A campus security control system fusing security monitoring and visitor management

Abstract

The application discloses a kind of campus security control systems of fusing security monitoring and visitor management, it is related to campus security control technical field, when visitor enters school, the identity information, school time and expected stay area of it are collected, visitor label is constructed;Visitor behavior reference model is matched based on historical behavior data, path trajectory set and behavior characteristic vector are generated;Real-time call monitoring image extracts current behavior characteristics, and reference characteristics are carried out similarity matching, and behavior deviation is calculated;Visitor current position and reference trajectory offset section are combined to carry out spatial analysis, and trigger abnormal behavior alarm instruction;And execute regional early warning, abnormal trajectory upload and high-risk label marking;Feedback data is used for the dynamic update and incremental learning of behavior model simultaneously;The method realizes the whole process dynamic monitoring and intelligent discrimination of visitor behavior, and improves the real-time performance and scalability of campus security system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of campus safety control, and particularly relates to a campus safety control system fusing security monitoring and visitor management. BACKGROUND

[0002] Currently, the personnel access management of campus places mainly relies on the registration of doorkeepers or the authentication of visitor systems, and some schools deploy camera security systems to realize basic video monitoring. However, the traditional security and visitor management systems operate independently of each other, lack data linkage, and cannot realize real-time early warning and linkage control on abnormal visitor behaviors.

[0003] More seriously, in low-grade primary schools or special education institutions, there is a large age difference and difficulty in identity recognition between visitors and students in school. Some abnormal visitors enter the campus by means of fuzzy identity or short-time contact, etc., which increases the risk of safety incidents. The traditional system is difficult to judge the real purpose of the visitor through a single visitor identity, and also cannot combine historical behavior patterns for early interception.

[0004] In addition, the current system cannot track the activity path of the visitor after entering the campus, especially in the corners of the campus where there are many monitoring blind spots or insufficient network signal coverage, such as old teaching buildings, outdoor activity areas, etc., which are easy to become potential risk blind spots, and lack a systematic and proactive risk identification mechanism. SUMMARY

[0005] The purpose of the present application is to provide a campus safety control system fusing security monitoring and visitor management to solve the problems in the background art.

[0006] In order to achieve the above purpose, the present application provides the following technical scheme: a campus safety control system fusing security monitoring and visitor management, comprising:

[0007] An information collection module: collecting visitor identity information V1 at the entrance of the campus, and recording the entering time T1 and the expected stay area L1 of the visitor, and constructing a visitor label set V;

[0008] A behavior matching module: according to the visitor identity information V1 and the historical visit record of the visitor, matching a corresponding visitor behavior reference model M1 from a visitor behavior model library;

[0009] A behavior modeling module: obtaining a typical activity path trajectory set P1 of the corresponding visitor in the campus and a behavior feature vector set F1 of each path node based on the model M1;

[0010] A behavior recognition module: calling real-time security monitoring image streams of the campus, extracting actual behavior feature vectors F2(t) of the visitor at each time period t, and performing similarity matching based on F2(t) and F1 to obtain a deviation index D(t).

[0011] Abnormality determination module: if D(t) exceeds the preset threshold, combined with the current actual position information L2(t) of the visitor and the offset section of the path trajectory set P1, automatically generate an abnormal behavior warning instruction;

[0012] Linkage early warning module: according to the abnormal behavior warning instruction, the area where the visitor is located is prewarned and prompted, and the abnormal trajectory is uploaded to the campus security terminal, and the visitor is marked as a high-risk label

[0013] Model updating module: update the feedback record of the corresponding model M1 in the visitor behavior model library.

[0014] Preferably, the matching of the corresponding visitor behavior reference model M1 from the visitor behavior model library comprises:

[0015] Structural coding is performed on the collected visitor identity information V1 to generate a feature vector for model retrieval , the feature vector includes visitor identity category, visit frequency and average stay time;

[0016] Based on the feature vector , similarity matching retrieval is performed in the visitor behavior model library, the cosine similarity algorithm is used to calculate the similarity score with the existing behavior model index, and the model with the highest score is selected as the reference model M1.

