behavior understanding layer through the campus private network encryption; Behavior understanding layer: Based on the spatio-temporal graph neural network engine, a spatio-temporal graph model is constructed, with bone key points as nodes and joint connections as spatial edges. The continuous frame spatial graph is connected in time sequence to form a spatio-temporal graph. The graph convolution network is used to learn the spatial relationship and time evolution law between nodes to complete behavior recognition, trailing detection and abnormal behavior distinction. The output behavior analysis results and abnormal event alarm signals are transmitted to the graph analysis layer together with the structured behavior data; Graph analysis layer: Based on the metadata tag library, a campus behavior knowledge graph is constructed. The Neo4j graph database is used to store entities, and the graph neural network GNN and community discovery algorithm are used for subgraph pattern mining to realize cross-regional collaborative early warning, dynamic risk scoring and visual heat map display. Risk score, early warning event and correlation analysis report are generated; Decision and execution layer: Configure hierarchical early warning strategy, use blockchain technology to store key operations to ensure traceability.

[0006] The above technical scheme further comprises: Further, the intelligent perception layer synchronously collects multi-modal data: Video stream acquisition: AI camera captures RGB video stream, preliminary noise reduction processing is performed through convolutional neural network CNN, and bone key points are extracted through pose estimation algorithm. Identity verification is performed through double-factor verification of bone key points and facial features in cooperation with face recognition module; Identity information acquisition: RFID / NFC card reader: Reads campus card UID, generates desensitization identifier through hash function, and performs double authentication of card swiping and bone key point verification in cooperation with access control permission system; Face recognition module: Extracts 512-dimensional feature vector using ArcFace algorithm, and performs live body detection through cosine similarity. When the cosine similarity is greater than 0.7, it is determined as a live body; Two-dimensional code scanning: Uses Zxing library to parse two-dimensional code content and verifies CRC code to ensure data integrity; Environmental sensor data: The original data collected by the temperature and humidity sensor is mapped to the [0, 1] interval through linear normalization. Temperature and humidity data are used to evaluate human comfort, and luminosity data are used to optimize AI camera exposure parameters.

[0007] Further, the intelligent perception layer adopts an HRNet network for 2D / 3D skeleton key point detection, and outputs a heat map of key points, each heat map channel corresponding to a joint, and the peak position of the heat map being decoded into a sub-pixel level coordinate through a soft-argmax operation, and being represented as wherein, is a heat map response value of the kth joint at a position (i, j), is a final coordinate.

[0008] Further, the behavior understanding layer realizes space-time graph modeling and dynamic behavior analysis based on a space-time graph neural network engine, learns the spatial correlation and time evolution law of the skeleton key points through a graph convolution network, and the specific steps are as follows: Node definition: the skeleton key points extracted by the intelligent perception layer are taken as nodes, and each node feature vector contains three-dimensional coordinates, velocity, and acceleration third derivative, and is represented as , wherein, is a sub-pixel level coordinate of the kth joint at time t, is a velocity of the kth joint at time t, is an acceleration of the kth joint at time t; Space edge definition: taking the natural joint connection relationship of the human body as a space edge, a spatial graph is constructed wherein, is a node set, is a spatial edge set; Space-time graph extension: the spatial graphs of continuous T frames are concatenated in time sequence to form a space-time graph wherein , , is a time edge, connecting the corresponding relationship of the same node at different times; Spatial graph convolution: a graph convolution network GCN is adopted to learn the spatial relationship between nodes, and is represented as wherein, is a spatial graph convolution output, is a neighbor node set of node i, is a learnable weight matrix, is an adjacency matrix, representing the connection strength of node i and j, is an activation function, represents a feature vector of node j; Time convolution: a one-dimensional convolution kernel is adopted to slide along the time dimension to capture the continuity of the action, and finally outputs wherein, is a time convolution kernel.

