A method for tracking and real-time archiving of personnel trajectories in a park based on face clustering

Abstract

This invention provides a method for tracking and real-time archiving of personnel trajectories in a park based on face clustering. The method includes: spatiotemporal calibration and security level labeling of multiple cameras; extraction of facial feature vectors using face detection and deep convolutional neural networks; incremental face clustering based on spatiotemporal constraints to establish cross-camera identity associations; trajectory completion based on a park access network map to detect tailgating, loitering behavior, and abnormal paths; real-time verification of the matching between personnel security levels and area security levels to detect accompanying access risks; dynamic updating of personnel digital profiles, construction of social relationship networks, and generation of visualized personnel portraits; and execution of tiered risk warnings to solidify a complete chain of evidence. This invention, by integrating spatiotemporal constraint clustering, trajectory completion, complex behavior analysis, real-time archiving, and accompanying access detection, achieves real-time detection and prevention of unauthorized access and abnormal behavior by personnel in confidential parks, thereby enhancing information security capabilities.

Description

Technical Field

[0001] This invention relates to the field of park security management technology, and in particular to a method for tracking and real-time archiving of park personnel trajectories based on facial clustering. Background Technology

[0002] With the acceleration of urbanization and the upgrading of park management models, the construction of smart parks is gradually transitioning from single-point intelligence to platform-based linkage. This is especially true in classified parks where there are extremely high requirements for personnel access management, internal activity monitoring, and protection of sensitive areas. Traditional security systems mainly rely on single-point methods such as access cards and video surveillance, which are insufficient to create a complete trajectory profile of personnel activities. Furthermore, they suffer from problems such as a disconnect between trajectory tracking and identity verification, a lack of dynamic risk warnings, insufficient clustering analysis capabilities, and a lack of real-time archiving mechanisms. Therefore, a method for personnel trajectory tracking and real-time archiving in classified parks is needed that can achieve real-time tracking of personnel trajectories throughout the park, dynamically verify the matching of personnel location with security levels, identify abnormal gatherings and unauthorized access behaviors, and simultaneously protect privacy. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method for tracking and real-time archiving of personnel trajectories in parks based on facial clustering, in order to solve the problems of disconnect between personnel identity and behavior, lack of dynamic risk warning, and difficulty in preventing leaks through behavioral analysis in existing classified parks.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solution: This invention provides a method for tracking and real-time archiving of personnel trajectories in a park based on face clustering, comprising the following steps: S1. Spatiotemporal calibration of multiple cameras deployed in the park is performed to establish a mapping relationship between pixel coordinates and park geographic coordinates, and the corresponding security level attributes of each area are marked to construct a park access network map; the park access network map includes camera location nodes, key intersection nodes, building entrance and exit nodes, and passages. S2, perform face detection on the video frames captured by each camera, extract face region images, and extract high-dimensional face feature vectors through a deep convolutional neural network; S3. Based on the park access network map, perform incremental clustering of the extracted facial feature vectors based on spatiotemporal constraints to establish cross-camera personnel identity associations; compare the clustering results with the park information database to determine the personnel identity and its corresponding confidentiality level, and establish initial personnel digital profiles; the park information database stores the identity information, facial feature vectors and their corresponding confidentiality level information of registered park personnel, etc. S4, Generate a sequence of personnel trajectory points based on the personnel identity association results; identify breakpoints in the trajectory sequence, complete the trajectory based on the park access network map to obtain complete trajectory data, and upload the complete trajectory data to the park information database to form the historical trajectory data of the corresponding personnel; S5. Based on the complete trajectory data, personnel confidentiality level, and area confidentiality level, perform behavior detection and real-time permission verification. The behavior detection includes: tailgating intrusion detection, loitering behavior detection, and path anomaly detection. The real-time permission verification includes: real-time location unauthorized access detection and trajectory backtracking unauthorized access detection. S6. Based on the complete trajectory data, behavior detection results, and real-time permission verification results, dynamically update the personnel digital profile; construct a social relationship network based on spatiotemporal co-occurrence frequency, wherein the spatiotemporal co-occurrence frequency is defined as: within the same area, the spatial distance between two personnel is less than a preset distance threshold and the duration of this state exceeds a preset time threshold; based on the social relationship network and the personnel confidentiality level, use statistical modeling methods to identify abnormal social relationships across confidentiality levels; S7, Based on the complete trajectory data and the social relationship network, detect accompanying access by personnel across different security levels; S8 establishes a multi-level risk warning mechanism based on the detected tailing intrusion, loitering behavior, abnormal path, real-time location overreach, trajectory backtracking overreach, and accompanying access results, and executes differentiated response strategies according to the preset risk level classification standards.

[0005] As a further improvement, the spatiotemporal calibration in S1 includes: obtaining the intrinsic and extrinsic parameter matrices of the camera through a camera calibration algorithm, establishing a conversion function from image pixel coordinates to park geographic coordinates; dividing the park map into grid cells, marking the confidentiality level of each grid cell, and establishing a grid index structure that supports efficient coordinate-based queries.