[0017] Preferably, the typical activity path trajectory set P1 of the corresponding visitor in the campus and the behavior feature vector set F1 of each path node are obtained based on the model M1, which comprises:

[0018] The historical trajectory template data is extracted from the reference model M1, and the commonly used path corresponding to the visitor identity category is mapped to the geographic coordinates based on the campus electronic map to generate a preliminary path set ;

[0019] The continuous coordinate points in are clustered and divided to identify key stay nodes and turning nodes, and a simplified path trajectory set P1 is constructed;

[0020] The behavior features of each node are extracted, four-dimensional feature vectors including walking speed, stay time, direction change rate and interaction frequency are calculated, and a node behavior feature set F1 is formed;

[0021] The generated path trajectory set P1 and the feature vector set F1 are time series bound to construct a behavior reference template that can be dynamically updated with the visitor.

[0022] ​Preferably, the real-time security monitoring image stream of the campus is retrieved, the actual behavior feature vector F2(t) of the visitor at each time period t is extracted, and similarity matching is performed between F2(t) and F1 based on F2(t) and F1 to obtain a deviation index D(t), including:

[0023] Real-time image frame sequences are collected and human body detection and target tracking are performed to obtain the moving track and key frame image of the visitor within a time period t;

[0024] Image feature extraction is performed on the image sequence to calculate the walking speed, stay duration, direction change rate, and interaction frequency to form the actual behavior feature vector F2(t);

[0025] F2(t) and the corresponding time point feature vector in F1 in the behavior reference template are matched in a one-to-one dimension manner, and the similarity score S(t) between the two vectors is calculated using the Euclidean distance;

[0026] The deviation index D(t) is calculated in reverse according to the similarity score.

[0027] Preferably, the deviation index D(t) is calculated in reverse according to the similarity score, including:

[0028] The calculated similarity score S(t) is taken as input to perform a reverse conversion function to obtain an initial deviation value;

[0029] The deviation threshold Dmax is dynamically set in combination with the visitor identity category and the area type;

[0030] D(t) is weighted and averaged in a sliding window manner within a continuous time period to obtain a smoothed deviation sequence ;

[0031] If the deviation index D(t) of the visitor at each of the continuous N time points exceeds Dmax, the behavior abnormal high-risk event is recognized.

[0032] Preferably, the current actual position information L2(t) of the visitor and the offset section of the path track set P1 are combined to automatically generate an abnormal behavior alarm instruction, including:

[0033] The visitor position L2(t) corresponding to the current time point t is compared with the behavior reference path track set P1 in terms of geographic coordinates, the shortest spatial distance d(t) is calculated, and it is judged whether it falls into the offset section of P1;

[0034] If d(t) is greater than a set spatial tolerance threshold r, and the deviation point does not belong to the expected stay area L1, it is determined as a track deviation event;

[0035] ​Combine the current deviation index D(t) with the trajectory deviation state, match the risk level trigger rule, and judge whether the condition for generating an alarm is met;

[0036] If the trigger condition is met, an abnormal behavior alarm instruction is generated, which includes the visitor identification, the abnormal type, the occurrence position, and the timestamp.

[0037] Preferably, the visitor is simultaneously marked as a high-risk label , comprising:

[0038] Receiving the generated abnormal behavior alarm instruction, parsing the visitor identification, the abnormal type, the position L2(t), and the timestamp t information contained therein;

[0039] Publishing a warning information to the teaching building or floor range corresponding to the current position of the visitor, prompting the on-site personnel to pay attention to the abnormal behavior;

[0040] Cutting the trajectory data during the abnormal behavior from the path trajectory set P1 and marking it as an abnormal segment P1M, uploading it to the campus security terminal platform and attaching a risk level label;

[0041] Updating the identity label of the visitor from the original visitor label V to the high-risk label and storing it in the visitor management database.

[0042] Preferably, the feedback record of the corresponding model M1 in the visitor behavior model library includes:

[0043] Receiving the high-risk label and the corresponding abnormal trajectory segment P1M as the feedback input sample of the visitor behavior reference model M1;

[0044] Performing trajectory difference analysis on P1M and the original behavior reference path trajectory set P1, extracting a difference feature vector ΔF, and accumulating behavior deviation weights according to the frequency of its occurrence;

[0045] Adjusting the corresponding feature dimension weight of the node behavior feature vector set F1 in the visitor behavior reference model M1;

[0046] When the cumulative feedback sample reaches a preset threshold, trigger the local retraining process of the visitor behavior reference model M1, update the trajectory pattern and feature weight distribution using an incremental learning algorithm, and save the result to the visitor behavior model library.