[0009] Further, the The input full connection layer outputs a behavior class probability distribution through a softmax function, denoted as wherein, is a set of behavior classes, is a confidence score of class c, is other classes, is class c and is a weight vector of class c, is class c and is a bias term of class c, the model parameters are optimized by a cross-entropy loss function, and the probability distribution output by the full connection layer is used to determine the specific class to which the current action belongs, if the confidence score of a certain class is the highest and exceeds the threshold value of the corresponding class, an abnormal behavior alarm is triggered; The behavior type label output by the full connection layer is combined with the spatiotemporal synchrony index to determine the tailing behavior, the behavior of deliberately imitating the previous card swiping action is identified through spatiotemporal synchrony analysis, and the spatiotemporal synchrony index is defined as wherein, and are the skeletal key point position vectors of the follower and the followed person at time t, is the phase difference between the two actions, when and the action type matches, it is determined as a tailing behavior;

[0010] Further, the specific steps of constructing the campus behavior knowledge graph based on the metadata label library and performing subgraph pattern mining by the graph analysis layer are as follows: Metadata label generation: obtain structured data from the intelligent perception layer and the behavior understanding layer, and generate labels; Knowledge graph storage: store three types of entities and four types of relationships in a Neo4j graph database: Entities: personnel, area, and equipment; Relationships: spatiotemporal intersection, same row record, equipment association, and behavior association; Graph convolution network GNN feature learning: perform feature aggregation on the node embedding in the knowledge graph, denoted as wherein, is the embedding of node v at the l-th layer, is a set of neighbor nodes, is the node degree, is a ReLU activation function, and the high-order feature representation of the node is learned through a multi-layer graph convolution network GNN; Community discovery algorithm: identify abnormal subgraphs through modularity optimization, and the modularity is denoted as ,in, Let i be the connection weight between nodes i and j. Let m be the node degree and m be the total weight. The indicator function is used to iteratively optimize the Q-value.

[0011] Furthermore, the graph analysis layer integrates three dimensions: individual historical behavior, network characteristics, and regional risk coefficients. It uses a weighted linear model to complete dynamic risk scoring, maps the regional risk scores to the campus map, and uses color gradients to display the risk distribution.

[0012] Decision-making and execution level: Configure tiered early warning strategies and use blockchain technology to store evidence of key operations to ensure traceability.

[0013] Furthermore, the decision-making and execution layer defines a three-level early warning threshold based on the dynamic risk score of 0-100 points output by the weighted linear model, and uses an adaptive threshold algorithm to dynamically adjust the early warning threshold based on historical event data. The initial three-level early warning thresholds are: A dynamic risk score of ≥80 indicates high risk: an alarm will be automatically triggered and a notification will be sent to the security personnel's mobile device, while video tracking and on-site handling will be initiated simultaneously. A dynamic risk score of 60 or less and less than 80 indicates medium risk: logs are recorded and analysis reports are generated for subsequent review by security personnel; A dynamic risk score of <60 indicates low risk: the monitoring system will continuously track the risk without triggering an active response.

[0014] The present invention has the following beneficial effects: In this invention, key points of the human skeleton are extracted to construct a spatiotemporal graph model, directly learning the essential physical structure of human movement. Compared to traditional models based on fixed rules or shallow machine learning, ST-GCN can accurately capture the spatial relationships and temporal evolution of complex movements, exhibiting strong robustness to changes in perspective and clothing, significantly reducing the misjudgment of normal running as abnormal or the missed detection of concealed tailgating. Through spatiotemporal synchronization analysis, the system can identify whether a subsequent person is deliberately imitating the card-swiping action of the preceding person, accurately distinguishing between tailgating and normal walking, reducing the false alarm rate. Attached Figure Description

[0015] Figure 1 This is a system block diagram of an artificial intelligence-based smart campus access control system proposed in this invention. Detailed Implementation