[0006] As a further improvement, the incremental clustering based on spatiotemporal constraints in S3 includes: for the extracted facial feature vector, calculating its cosine similarity with each existing cluster center, and filtering candidate cluster sets whose similarity exceeds a threshold; performing time accessibility verification on the candidate clusters, calculating the ratio of the actual time difference to the minimum travel time, wherein the minimum travel time is obtained by dividing the shortest path distance in the park access network map by the maximum movement speed of people, or dynamically determined based on the statistical quantile of historical travel time; performing spatial continuity verification, verifying the existence of a connected path based on the park access network map, and eliminating spatially inaccessible candidates; assigning the facial feature vector to the cluster with the highest similarity among the candidate clusters that have passed the verification, and creating a new cluster if there are no candidate clusters that have passed the verification.

[0007] As a further improvement, the trajectory completion in S4 includes: identifying breakpoints in the trajectory sequence where the time gap exceeds a threshold; using the point before the breakpoint as the starting point and the point after the breakpoint as the ending point, calculating possible completion paths on the park's access network map using the shortest path algorithm; combining the historical trajectory patterns of personnel to perform confidence weighting on the completion paths, prioritizing completion results with high overlap with historical paths; verifying whether the travel time of the completion path matches the time gap of the breakpoint, and if they match, inserting the completion path into the trajectory sequence.

[0008] As a further improvement, the behavior detection in S5 includes: The tailgating intrusion detection: monitors the behavior of the current person closely following the person in front of them into the access control of the confidential unit within a preset time threshold without performing independent identity authentication. If the behavior is met, it is determined to be tailgating intrusion. The loitering behavior detection: calculates the density of trajectory points of personnel in unauthorized areas per unit time. If it exceeds a preset density threshold, it is determined to be loitering behavior. The path anomaly detection involves monitoring the deviation between the actual trajectory path of personnel with low security levels and the optimal path. The optimal path is the shortest path calculated based on the park's access network map, and all areas along the path meet the personnel's current security level permissions. The deviation is defined as the ratio of the actual path length to the optimal path length. If a person continuously passes through the periphery of multiple high security level areas and the deviation exceeds the threshold, it is determined to be a path anomaly.

[0009] As a further improvement, the real-time permission verification in S5 includes: The real-time location overreach detection compares the security level of the area where the person is currently located with the person's own security level. If the security level of the area is higher than the person's security level, it is determined that an overreach has been committed. The trajectory backtracking unauthorized access detection: Based on the personnel's historical trajectory data, backtracking detection is performed to check whether there are access records with an area confidentiality level higher than the personnel's historical confidentiality level.

[0010] As a further improvement, the accompanying access in S7 includes: The system detects that the spatial distance between personnel at the first level of security and personnel at the second level of security remains less than a threshold within a set time window, wherein the first level of security is higher than the second level of security. Based on the identification results of abnormal social relationships in the social relationship network, it determines whether personnel at the second level of security have independent access to the current area. If not, it is determined to be accompanying access.

[0011] A campus personnel trajectory tracking and real-time archiving system based on face clustering includes: The multi-source video acquisition module is used to acquire video data through multiple cameras deployed within the park; The spatiotemporal configuration module is used to perform spatiotemporal calibration on each camera, establish the mapping relationship between pixel coordinates and geographical coordinates of the park, mark the security level attributes of each area, and construct a network map of access to the park. The face detection and feature extraction module is used to detect faces in video frames, extract face region images, and extract face feature vectors through a deep convolutional neural network. The face clustering and identity association module is used to perform incremental clustering of extracted face feature vectors based on spatiotemporal constraints based on the park access network map, and establish cross-camera personnel identity association. The trajectory reconstruction and spatiotemporal analysis module is used to generate a sequence of personnel trajectory points based on the personnel identity association results, complete the trajectory in the monitoring blind spot, extract behavioral features, and detect tailgating, loitering behavior, and abnormal path behavior. The security level matching and verification module is used to establish a real-time matching and verification mechanism between personnel security levels and regional security levels based on personnel digital files and social relationship networks, and to detect unauthorized access and accompanying access behaviors. The real-time data aggregation and personnel profiling module is used to create dynamically updated digital profiles of personnel based on their trajectory point sequences and behavioral characteristics, integrating basic information, facial features, trajectory history, behavioral patterns, and social relationship networks. The risk warning and response module is used to establish a multi-level risk warning mechanism based on tailgating behavior, loitering behavior, abnormal path behavior, and real-time location verification and accompanying orientation verification results, and to execute differentiated response strategies according to preset risk level classification standards.