[0047] In the above technical solution, the present application provides technical effects and advantages:

[0048] 1. The campus security control method fuses security monitoring and visitor behavior modeling, realizes the whole-process closed-loop management of visitor from school registration, behavior trajectory tracking, behavior feature recognition to abnormal early warning and model self-update. Compared with the existing security system which can only rely on fixed path and access control, the application can dynamically identify the actual behavior state of the visitor during the school period, and perform personalized comparison and risk judgment based on the historical behavior model.

[0049] 2. The behavior deviation degree calculation, trajectory deviation analysis and high-risk label mechanism proposed in the application not only has a quantifiable risk judgment basis, but also continuously learns and optimizes through feedback driving model, effectively solving the problem of static model in the prior art which is difficult to adapt to diversified visitor behavior. Through multi-module linkage and data closed-loop update, the application significantly enhances the visitor safety control capability in the campus scene, and is suitable for visitor management in different types of campuses. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0051] Figure 1 The system module flowchart of the present application.

[0052] Figure 2 The method flowchart of the present application.

[0053] Figure 3 The visitor behavior reference model updating flowchart of the present application. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0055] Embodiment, please refer to Figure 1 、 2 , 3, the campus security control system fusing security monitoring and visitor management described in the embodiment, comprising:

[0056] An information collection module: collect visitor identity information V1 at the campus entrance, and record the entry time T1 and the expected stay area L1, and construct a visitor tag set V;

[0057] A behavior matching module: according to the visitor identity information V1 and the historical visit record, a corresponding visitor behavior reference model M1 is matched from the visitor behavior model library;

[0058] A behavior modeling module: based on the model M1, a typical activity path trajectory set P1 of the corresponding visitor in the campus is obtained, and a behavior feature vector set F1 of each path node is obtained;

[0059] A behavior recognition module: real-time security monitoring image stream of the campus is called, actual behavior feature vector F2(t) of the visitor in each time period t is extracted, and similarity matching is performed based on F2(t) and F1, and a deviation index D(t) is obtained;

[0060] An abnormality determination module: if D(t) exceeds a preset threshold, combined with the current actual position information L2(t) and the offset section of the path trajectory set P1, an abnormal behavior alarm instruction is automatically generated;

[0061] A linkage early warning module: according to the abnormal behavior alarm instruction, the area where the visitor is located is warned and prompted, and the abnormal trajectory is uploaded to the campus security terminal, and the visitor is marked as a high-risk tag ;

[0062] A model updating module: update the feedback record of the corresponding model M1 in the visitor behavior model library.

[0063] In the embodiment of the application, the information collection module is used for collecting the identity information and the initial state of the behavior of the visitor entering the campus, as the basis for subsequent behavior modeling and analysis.

[0064] Specifically, the information collection module includes an identity recognition unit, a time recording unit and a region preset unit:

[0065] The identity recognition unit is used to collect the identity information V1 of the visitor when the visitor passes through the campus entrance gate, the access control system or the visitor registration terminal, and the identity information includes but is not limited to visitor name, ID number, contact information, visit object, visit reason and other data, which can be automatically obtained through ID card scanning, face recognition or visitor registration two-dimensional code;

[0066] The time recording unit is used to record the entry time T1 of the visitor synchronously as the starting time point of the behavior analysis. The time recording can be obtained by timestamp binding with the trigger signal of the access control device or the face recognition camera;

[0067] The area preset unit is configured to determine the expected activity area L1 of the visitor according to the visitor registration information, the location of the visited person, or a system preset rule. The L1 can be a certain teaching building, an office, or a plurality of designated functional areas, serving as a geographical restriction parameter of the visitor tag.