[0016] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0017] Please refer to Figure 1 As shown in the drawings, the present application is a kind of intelligent campus access control system based on artificial intelligence, comprising: Intelligent sensing layer: deploy AI camera with skeleton key point extraction (such as Huawei HoloSens series), integrate RFID / NFC card reader, two-dimensional code scanning module and environmental sensor (temperature and humidity, illumination monitoring module), use pose estimation algorithm to extract 17-25 skeleton key points (head, shoulder, elbow, wrist, etc.) of human body in real time, and synchronously collect multi-modal identity information such as campus card, face and two-dimensional code and environmental metadata (such as weather, time period, passenger flow), the original data collected is compressed and desensitized in real time by edge computing node, and then transmitted to behavior understanding layer through campus private network encryption; Behavior understanding layer: construct spatio-temporal graph model based on space-time graph neural network (ST-GCN) engine, take skeleton key points as nodes and joint connection as space edge, form spatio-temporal graph by connecting continuous frame space graph in time sequence, learn the spatial relationship (such as hand-foot coordinated action) and time evolution law (such as action continuity) between nodes by using graph convolution network, complete behavior recognition (running, squatting, falling, fighting, etc. Action recognition accuracy ≥ 95%), tail detection (identify the behavior of deliberately imitating the card swiping action of the previous person by analyzing the time-space synchronization, false positive rate ≤ 0.5%) and abnormal behavior distinction (such as "stand and wait for people" and "suspicious wandering" heat map visualization difference analysis), the output behavior analysis result (behavior type, confidence, timestamp) and abnormal event alarm signal and structured behavior data (such as passing time, action type, area location) are transmitted to graph analysis layer together; Graph analysis layer: construct campus behavior knowledge graph based on metadata label library (containing weather, time period, passenger flow, same trip relationship, etc. 20+dimensional label), store entities (personnel, area, equipment) and relationships (spatio-temporal intersection, same trip record) in Neo4j graph database, apply graph neural network GNN and community discovery algorithm (such as Louvain algorithm) for subgraph pattern mining, realize cross-region collaborative early warning (such as identifying the "reconnaissance circle" mode formed by multiple strangers in sensitive area, or the "illegal card borrowing" behavior of the card of a graduated student frequently linked with a student in school), dynamic risk score (real-time calculation of 0-100 risk score by comprehensively considering individual historical behavior, correlation network characteristics and regional risk coefficient) and visual heat map display, and generate risk score, early warning event and correlation analysis report; Decision and execution layer: configure hierarchical early warning strategy (such as high-risk events automatically trigger alarm and push to security personnel mobile terminal, low-risk events record log for subsequent analysis), use blockchain technology to store evidence of key operations (permission change, abnormal alarm) to ensure traceability.

[0018] In one embodiment, the intelligent perception layer synchronously collects multi-modal data: Video stream acquisition: AI camera (such as Huawei HoloSens) captures RGB video stream at a rate of 30 frames / second, preliminary noise reduction processing is performed through convolutional neural network CNN, and skeletal key points are extracted through pose estimation algorithm, identity verification is performed through double-factor verification of skeletal key points and facial features in cooperation with face recognition module; Identity information acquisition: RFID / NFC card reader: reads campus card UID, generates desensitization identifier through hash function, and performs double authentication of card swiping and skeletal key point verification through access control permission system linkage to prevent tailing or card theft; Face recognition module: extracts 512-dimensional feature vector using ArcFace algorithm, and performs live body detection through cosine similarity, and determines as live body when cosine similarity > 0.7; QR code scanning: uses Zxing library to analyze QR code content, and checks CRC code to ensure data integrity; Environmental sensor data: temperature and humidity sensor collects raw data, which is mapped to [0, 1] interval through linear normalization, temperature and humidity data are used to evaluate human comfort, and illuminance data are used to optimize AI camera exposure parameters.

[0019] In one embodiment, the intelligent perception layer uses HRNet network to perform 2D / 3D skeletal key point detection, outputs a heat map of key points, and each heat map channel corresponds to a joint node. The peak position of the heat map is decoded to a sub-pixel level coordinate through a soft-argmax operation, and is represented as wherein, is the heat map response value of the kth joint node at position (i, j), is the final coordinate.

[0020] In one embodiment, the behavior understanding layer realizes spatio-temporal graph modeling and dynamic behavior analysis based on a spatio-temporal graph neural network engine, learns the spatial correlation and temporal evolution law of skeletal key points through a graph convolution network, and the specific steps are as follows: Node definition: the skeletal key points extracted by the intelligent perception layer are nodes, and each node feature vector includes three-dimensional coordinates, velocity, and acceleration third derivative, represented as , wherein, Let be the sub-pixel coordinates of the k-th joint at time t. Let be the velocity of the k-th joint at time t. Let be the acceleration of the k-th joint at time t; Spatial boundary definition: Using the natural joint connections of the human body (such as shoulder-elbow, hip-knee) as spatial boundaries, construct a spatial diagram. ,in, For a set of nodes, For the set of spatial edges; Spatiotemporal graph extension: Concatenate the spatial graphs of consecutive T frames in time series to form a spatiotemporal graph. ,in , , A time edge connects the corresponding relationships of the same node at different times; Spatial Graph Convolution: A Graph Convolutional Network (GCN) is used to learn the spatial relationships between nodes, represented as... ,in, For spatial graph convolution output, Let i be the set of neighboring nodes. For learnable weight matrix, Let be the adjacency matrix, representing the connection strength between nodes i and j. For activation function, This represents the feature vector of node j; Temporal convolution: A one-dimensional convolutional kernel slides along the time dimension to capture the continuity of actions, ultimately outputting... ,in, This is the temporal convolution kernel.