[0012] The beneficial effects of this invention are: This invention effectively filters physically unreachable face associations by introducing spatiotemporal constraints, including temporal accessibility and spatial continuity. It overcomes the problem of erroneous associations that are easily generated by traditional pure feature similarity clustering in cross-camera scenarios, and improves the continuity and accuracy of cross-camera personnel identity tracking.

[0013] By constructing a network map of the park's access routes and integrating shortest path reasoning and access time verification, a trajectory completion method suitable for monitoring blind spots is provided. This method can reconstruct the complete movement trajectory of personnel even when camera coverage is discontinuous, eliminate monitoring breakpoints, and ensure the integrity and availability of trajectory data.

[0014] Breaking through the limitations of traditional static access control permissions, a three-dimensional real-time verification mechanism of "personnel location-security level-time" is established. This mechanism not only enables real-time unauthorized access detection, historical trajectory tracing, and accompanying access identification, but also achieves dynamic and refined access control, thereby improving the level of intelligence in access control of classified areas.

[0015] Through a real-time data aggregation mechanism, personnel files are continuously updated with each activity, supporting online learning of behavioral patterns and anomaly detection. This evolves from static identity management to dynamic behavioral understanding, enhancing the ability to perceive internal threats. By analyzing social networks and detecting aggregation patterns, it identifies risky behaviors such as unusual contact and illegal gatherings before actual information leaks, achieving a shift from "post-event tracing" to "pre-event prevention." Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a method for tracking and real-time archiving of park personnel based on face clustering, as proposed in this invention.

[0018] Figure 2 This is a schematic diagram of the framework of a park personnel trajectory tracking and real-time data aggregation system based on face clustering, according to the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] Reference Figure 1-2 As shown, a method for tracking and real-time archiving of park personnel trajectories based on face clustering includes: S1. Spatiotemporal calibration is performed on multiple cameras deployed within the park to establish a mapping relationship between pixel coordinates and park geographic coordinates, and the security level attributes corresponding to each area are marked to construct a park access network map. Specifically, the park access network map includes camera location nodes, key intersection nodes, building entrance and exit nodes, and passageways. A video surveillance network covering the entire park is deployed, including cameras in public areas, cameras in classified areas, and cameras in boundary areas; spatiotemporal calibration is performed on each camera to establish a mapping relationship between pixel coordinates and park geographic coordinates, and the security level attributes corresponding to the field of view coverage area of ​​each camera are recorded. The spatiotemporal calibration includes: obtaining the intrinsic and extrinsic parameter matrices of the cameras through a camera calibration algorithm (such as the Zhang Zhengyou calibration method), and establishing image pixel coordinates (…). ) to the park's geographical coordinates ( The conversion function is used. The park map is divided into grid cells, with the grid size set to 2m × 2m based on the positioning accuracy requirements. The security level of each grid cell is marked. Where n is the number of security levels (e.g., ordinary, internal, confidential, top secret). Establish a grid index structure that supports efficient coordinate-based queries to quickly retrieve the security level of any location within a given area.

[0021] Constructing a directed graph of the access network in the park V is a set of nodes, including: camera location nodes. Key intersection nodes Building entrance / exit nodes Elevator / staircase nodes E is the set of edges, representing feasible pathways (pedestrian walkways, driveways, stairs, corridors, etc.), with edge weights. Indicates travel time. Edge weight calculation: ,in For Euclidean distance, The speeds are corresponding to different types of passageways. Typically, the speed for pedestrian walkways is 1.2 m / s, for vehicular traffic lanes it's 5 m / s, and for stairs it's 0.8 m / s. Access control verification is delayed; typically, it's 0 seconds for regular areas and 3-5 seconds for confidential areas. The network graph supports dynamic updates: when temporary construction or road closures occur in the park, administrators can disable the corresponding edges through the configuration interface, and the system will automatically recalculate the shortest path.

[0022] S2 performs face detection on video frames captured by each camera, extracts face region images, and extracts high-dimensional face feature vectors using a deep convolutional neural network. Specifically, a multi-task cascaded convolutional neural network is used for face detection to obtain face bounding boxes and key point locations. The detection process includes: first, quickly generating candidate windows and obtaining the bounding box regression vectors of face regions through a fully convolutional network; second, refining the candidate windows and eliminating a large number of non-face windows; and finally, refining and outputting the locations of 5 face key points.

[0023] Liveness detection is performed using 3D structured light or infrared binocular vision technology, analyzing facial depth information and natural micro-expressions to determine if it is a real face; specifically, an RGB+Depth dual-modal fusion strategy is employed. For cameras equipped with depth sensors, facial depth maps are acquired using 3D structured light, and a depth consistency score is calculated. For ordinary RGB cameras, a liveness detection method based on remote photoplethysmography (RPPG) is used to determine authenticity by analyzing minute periodic changes in facial skin color. The liveness detection threshold is dynamically adjusted according to the risk level of the scene, with the exception of highly confidential areas. Confidential area ordinary area .