[0068] After the information collection module completes the recording of V1, T1, and L1, the information collection module encapsulates the three to form a visitor tag set V={V1, T1, L1} and stores the tag set in a visitor behavior analysis database for subsequent behavior model matching and monitoring data comparison.

[0069] In a preferred embodiment of the present application, the behavior matching module is configured to dynamically match a reference model M1 most suitable for the current visitor from a visitor behavior model library according to the historical behavior characteristics of the visitor after the initial identity of the visitor is collected. Specifically, the behavior matching module includes:

[0070] First, the collected visitor identity information V1 is subjected to structured coding processing to generate a feature vector for behavior model matching. The feature vector is a multidimensional numerical vector, which includes but is not limited to the following three dimensions:

[0071] Visitor identity category: According to the identity field (such as “parent”, “supplier”, “maintenance personnel”, etc.) in V1, a preset identity category coding rule is used, for example: parent=1, supplier=2, maintenance personnel=3; visit frequency: the visit frequency f of the visitor in the past 30 days is counted as the frequency dimension value; average stay duration: the average value t_avg (unit: minutes) of the stay time of each visit in the historical visit is counted, and a floating point number is taken to one decimal place. The final obtained feature vector is in the form of: [C, f, t_avg], where C is the identity category code, f is the visit frequency, and t_avg is the average stay duration. The visitor behavior model library is a multidimensional feature index database, and each model M(i) is composed of the following contents: model number M(i), corresponding to a unique visitor behavior sample; historical feature vector ; corresponding path trajectory set P(i); behavior feature vector set F(i) (used for subsequent similarity comparison). The library is constructed based on historical visitor records, automatically updated incrementally every day, and similar visitor behaviors are merged through a clustering method (such as K-means) to improve the compactness and coverage of the model.

[0072] After obtaining , a similarity search is performed in the behavior model library. Specifically: a cosine similarity algorithm is called to compare the current feature vector with all historical model vectors Calculate the similarity; the cosine similarity calculation method is: taking and respectively as three-dimensional vectors, calculating the cosine of the angle, the value range is between 0 and 1, the larger the value represents the higher the similarity; select the model M(i) with the maximum similarity score (i.e. the maximum cosine of the angle) and take it as the reference behavior model M1 corresponding to the current visitor. For example, when the current visitor's is [2, 5, 45.5], compare it with all vectors in the model library one by one, calculate the cosine of each vector, and finally select the one with the highest score as M1.

[0073] After matching, the selected model M1 is associated with the label set V of the current visitor to form a data pair (V, M1), and is pushed to the behavior recognition module for subsequent generation of the path trajectory set P1 and the behavior feature vector set F1 of the visitor in the campus.

[0074] In the preferred embodiment of the present application, after the behavior modeling module matches the corresponding visitor's behavior reference model M1, it further performs path modeling and behavior feature extraction operations to construct a typical path trajectory set P1 and a node behavior feature vector set F1 suitable for the visitor type, which are used for subsequent real-time behavior deviation judgment and anomaly detection. Specifically, it includes:

[0075] First, extract the historical trajectory template data of the corresponding identity category visitor from the selected reference model M1. The trajectory template is composed of multiple positioning records of historical visitors in the campus, and each record contains a timestamp, location information and corresponding identity label.

[0076] Call the campus electronic map module (GIS engine) to map the above location data to the corresponding two-dimensional geographic coordinate point sequence. For example, if the visitor is a "parent" category, the mapped path will preferentially cover the teaching building A, the administrative building B and the school gate area. After merging all similar trajectories, a preliminary path set is obtained, where each path is an ordered set composed of multiple geographic coordinate points, representing the common travel route of this type of visitor.

[0077] Since the preliminary path set contains redundant coordinate points and abnormal deviation trajectories, in order to improve the efficiency of behavior modeling, the step of is performed.

[0078] A spatial clustering algorithm based on density (such as the DBSCAN algorithm) is used to classify points with a spatial distance less than a preset radius r (for example, 2 meters) and a high residence density as the same cluster. According to the clustering results, the following two types of key nodes are identified:

[0079] Residence nodes: such as classroom entrances, waiting areas, etc.

[0080] Turning points: such as stairway, crossroads, etc.