[0021] In one embodiment, The input is a fully connected layer, and the output is a probability distribution of behavior categories through a softmax function, represented as follows: ,in, This is a set of behavior categories (such as running, squatting, falling down, fighting, etc.). For the confidence level of category c, For other categories, and Category C and The weight vector, and Category C and The bias term optimizes the model parameters through the cross-entropy loss function (optimized by the backpropagation algorithm). Based on the probability distribution output by the fully connected layer, it determines the specific category of the current action, such as "running", "squatting", or "falling". If the confidence score of a certain category is the highest and exceeds the threshold of the corresponding category, an abnormal behavior alarm is triggered. The behavior type label output by the fully connected layer is combined with the spatiotemporal synchrony index to determine tailing behavior, the behavior of deliberately imitating the card swiping action of the previous person is identified through spatiotemporal synchrony analysis, and the spatiotemporal synchrony index is defined as wherein, and are the skeletal key point position vectors of the follower and the followed person at time t, is the phase difference between the two actions, and when the phase difference is greater than a threshold value and the action types match, it is determined that the behavior is tailing behavior; A behavior feature heat map is generated through gradient weighted class activation mapping (Grad-CAM), and abnormal behavior is distinguished by comparing the differences in the heat maps. For example, the heat map of “standing and waiting for people” shows a relaxed posture (such as drooping shoulders and focused vision), while the heat map of “suspicious loitering” shows a tense posture (such as raised shoulders and scanning vision).

[0022] In one embodiment, the specific steps of constructing a campus behavior knowledge graph and performing subgraph pattern mining based on the metadata label library by the graph analysis layer are as follows: Metadata label generation: structured data is obtained from the intelligent perception layer and the behavior understanding layer, and labels are generated, including but not limited to: time period label: mapping the passage time to 8 time periods; people flow label: counting the number of people passing through per unit time through regional cameras, and using Z-score standardization, when the standardized data is > 2, it is marked as “high flow area”; same trip relationship label: based on frequent item set mining (such as Apriori algorithm), if two people travel together for ≥ 3 times within 7 days, a “frequent same trip” relationship is established; Knowledge graph storage: using Neo4j graph database to store three types of entities and four types of relationships: Entities: personnel (students / faculty / visitors), areas (teaching buildings / laboratories / dormitories), equipment (access control / cameras); Relationships: spatiotemporal intersection (such as “personnel staying in area A”), same trip record (such as “two people traveling together”), equipment association (such as “camera C monitoring area B”), and behavior association (such as “personnel X having a fall behavior in area Y”); Graph convolutional network GNN feature learning: feature aggregation is performed on the node embedding in the knowledge graph, represented as wherein, is the embedding of node v at the lth layer, is the neighbor node set, is the node degree, is the ReLU activation function, and high-order feature representation of the node is learned through a multi-layer graph convolutional network GNN.

[0023] Community discovery algorithm: abnormal subgraphs are identified through modularity optimization, and the modularity is represented as wherein, is the connection weight of nodes i and j, is the node degree, and m is the total weight, is the indicator function (1 for the same community, otherwise 0), by iteratively optimizing the Q value, the "scout circle" pattern formed by multiple strangers in the sensitive area or the "illegal card borrowing" behavior of the graduated student card frequently linked with the students in school is identified.

[0024] In one embodiment, the graph analysis layer integrates three dimensions of individual historical behavior (number of individual historical abnormal behavior times), associated network features (number of high-risk nodes in the associated network), and regional risk coefficient (frequency of historical events in the region), adopts a weighted linear model to complete dynamic risk scoring, maps the regional risk score to the campus floor plan, and adopts color gradient (red-yellow-green) to display the risk distribution.

[0025] The decision and execution layer: configure a hierarchical early warning strategy (such as automatically triggering an alarm and pushing it to the mobile terminal of security personnel for high-risk events, and recording logs for subsequent analysis for low-risk events), and use blockchain technology to store evidence of key operations (permission changes, abnormal alarms) to ensure traceability.