[0024] The validated face region image is aligned and transformed using an affine transformation to map key points to standard positions. The normalized 112×112 pixel face image is then input into a pre-trained deep convolutional neural network to extract a 512-dimensional face feature vector. The values ​​of each dimension are floating-point numbers; the deep convolutional neural network is trained using an angular marginal loss function to enhance the inter-class discriminative power of feature vectors from different individuals; the extracted feature vectors are L2 normalized. ,make This facilitates the calculation of cosine similarity.

[0025] S3, based on the park access network map, incremental clustering based on spatiotemporal constraints is performed on the extracted facial feature vectors to establish cross-camera personnel identity associations; the clustering results are compared with the park information database to determine the personnel identity and its corresponding confidentiality level, and an initial personnel digital profile is established; the park information database stores the identity information, facial feature vectors, and their corresponding confidentiality level information of registered park personnel, etc. Specifically, the spatiotemporal constraints include time accessibility constraints and spatial continuity constraints; the time accessibility constraint calculates the minimum travel time based on the shortest path distance and the maximum movement speed of personnel in the park access network map, and filters according to the ratio of the actual time difference to the minimum travel time to exclude physically impossible time associations; the spatial continuity constraint verifies whether there is a connected path between two points based on the park access network map to exclude spatially inaccessible associations; incremental face clustering includes: initializing an empty cluster set. Each cluster Includes: Cluster center feature vector Member feature set and its spatiotemporal attributes; the spatiotemporal attributes include: timestamp t, geographic location (x, y), and camera identifier for each feature. wait; Incremental clustering process: For newly extracted face feature vectors Calculate its relationship with each existing cluster center. Cosine similarity:

[0026] All feature vectors have undergone L2 normalization. Filtering is performed based on similarity exceeding a threshold. The candidate cluster set,

[0027] In this embodiment, if there is no historical statistical data, then the threshold is... =0.65, and its corresponding cosine distance is 0.35.

[0028] The spatiotemporal constraints include: time reachability verification, and calculation of facial feature vectors. timestamp With candidate clustering Timestamp of recent features Time difference Minimum passage time Employing a dynamic threshold strategy: Strategy A: If there is no historical data or the amount of historical data is less than 30 records, then the ratio of the shortest path distance of the road network to the maximum movement speed of people is used, i.e. ,in From the park access network map Location to The shortest path distance to the nearest location The maximum movement speed of personnel is set to 2.0 m / s in this embodiment; The congestion factor is dynamically adjusted based on the real-time pedestrian density within the park: when sparse... Normally 2. During crowded times .

[0029] Strategy B: If historical travel time statistics exist, then take the 85th percentile of the historical travel time for the same route (i.e., from the starting camera to the ending camera) within the past 30 days as the minimum travel time. ,in This represents the 85th percentile, meaning 85% of historical travel times are less than this value. This strategy adapts to the actual travel habits of people in the park, improving the accuracy of time constraints. Time accessibility determination: If... If the time is unreachable, the candidate is excluded.

[0030] Spatial continuity verification: Verification based on the campus access network diagram from Check if there is a connected path from the current location to the nearest location in the candidate cluster. Use a depth-first search or breadth-first search algorithm to determine the connectivity between the two points. If the two points are located in different buildings and there is no feasible path (e.g., separated by a wall), they are considered spatially discontinuous and excluded from the candidate list.

[0031] After time-constrained filtering, the candidate clusters that pass the dual-constraint verification are sorted in descending order of similarity. Assign the cells to the cluster with the highest similarity and update the cluster centers:

[0032] Where n is the number of members in the original cluster. The spatiotemporal attribute range of the cluster is also updated. If no candidate cluster passes the validation, a new cluster is created. ,initialization .

[0033] The clustering results are compared with the park's information database, and the maximum similarity with the features in the database is calculated. .like If the identity is confirmed, it is marked as the identity of the corresponding internal personnel, and their confidentiality level is obtained. In this embodiment, the identity confirmation threshold is set to 0.80; otherwise, it is marked as a temporary visitor identity, a temporary ID is assigned, and a new profile is created.

[0034] S4, Generate a sequence of personnel trajectory points based on the personnel identity association results; identify breakpoints in the trajectory sequence, complete the trajectory based on the park access network map to obtain complete trajectory data, and upload the complete trajectory data to the park information database to form the historical trajectory data of the corresponding personnel. Specifically, the trajectory points are defined as six-tuples. ,in Here, t represents the geographical coordinates of the park, t represents the timestamp, and id represents the personnel identification identifier associated with the cluster. For camera identification, It is classified as a regional security level.

[0035] The trajectory points are arranged in chronological order to form the original trajectory sequence. In this embodiment, the time difference between adjacent points is required. If the value exceeds the limit, it is considered a trajectory break.