[0081] These key nodes are organized in a sequential manner to form a simplified path trajectory set P1, which describes the typical movement patterns of visitors in the campus.

[0082] For each key node in the path trajectory set P1, the corresponding monitoring image segment of the area is called to extract the visitor behavior features. The following four indicators are calculated as the dimensions of the node behavior feature vector:

[0083] Walking speed: calculated by the displacement of the visitor between image frames and the time difference, unit: m / s;

[0084] Stay time: statistics of the visitor's stay time at the node, unit: seconds;

[0085] Direction change rate: analysis of the angle change of the path direction entering and exiting the node to quantify the degree of direction deviation;

[0086] Interaction frequency: statistics of the frequency of close behavior or communication between visitors at the node, which can be realized by human body detection and posture recognition algorithm.

[0087] The above four feature values are combined to form a four-dimensional behavior feature vector , where v is the walking speed, t is the stay time, θ is the direction change rate, and f is the interaction frequency. After all the node feature vectors are collected, a complete behavior feature vector set F1 is formed.

[0088] Finally, each key node in the path trajectory set P1 is bound to the corresponding behavior feature vector Fi, and a time series relationship is established according to the order of node appearance. The final dynamic behavior reference template structure for behavior recognition is formed:

[0089] The template format is: M1={ (P1_1, F1_1), (P1_2, F1_2),..., (P1_n, F1_n)}, where P1_n represents the nth path node, and F1_n is the corresponding feature vector.

[0090] In a preferred embodiment of the present application, the behavior recognition module is used to compare the actual behavior of the visitor in the campus with the reference behavior features generated by the behavior modeling module, and through similarity calculation and deviation judgment, to identify whether the visitor has abnormal behavior trend.

[0091] First, through the network security camera deployed in each area of the campus, the monitoring image stream corresponding to the current location of the visitor is real-time retrieved, and the continuous image frame sequence within the time period t is extracted.

[0092] The deep learning human detection algorithm based on YOLOv5 or HRNet is used to match the identity of the personnel appearing in the image frame and track the target, so as to obtain the motion trajectory point set and key behavior frame of the target visitor in the time period t, and provide data input for subsequent feature extraction.

[0093] The above image frame is analyzed for behavior, and the following four dynamic features are extracted to construct the actual behavior feature vector F2(t) of the time period:

[0094] Walking speed v(t): calculated by target trajectory point displacement and time interval, unit: meters per second;

[0095] Stay time t_s(t): the time spent at the adjacent two key nodes, unit: seconds;

[0096] Direction change rate θ(t): based on the forward direction vector, the average of the direction angle change between each frame is calculated to reflect the trajectory stability;

[0097] Interaction frequency f(t): by analyzing the average distance between the visitor and others and the face orientation in the image, the number of close-range interaction behaviors is counted.

[0098] The finally constructed F2(t) is a four-dimensional vector: [v(t), t_s(t), θ(t), f(t)], which is used for comparison with the reference behavior feature F1.

[0099] F2(t) and the behavior reference feature vector F1(t) generated by the reference model M1 in the behavior modeling module are compared one by one. The similarity is calculated using the Euclidean distance algorithm. Specifically: let F2(t) = [x1, x2, x3, x4], F1(t) = [y1, y2, y3, y4], then the similarity score ; the similarity S(t) ranges from 0 to 1, and the closer the value is to 1, the closer the current behavior is to the expected behavior. This similarity score will be used as the basis for the deviation calculation.

[0100] The S(t) is converted into the deviation index D(t) by using the reverse mapping method to reflect the degree of abnormality of the current behavior. The conversion function is: , where the value range of D(t) is 0 to 1, and the larger D(t) is, the more serious the behavior deviation is.

[0101] According to the visitor identity category (such as parents, off-campus technical personnel) and the activity area type (such as teaching area, administrative area), the deviation threshold Dmax is set; the setting of Dmax is based on the historical behavior deviation mean μ and standard deviation σ of the visitor in similar scenarios, and the recommended setting is: where μ and σ can be obtained by model training phase statistics.