[0026] In one embodiment, the decision and execution layer defines three early warning thresholds according to the 0-100 dynamic risk score value output by the weighted linear model, and uses an adaptive threshold algorithm to dynamically adjust the early warning threshold according to historical event data. The initial three early warning thresholds are: Dynamic risk score value ≥ 80 is high risk: automatically trigger an alarm and push it to the mobile terminal of security personnel (such as mobile phone APP, intercom), and at the same time start video tracking and on-site disposal; 60 ≤ dynamic risk score value < 80 is medium risk: record logs and generate analysis reports for subsequent verification by security personnel (such as calling monitoring video, checking personnel identity); Dynamic risk score value < 60 is low risk: only monitor the system to track continuously, without triggering active response.

[0027] Although embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and alterations can be made thereto without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart campus access control system based on artificial intelligence, characterized in that, include: Intelligent perception layer: Deploy AI cameras with skeletal key point extraction capabilities, integrate RFID / NFC card readers, QR code scanning modules and environmental sensors, use pose estimation algorithms to extract human skeletal key points in real time, and simultaneously collect multimodal identity information and environmental metadata. The collected raw data is compressed and desensitized in real time by edge computing nodes, and then transmitted to the behavior understanding layer through the campus private network in an encrypted manner. Behavior Understanding Layer: Based on the spatiotemporal graph neural network engine, a spatiotemporal graph model is constructed. Skeletal key points are used as nodes and joint connections are used as spatial edges. The continuous frame spatial graphs are concatenated in time series to form a spatiotemporal graph. The graph convolutional network is used to learn the spatial relationships and temporal evolution rules between nodes to complete behavior recognition, tail detection and abnormal behavior differentiation. The output behavior analysis results and abnormal event alarm signals are transmitted together with the structured behavior data to the graph analysis layer. Graph Analysis Layer: Based on the metadata tag library, a campus behavior knowledge graph is constructed. The Neo4j graph database is used to store entities. The graph neural network (GNN) and community detection algorithm are applied to mine subgraph patterns, enabling cross-regional collaborative early warning, dynamic risk scoring, and visualization of heat maps. Risk scores, early warning events, and correlation analysis reports are generated. Decision-making and execution level: Configure tiered early warning strategies and use blockchain technology to store evidence of key operations to ensure traceability.

2. The intelligent campus access control system based on artificial intelligence according to claim 1, characterized in that, The intelligent sensing layer synchronously collects multimodal data: Video stream acquisition: The AI ​​camera captures RGB video streams, performs preliminary noise reduction through a convolutional neural network (CNN), extracts skeletal key points through a pose estimation algorithm, and collaborates with the face recognition module to perform identity verification through two-factor authentication of skeletal key points and facial features. Identity information collection: RFID / NFC card reader: Reads the campus card UID, generates a desensitized identifier through a hash function, and links with the access control system to perform dual authentication of card swiping and skeletal key point verification; Face recognition module: The ArcFace algorithm is used to extract a 512-dimensional feature vector, and liveness detection is performed by using cosine similarity. When the cosine similarity is > 0.7, the face is considered to be live. QR code scanning: The Zxing library is used to parse the QR code content, and the CRC code is verified to ensure data integrity; Environmental sensor data: Temperature and humidity sensors collect raw data, which is mapped to the [0,1] interval through linear normalization. Temperature and humidity data are used to assess human comfort. Illuminance data is logarithmically transformed to compress the dynamic range. Illuminance data is used to optimize the exposure parameters of the AI ​​camera.

3. The intelligent campus access control system based on artificial intelligence according to claim 2, characterized in that, The intelligent perception layer uses an HRNet network for 2D / 3D skeletal keypoint detection, outputting heatmaps of the keypoints. Each heatmap channel corresponds to one keypoint. The peak positions of the heatmaps are decoded into sub-pixel coordinates through a soft-argmax operation, represented as follows: ,in, Let be the heatmap response value of the k-th joint at position (i,j). These are the final coordinates.