[0036] Trajectory completion includes: identifying time gaps in the trajectory sequence. Exceeding the threshold The break point, in this embodiment preferably For each breakpoint, execute the trajectory completion process: based on the park access network map. ,in These are location nodes, which include camera locations and key intersections. The edge represents a feasible path, with its weight equal to the travel time; the edge weight is the point before the break point. For the starting point and the subsequent point Using the destination as the endpoint, the shortest path algorithm is used to calculate the most likely complete path on the park's traffic network map. To further improve the rationality of the completed paths, the system maintains a database of normal trajectory patterns for each individual, recording their historically high-frequency paths. During completion, multiple candidate completed paths are weighted according to their "historical frequency of occurrence."

[0037] in For path length, As the weighting coefficient, in this embodiment, we set... , This is an indicator function. Historically high-frequency paths are prioritized for completion; if a historical path differs from the shortest path, the historical path takes precedence, reflecting individual behavioral preferences.

[0038] Verify the passage time of the completed path. Is it related to the time interval of fracture? The trajectory is matched (with an allowable error of ±20% in this embodiment). If a match is found, the completed path is inserted into the trajectory sequence and marked as "complete segment" to distinguish it from the measured segment; if no match is found, it is marked as "cannot be completed". The completed trajectory data is then uploaded to the park's information database as a record of the person's historical trajectory.

[0039] S5, based on complete trajectory data, personnel confidentiality level, and area confidentiality level, performs behavior detection and real-time permission verification. Behavior detection includes: tailgating intrusion detection, loitering behavior detection, and path anomaly detection; real-time permission verification includes: real-time location unauthorized access detection and trajectory backtracking unauthorized access detection.

[0040] Tail-and-run detection: Monitors continuous detection events within the access control area of ​​classified units. If the event occurs within a preset time window threshold... If two individuals with different IDs are detected entering the access control system sequentially, and the latter individual fails to perform independent identity authentication (i.e., the latter individual's face is not detected to remain in front of the access control card reader for ≥1 second), then this is determined to be tailgating. In this embodiment, a preferred setting is... If the condition is met within 1 to 3 seconds, it is determined to be a tailgating intrusion, triggering a Level 1 warning.

[0041] Loitering behavior detection: Calculates the time spent by personnel within a unit of time. Unauthorized internal areas (i.e.) , For regional security level, Track point density (for personnel confidentiality level) ,in The number of trajectory points. This represents the area of ​​the region. If... Exceeding the preset density threshold Furthermore, based on the analysis of the historical trajectory of the person being monitored, if the area is not their regular work area, it is determined to be loitering behavior, triggering a level-two warning. In this embodiment, preferably... Set to 5 minutes. Set to 10 points / minute.

[0042] Path anomaly detection: Monitor the actual trajectory paths of personnel with low security levels. with the optimal path The deviation is defined as follows: the optimal path is calculated based on the campus network map to find the shortest path from the starting point to the ending point, where the security level of all areas along the path does not exceed the current security level of the individual. The deviation is defined as the ratio of the actual path length to the optimal path length. If the actual path is more than 50% longer than the optimal path, and the actual path passes through multiple high-security areas with an outer distance of less than 20m, it is judged as a path anomaly, triggering a level 3 warning.

[0043] Real-time location access control: Real-time acquisition of the security level of the area where a person is currently located. This is related to the individual's own level of confidentiality. Perform a comparison. If... If so, it is determined to be real-time unauthorized access.

[0044] Trajectory backtracking unauthorized access detection: When a person's confidentiality level changes due to personnel changes or security audits (the system receives external input of the level change event), it automatically backtracks all access records in the person's historical trajectory whose timestamps are earlier than the change time, marks the records in which the regional confidentiality level exceeds the new confidentiality level as historical unauthorized access, and generates a violation report for auditing.

[0045] Spatiotemporal consistency verification: Checking the files of the same person. Within a preset short time window Are there two objects that are more than a preset distance threshold apart? The capture point is preferably set in this embodiment. =2 seconds =50 meters, if it exists, it is judged as a risk of identity theft, triggering a level one warning.

[0046] S6 dynamically updates personnel digital profiles based on complete trajectory data, behavior detection results, and real-time permission verification results; it constructs a social relationship network based on spatiotemporal co-occurrence frequency, where spatiotemporal co-occurrence frequency is defined as: within the same area, the spatial distance between two personnel is less than a preset distance threshold, and the duration of this state exceeds a preset time threshold; based on the social relationship network and personnel confidentiality level, it uses statistical modeling methods to identify abnormal social relationships across confidentiality levels. Specifically, the personnel digital profile includes: basic information such as identity identifier, name, department, and current confidentiality level. Contact information, etc.; facial features include feature vector set, quality score, anti-counterfeiting verification record, update time, etc.; trajectory history includes complete trajectory sequences of the most recent N days. The system includes: regional access statistics, duration of stay distribution, and entry frequency; behavioral patterns, including a normal trajectory pattern library, abnormal behavior records, and tags for tailing / loitering / abnormal paths; social relationship networks, including a list of co-occurring individuals, relationship weights, community affiliation, and cross-level connection markers; and risk status, including current risk score, risk tag set, warning history, and handling records.