[0102] Considering that short-time behavior disturbance or device jittering can cause temporary fluctuations of D(t), the present application introduces a sliding time window smoothing algorithm, which specifically includes: setting the sliding window length to T seconds, and the window contains multiple time points D(t); performing weighted average on all D(t) values in the window to obtain the smoothed deviation degree ; the weight can be distributed in a decreasing manner over time, and the value close to the current time point has a higher weight, for example, an exponential decay weight distribution is adopted.

[0103] When it is judged that the deviation degree D(t) at the continuous N time points (such as N = 3) exceeds the current set threshold value Dmax, an abnormally high-risk event marker is triggered.

[0104] In the preferred embodiment of the present application, the abnormality determination module is used to determine whether the abnormal behavior alarm condition is reached and automatically generate the corresponding alarm instruction in the case that the visitor behavior deviates from the reference trajectory, by comprehensively comparing the current actual position information and the path trajectory set. This module relies on spatial position analysis and rule matching strategy to realize accurate determination of behavior risk.

[0105] First, the visitor actual position information L2(t) corresponding to the current time point t is compared with the reference path trajectory set P1 generated by the behavior modeling module at the geographic coordinate level.

[0106] Specifically, assuming that P1 is a set of multiple node coordinate points, and L2(t) is the two-dimensional coordinate value (x, y) of the current position information, the system obtains the minimum distance value d(t) by calculating the Euclidean distance between L2(t) and all nodes in P1, which represents the shortest spatial deviation between the current visitor and the expected path.

[0107] If d(t) is less than or equal to the preset spatial tolerance threshold r (for example, set to 3 meters), it is considered that the visitor is still within the expected path range; if d(t) is greater than r, it enters the deviation state judgment link. The tolerance threshold r can be dynamically adjusted according to the campus environment structure, monitoring coverage radius, and historical path deviation distribution data.

[0108] On the basis of determining that d(t) exceeds the threshold r, the system further judges whether the deviation point falls within the visitor's expected stay area L1. The expected stay area L1 is usually generated when the visitor registers, indicating the normal access range (such as administrative building, laboratory, etc.).

[0109] ​If L2(t) deviates from P1 but is within the range of L1, it is considered as a reasonable path change and no trajectory deviation flag is triggered; if L2(t) is neither within the path buffer of P1 nor in the stay area defined by L1, it is determined as a trajectory deviation event, and information such as the deviation time point, position and minimum distance d(t) is recorded for subsequent rule matching.

[0110] Once the trajectory deviation event is marked, the system will combine the current behavior deviation degree indicator D(t) as an input parameter to enter the risk rule matching engine for alarm condition judgment.

[0111] The rule engine uses a configurable rule set based on the IF-THEN structure, and example rules include:

[0112] IF D(t)≥0.6 AND d(t)≥5 meters AND the current area is a restricted access area THEN risk level = high;

[0113] IF D(t)≥0.4 AND deviation duration≥60 seconds THEN risk level = medium;

[0114] Where D(t) is the behavior deviation degree calculated in the behavior recognition module, d(t) is the position deviation. The deviation duration can be calculated by accumulating the deviation state in the continuous time period.

[0115] If the matching result meets the preset high or medium risk level, it is determined that the current visitor behavior has the condition to trigger the alarm.

[0116] After the rule matching judgment meets the alarm triggering condition, a structured abnormal behavior alarm instruction is immediately generated, including the following fields:

[0117] Visitor identification: such as visitor ID, name, registration information;

[0118] Abnormal type: such as "high deviation behavior", "path deviation to restricted area", etc.

[0119] Occurrence position: geographic coordinates or floor number of the current position L2(t);

[0120] Timestamp: the exact time point t when the abnormal behavior is triggered;

[0121] Risk level: the level label output by the rule engine (such as medium, high).

[0122] In a preferred embodiment of the present application, after receiving the abnormal behavior alarm instruction, the linkage early warning module not only performs the area warning and trajectory uploading operation, but also further performs risk level assessment on the visitor triggering the abnormal behavior, and updates its identity tag to a high risk label , to achieve continuous dynamic supervision of suspicious personnel. The process includes the following four steps:

[0123] First, receive the abnormal behavior alarm instruction generated by the anomaly determination module, which is a structured data object containing the following key fields:

[0124] Visitor identification information: such as visitor unique ID, registration number, or face recognition code;

[0125] Abnormal type: such as "path deviation + high behavior deviation";

[0126] Current location L2(t): represented by latitude and longitude coordinates or building number + floor number;

[0127] Timestamp t: the exact time point of the abnormal event, in milliseconds.