4. The intelligent campus access control system based on artificial intelligence according to claim 1, characterized in that, The behavior understanding layer is based on a spatiotemporal graph neural network engine to realize spatiotemporal graph modeling and dynamic behavior analysis. It learns the spatial correlation and temporal evolution of skeletal key points through graph convolutional networks. Specific steps: Node definition: Nodes are defined using the skeletal keypoints extracted by the intelligent perception layer. Each node's feature vector contains three-dimensional coordinates, velocity, and the third derivative of acceleration, expressed as... , ,in, Let be the sub-pixel coordinates of the k-th joint at time t. Let be the velocity of the k-th joint at time t. Let be the acceleration of the k-th joint at time t; Spatial boundary definition: Using the natural joint connections of the human body as spatial boundaries, construct a spatial diagram. ,in, For a set of nodes, For the set of spatial edges; Spatiotemporal graph extension: Concatenate the spatial graphs of consecutive T frames in time series to form a spatiotemporal graph. ,in , , A time edge connects the corresponding relationships of the same node at different times; Spatial Graph Convolution: A Graph Convolutional Network (GCN) is used to learn the spatial relationships between nodes, represented as... ,in, For spatial graph convolution output, Let be the set of neighboring nodes of node i. For learnable weight matrix, Let be the adjacency matrix, representing the connection strength between nodes i and j. For activation function, Represents the feature vector of node j; Temporal convolution: A one-dimensional convolutional kernel slides along the time dimension to capture the continuity of actions, ultimately outputting... ,in, This is the temporal convolution kernel.

5. The intelligent campus access control system based on artificial intelligence according to claim 4, characterized in that, Will The input is a fully connected layer, and the output is a probability distribution of behavior categories through a softmax function, represented as follows: ,in, For a set of behavior categories, For the confidence level of category c, For other categories, and Category C and The weight vector, and Category C and The bias term optimizes the model parameters through the cross-entropy loss function. The probability distribution output by the fully connected layer determines the specific category to which the current action belongs. If the confidence score of a certain category is the highest and exceeds the threshold of the corresponding category, an abnormal behavior alarm is triggered. The behavior type label output by the fully connected layer is combined with a spatiotemporal synchronization index to determine tailing behavior. Spatiotemporal synchronization analysis identifies behaviors that deliberately mimic previous card-swiping actions. The spatiotemporal synchronization index is defined as follows: ,in, and Let be the skeletal keypoint position vectors of the follower and the followed, respectively, at time t. For the phase difference of the two actions, when If the action type matches the threshold, it is determined to be a tailing behavior; A heatmap of behavioral features is generated by gradient-weighted class activation mapping, and abnormal behaviors are distinguished by comparing the differences in the heatmaps.

6. The intelligent campus access control system based on artificial intelligence according to claim 1, characterized in that, The specific steps for constructing a campus behavior knowledge graph and performing subgraph pattern mining based on the metadata tag library in the graph analysis layer are as follows: Metadata tag generation: Obtain structured data from the intelligent perception layer and the behavior understanding layer to generate tags; Knowledge graph storage: Neo4j graph database is used to store three types of entities and four types of relationships. Entities: personnel, areas, equipment; Relationships: Spatiotemporal intersection, peer records, device association, behavioral association; Graph Neural Network (GNN) Feature Learning: This involves aggregating features from node embeddings in a knowledge graph, represented as... ,in, For the embedding of node v at layer l, For the set of neighboring nodes, For node degree, The ReLU activation function is used to learn the high-order feature representations of nodes through a multi-layer graph convolutional network (GNN). Community detection algorithm: Identifies anomalous subgraphs by optimizing modularity, where modularity is represented as... ,in, Let i be the connection weight between nodes i and j. Let m be the node degree and m be the total weight. The indicator function is used to iteratively optimize the Q-value.

7. The intelligent campus access control system based on artificial intelligence according to claim 6, characterized in that, The graph analysis layer integrates three dimensions: individual historical behavior, network characteristics, and regional risk coefficient. It uses a weighted linear model to complete dynamic risk scoring, maps the regional risk scores to the campus map, and uses color gradients to display the risk distribution. Decision-making and execution level: Configure tiered early warning strategies and use blockchain technology to store evidence of key operations to ensure traceability.

8. The intelligent campus access control system based on artificial intelligence according to claim 7, characterized in that, The decision-making and execution layer defines a three-level early warning threshold based on the dynamic risk score of 0-100 points output by the weighted linear model, and uses an adaptive threshold algorithm to dynamically adjust the early warning thresholds based on historical event data. The initial three-level early warning thresholds are: A dynamic risk score of ≥80 indicates high risk: an alarm will be automatically triggered and a notification will be sent to the security personnel's mobile device, while video tracking and on-site handling will be initiated simultaneously. A dynamic risk score of 60 or less and a score of less than 80 indicates medium risk: logs are recorded and analysis reports are generated for subsequent review by security personnel; A dynamic risk score of <60 indicates low risk: the monitoring system will continuously track the risk without triggering an active response.

Images