[0047] Social relationship networks are constructed based on spatiotemporal co-occurrence frequency: co-occurrence is defined as: within a preset time window Inside, the spatial distance between two people is less than Furthermore, the duration of this state exceeds a preset time. In this embodiment, it is preferably set to... Set to 30 minutes. The distance is 10 meters. Count the number of co-occurrences. Construct a weighted undirected graph =( , Edge weight ,in , The total number of detections for the first person i and the second person j are respectively, and this weight reflects the degree of co-occurrence between the two.

[0048] Anomaly Social Relationship Identification: For each individual, the co-occurrence frequency with personnel at different security levels is calculated, and a co-occurrence frequency distribution is constructed. Based on the normal distribution assumption, cross-security level co-occurrence relationships with a co-occurrence frequency higher than "the mean co-occurrence frequency of personnel at the same security level plus three standard deviations" are marked as anomaly social connections. This model can effectively identify potential information leakage risk paths, such as low-security personnel frequently co-occurring with high-security personnel outside of classified areas. Community detection algorithms are used to identify close social groups and mark anomaly social connections across security levels.

[0049] S7, based on complete trajectory data and social relationship networks, detects accompanying visits by personnel across different security levels. Specifically, it detects visits by personnel at the first and second security levels within a set time window. The internal space distance is consistently less than the threshold. The first level of security is higher than the second level of security. In this embodiment, it is preferably set to... Set to 10 minutes. Accompanying behaviors are defined as those occurring at sampling points 5 meters or more in total, meeting the criteria. Based on abnormal social relationships identified in the social relationship network, it is determined whether individuals with low confidentiality levels possess independent access to the current area (i.e., whether the individual's confidentiality level is greater than or equal to the current area's confidentiality level). If individuals with low confidentiality levels do not possess independent access, they are classified as high-risk companions; if they possess independent access but exhibit abnormal social relationships, they are classified as medium-risk companions, recorded, and monitored more frequently.

[0050] Further, through long-term data statistics, abnormal accompanying patterns are identified: if the co-occurrence frequency of personnel at different levels is significantly higher than the average co-occurrence frequency of personnel at the same level, it is judged as a frequency abnormality; if the accompanying visits occur during non-working hours (such as nighttime or holidays), it is judged as a time period abnormality; if the accompanying stay occurs outside core areas such as top-secret laboratories and archives, it is judged as a regional abnormality.

[0051] S8, based on the detected tailing intrusion, loitering behavior, abnormal paths, unauthorized real-time location access, unauthorized trajectory backtracking, and accompanying access results, establishes a multi-level risk warning mechanism and executes differentiated response strategies according to preset risk level classification standards. Specifically, the graded warning includes: Level 1 warning (highest risk): the triggering conditions are unauthorized real-time location access to top-secret areas, attempts to tailgate into classified areas, identity impersonation risk (abnormal spatiotemporal consistency verification), and abnormal breaches of physical isolation facilities, etc. The response strategy is to immediately trigger audible and visual alarms, lock relevant access control, push notifications to the security command center's large screen and handheld terminals, and initiate full-park tracking. Level 2 Warning (Medium Risk): Triggering conditions include unauthorized access to classified areas, high-risk accompanying visits, unauthorized entry into classified areas during unauthorized periods, and trajectory patterns significantly deviating from the individual's historical baseline, such as loitering or abnormal paths. The response strategy is to push warning information to security personnel in the responsible area, call in associated cameras to track and capture images, and record key monitoring. Level 3 Warning (General Risk): Triggering conditions include access to unauthorized areas, prolonged stay at the boundary of sensitive areas, and abnormal contact with key personnel. The response strategy is to record and file the information in the system, increase the frequency of detection, and include the individual in the behavior analysis for key monitoring.

[0052] Furthermore, this invention also provides a system for tracking and real-time archiving of personnel trajectories in classified parks based on spatiotemporal constrained face clustering, used to execute the above-mentioned method. The system includes: The multi-source video acquisition module is used to acquire video streams from various areas of the park in real time through multiple cameras deployed in the park, and output the raw video data to the face detection and feature extraction module.

[0053] The spatiotemporal configuration module is used to perform spatiotemporal calibration on each camera, establish the mapping relationship between camera coordinates and geographical coordinates of the park, mark the corresponding security level attributes of each area, and construct a park access network map; and synchronize the network map data to the face clustering and identity association module, trajectory reconstruction and spatiotemporal analysis module, and security level matching and verification module.

[0054] Face detection and feature extraction module: used to detect faces in video frames, extract face region images, and extract face feature vectors through deep convolutional neural networks; the feature vectors and corresponding spatiotemporal information (such as timestamps, camera positions, etc.) are sent to the face clustering and identity association module.