[0128] After parsing the instruction, bind the visitor ID with the current positioning data to locate its physical location and identity attributes for subsequent operations.

[0129] To implement on-site intervention and risk notification, the system calls the campus broadcast control module or information publishing interface to push warning information to the physical area corresponding to the visitor's current location. The specific methods include: if the location is inside a teaching building, push to the corresponding floor electronic screen to display "abnormal behavior reminder"; if the location is an outdoor area, the system can issue voice or image reminders through voice broadcast devices or LED display screens. The broadcast content can be configured according to the abnormal level, such as "medium risk reminder" or "high risk warning", and the playback duration and frequency can be configured.

[0130] Further extract the trajectory segment in the abnormal behavior occurrence time period from the visitor's behavior reference path trajectory set P1, denoted as abnormal segment P1M. The extraction method is: according to the alarm time point t and the set front and back compensation window T (such as 10 seconds before and after), cut out the corresponding path segment; the obtained trajectory segment P1M includes continuous position point sequence, behavior feature change, stay and turning data, etc.; P1M is packaged together with visitor ID and risk level label as a trajectory anomaly report, uploaded to the campus security terminal platform or security console. The platform is used for subsequent dispatch of security personnel, locking of camera view or retrieval of video backtracking.

[0131] Call the visitor management database interface to update the current visitor's identity label V to a high-risk label and complete persistent storage. The label structure is updated as follows: the original label V includes: {visitor ID, identity category, entry time T1, expected area L1}; the high-risk label The extended field is: {original V information, risk level, abnormal trajectory ID, marking time T, processing state Flag}. Among them: the risk level is based on the rule engine matching result (such as medium, high); the abnormal trajectory ID is associated with the P1M number uploaded to the security platform; the processing state Flag is used to mark whether it has been responded or intervened.

[0132] In a preferred embodiment of the present application, the model updating module is used to feed back the high-risk behavior generated by the visitor in the actual behavior process to the corresponding behavior reference model M1, realizing the self-learning and dynamic optimization of the behavior model. Through a series of processing steps such as abnormal trajectory difference analysis, feature weight correction and incremental learning update, the model M1 improves the recognition accuracy of similar behaviors in the future.

[0133] When the linkage early warning module marks a visitor as a high-risk label , the system synchronously extracts the trajectory segment P1M corresponding to its abnormal behavior, and takes both as the feedback update input of the model M1.

[0134] P1M is a continuous trajectory segment extracted within the time window of triggering abnormal behavior alarm, containing a sequence of spatial coordinates and behavior feature changes; The visitor identification field carried in P1M is used to accurately locate the corresponding behavior reference model M1; the input sample is structured as follows: , and temporarily stored in the model updating buffer area, waiting for subsequent processing. In order to accurately identify the deviation features between the abnormal behavior and the reference model, difference analysis is performed on the original path trajectory set P1 in P1M and M1. The specific operation is as follows:

[0135] The trajectory nodes in P1 and P1M corresponding to the same time window or the same space region are compared in coordinates, and the key indicators such as spatial offset Δd, direction offset angle Δθ, and stay time difference Δt_s are calculated;

[0136] The above difference values are converted into a unified four-dimensional difference feature vector ΔF = [Δv, Δt_s, Δθ, Δf], where each dimension corresponds to a behavior feature in F1;

[0137] The frequency of each type of ΔF in the historical feedback sample is counted to form a behavior deviation statistical matrix, which is used to evaluate the weight correction amplitude of the feature dimension.

[0138] The deviation weight can be calculated by frequency normalization, that is, the weight adjustment value of each feature dimension ∝ appearance frequency / total feedback times.