[0055] Face clustering and identity association module: Based on the park access network map, it performs spatiotemporally constrained incremental clustering on the extracted face feature vectors to establish cross-camera personnel identity association; generates personnel identity identifiers and outputs them to the trajectory reconstruction and spatiotemporal analysis module and the real-time clustering and personnel profiling module.

[0056] The trajectory reconstruction and spatiotemporal analysis module is used to generate a sequence of personnel trajectory points based on the personnel identity association results, complete the trajectory in the monitoring blind spot, extract behavioral features, and detect tailing intrusion behavior, loitering behavior, and abnormal path behavior; the trajectory data and behavior detection results are output to the real-time file aggregation and personnel profiling module and the risk warning and response module, respectively.

[0057] Real-time file aggregation and personnel profiling module: It is used to establish dynamically updated digital files of personnel based on personnel trajectory point sequences and behavioral characteristics, and integrate basic information, facial features, trajectory history, behavioral patterns and social relationship networks; it constructs a social relationship network based on spatiotemporal co-occurrence, identifies abnormal social connections across confidentiality levels, and pushes the file and social network data to the confidentiality level matching and verification module and the risk warning and response module.

[0058] Confidentiality Level Matching and Verification Module: This module is used to establish a real-time matching and verification mechanism between personnel confidentiality levels and regional confidentiality levels based on personnel digital files and social relationship networks, detecting unauthorized access and accompanying access behaviors; and outputting the results of unauthorized access and accompanying risks to the risk warning and response module.

[0059] Risk warning and response module: It is used to establish a multi-level risk warning mechanism based on tailgating and intrusion behavior, loitering behavior, abnormal path behavior, and real-time location verification and accompanying access verification results, and to execute differentiated response strategies according to preset risk level classification standards.

[0060] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A park personnel trajectory tracking and real-time archiving method based on face clustering, characterized in that, Includes the following steps: S1. Spatiotemporal calibration of multiple cameras deployed in the park is performed to establish a mapping relationship between pixel coordinates and park geographic coordinates, and the corresponding security level attributes of each area are marked to construct a park access network map; the park access network map includes camera location nodes, key intersection nodes, building entrance and exit nodes, and passages. S2, perform face detection on the video frames captured by each camera, extract face region images, and extract high-dimensional face feature vectors through a deep convolutional neural network; S3. Based on the park access network map, perform incremental clustering of the extracted facial feature vectors based on spatiotemporal constraints to establish cross-camera personnel identity associations; compare the clustering results with the park information database to determine the personnel identity and its corresponding confidentiality level, and establish initial personnel digital profiles; the park information database stores the identity information, facial feature vectors and their corresponding confidentiality level information of registered park personnel, etc. S4, Generate a sequence of personnel trajectory points based on the personnel identity association results; identify breakpoints in the trajectory sequence, complete the trajectory based on the park access network map to obtain complete trajectory data, and upload the complete trajectory data to the park information database to form the historical trajectory data of the corresponding personnel; S5. Based on the complete trajectory data, personnel confidentiality level, and area confidentiality level, perform behavior detection and real-time permission verification. The behavior detection includes: tailgating intrusion detection, loitering behavior detection, and path anomaly detection. The real-time permission verification includes: real-time location unauthorized access detection and trajectory backtracking unauthorized access detection. S6. Based on the complete trajectory data, behavior detection results, and real-time permission verification results, dynamically update the personnel digital profile; construct a social relationship network based on spatiotemporal co-occurrence frequency, wherein the spatiotemporal co-occurrence frequency is defined as: within the same area, the spatial distance between two personnel is less than a preset distance threshold and the duration of this state exceeds a preset time threshold; based on the social relationship network and the personnel confidentiality level, use statistical modeling methods to identify abnormal social relationships across confidentiality levels; S7, Based on the complete trajectory data and the social relationship network, detect accompanying access by personnel across different security levels; S8 establishes a multi-level risk warning mechanism based on the detected tailing intrusion, loitering behavior, abnormal path, real-time location overreach, trajectory backtracking overreach, and accompanying access results, and executes differentiated response strategies according to the preset risk level classification standards.

2. The method for tracking and real-time archiving of park personnel trajectories based on face clustering according to claim 1, characterized in that, The spatiotemporal calibration in S1 includes: obtaining the intrinsic and extrinsic parameter matrices of the camera through a camera calibration algorithm, establishing a conversion function from image pixel coordinates to park geographic coordinates; dividing the park map into grid cells, marking the confidentiality level of each grid cell, and establishing a grid index structure that supports efficient coordinate-based queries.

3. The method for tracking and real-time archiving of park personnel trajectories based on face clustering according to claim 1, characterized in that, The incremental clustering based on spatiotemporal constraints in S3 includes: for the extracted facial feature vector, calculating its cosine similarity with each existing cluster center, and filtering candidate clusters whose similarity exceeds a threshold; performing time accessibility verification on the candidate clusters, calculating the ratio of the actual time difference to the minimum travel time, wherein the minimum travel time is obtained by dividing the shortest path distance in the park access network map by the maximum movement speed of people, or dynamically determined based on the statistical quantile of historical travel time; performing spatial continuity verification, verifying the existence of a connected path based on the park access network map, and eliminating spatially inaccessible candidates; assigning the facial feature vector to the cluster with the highest similarity among the candidate clusters that have passed the verification, and creating a new cluster if there are no candidate clusters that have passed the verification.

4. The method for tracking and real-time archiving of park personnel trajectories based on face clustering according to claim 1, characterized in that, The trajectory completion in S4 includes: identifying breakpoints in the trajectory sequence where the time gap exceeds a threshold; using the point before the breakpoint as the starting point and the point after the breakpoint as the ending point, calculating possible completion paths on the park's access network map using the shortest path algorithm; combining the historical trajectory patterns of personnel to perform confidence weighting on the completion paths, prioritizing completion results with high overlap with historical paths; verifying whether the travel time of the completion path matches the time gap of the breakpoint, and if they match, inserting the completion path into the trajectory sequence.

5. The method for tracking and real-time archiving of park personnel trajectories based on face clustering according to claim 1, characterized in that, The behavior detection in S5 includes: The tailgating intrusion detection: monitors the behavior of the current person closely following the person in front of them into the access control of the confidential unit within a preset time threshold without performing independent identity authentication. If the behavior is met, it is determined to be tailgating intrusion. The loitering behavior detection: calculates the density of trajectory points of personnel in unauthorized areas per unit time. If it exceeds a preset density threshold, it is determined to be loitering behavior. The path anomaly detection involves monitoring the deviation between the actual trajectory path of personnel with low security levels and the optimal path. The optimal path is the shortest path calculated based on the park's access network map, and all areas along the path meet the personnel's current security level permissions. The deviation is defined as the ratio of the actual path length to the optimal path length. If a person continuously passes through the periphery of multiple high security level areas and the deviation exceeds the threshold, it is determined to be a path anomaly.

6. The method for tracking and real-time archiving of park personnel trajectories based on face clustering according to claim 1, characterized in that, The real-time permission verification in S5 includes: The real-time location overreach detection compares the security level of the area where the person is currently located with the person's own security level. If the security level of the area is higher than the person's security level, it is determined that an overreach has been committed. The trajectory backtracking unauthorized access detection: Based on the personnel's historical trajectory data, backtracking detection is performed to check whether there are access records with an area confidentiality level higher than the personnel's historical confidentiality level.

7. The method for tracking and real-time archiving of park personnel trajectories based on face clustering according to claim 1, characterized in that, The accompanying access in S7 includes: The system detects that the spatial distance between personnel at the first level of confidentiality and personnel at the second level of confidentiality remains less than a threshold within a set time window, wherein the first level of confidentiality is higher than the second level of confidentiality. Based on the identification results of abnormal social relationships in the social relationship network, it determines whether personnel at the second level of confidentiality have independent access to the current area. If not, it determines that they have accompanied access.

8. A system for tracking and real-time archiving of park personnel trajectories based on facial clustering, characterized in that, The system is used to perform the method according to any one of claims 1 to 7, comprising: The multi-source video acquisition module is used to acquire video data through multiple cameras deployed within the park; The spatiotemporal configuration module is used to perform spatiotemporal calibration on each camera, establish the mapping relationship between pixel coordinates and geographical coordinates of the park, mark the security level attributes of each area, and construct a network map of access to the park. The face detection and feature extraction module is used to detect faces in video frames, extract face region images, and extract face feature vectors through a deep convolutional neural network. The face clustering and identity association module is used to perform incremental clustering of extracted face feature vectors based on spatiotemporal constraints based on the park access network map, and establish cross-camera personnel identity association. The trajectory reconstruction and spatiotemporal analysis module is used to generate a sequence of personnel trajectory points based on the personnel identity association results, complete the trajectory in the monitoring blind spot, extract behavioral features, and detect tailgating, loitering behavior, and abnormal path behavior. The security level matching and verification module is used to establish a real-time matching and verification mechanism between personnel security levels and regional security levels based on personnel digital files and social relationship networks, and to detect unauthorized access and accompanying access behaviors. The real-time data aggregation and personnel profiling module is used to create dynamically updated digital profiles of personnel based on their trajectory point sequences and behavioral characteristics, integrating basic information, facial features, trajectory history, behavioral patterns, and social relationship networks. The risk warning and response module is used to establish a multi-level risk warning mechanism based on tailgating behavior, loitering behavior, abnormal path behavior, and real-time location verification and accompanying orientation verification results, and to execute differentiated response strategies according to preset risk level classification standards.

Images