[0139] According to the ΔF extracted from the feedback sample and the cumulative deviation statistical result, the weight of each dimension of the behavior feature vector set F1 in the model M1 is dynamically adjusted. The adjustment method is as follows:

[0140] Let the feature of the i-th dimension in the original feature vector F1 be fi, and its corresponding weight be wi; if the offset of this dimension in AF frequently occurs, the system will update wi according to the following formula: the new weight , where a is the learning rate constant (e.g. 0.1), fi_impact is the weight increment factor after frequency normalization; the feature vector after weight adjustment is retained in the model M1 as the reference vector for the next behavior comparison.

[0141] Continue to count the number of feedback samples of all visitor behavior models, and when the cumulative feedback sample number of a model M1 reaches a set threshold Q (e.g. Q = 10 times), trigger the local retraining mechanism of the model. The retraining uses an incremental learning algorithm, and the main steps are as follows:

[0142] Use the historical normal trajectory data and the cumulative abnormal trajectory P1M samples to mix to form a training set; use a time series clustering-based learning method (such as a DTW dynamic time warping + K-means variant) to regenerate the typical path trajectory and the updated behavior feature vector set ; replace P1 and F1 in the original model M1 with the retraining result, and store the updated model in the visitor behavior model library.

[0143] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application.

Claims

1. A campus security control system integrating security monitoring and visitor management, characterized in that: include: Information collection module: Collect visitor identity information V1 at the campus entrance, record their entry time T1 and expected stay area L1, and construct visitor tag set V; Behavior matching module: Based on visitor identity information V1 and its historical school visit records, match the corresponding visitor behavior reference model M1 from the visitor behavior model library; Behavioral modeling module: Based on model M1, obtain the typical activity path trajectory set P1 of the corresponding visitor in the campus, and the behavioral feature vector set F1 of each path node; Behavior recognition module: Retrieves real-time campus security surveillance image stream, extracts the visitor's actual behavior feature vector F2(t) for each time period t, and performs similarity matching between F2(t) and F1 to obtain the deviation index D(t), including: Real-time acquisition of image frame sequences and execution of human detection and target tracking to obtain the visitor's movement trajectory and keyframe images within time period t; Image features are extracted from the image sequence, and walking speed, dwell time, rate of change of direction and frequency of interaction are calculated to form the actual behavior feature vector F2(t); The feature vectors at the corresponding time points in F1 of the behavioral reference template are matched one-to-one according to their dimensions, and the similarity score S(t) between the two vectors is calculated by Euclidean distance. The deviation index D(t) is calculated in reverse based on the similarity score, specifically including: The calculated similarity score S(t) is used as input to execute the inverse transformation function. The initial deviation value is obtained; The deviation threshold Dmax is dynamically set by combining the visitor's identity category and the type of region they are in; By applying a weighted average to D(t) over a continuous time period using a sliding window method, a smoothed deviation sequence is obtained. ; If N consecutive time points If all exceed Dmax, it is identified as a high-risk event of abnormal behavior; Anomaly detection module: If D(t) exceeds the preset threshold, it automatically generates an abnormal behavior alarm command by combining the visitor's current actual location information L2(t) with the offset segment of the path trajectory set P1. Linked Early Warning Module: Based on abnormal behavior alarm commands, the module issues early warnings for the area where the visitor is located, uploads the abnormal trajectory to the campus security terminal, and simultaneously marks the visitor as a high-risk individual. ; Model Update Module: Updates the feedback records of the corresponding model M1 in the visitor behavior model library.

2. The campus security control system integrating security monitoring and visitor management according to claim 1, characterized in that: The step of matching the corresponding visitor behavior reference model M1 from the visitor behavior model library includes: The collected visitor identity information V1 is structured and encoded to generate feature vectors for model retrieval. The feature vector This includes visitor identity category, visit frequency, and average stay duration; Based on feature vectors Perform similarity matching retrieval in the visitor behavior model library, call the cosine similarity algorithm to calculate its similarity score with the existing behavior model index, and select the model with the highest score as the reference model M1.

3. The campus security control system integrating security monitoring and visitor management according to claim 1, characterized in that: The process of obtaining the typical activity path trajectory set P1 of the corresponding visitor on campus based on model M1, and the behavioral feature vector set F1 of each path node, includes: Historical trajectory template data is extracted from the reference model M1, and geographic coordinate mapping is performed on the commonly used routes corresponding to visitor identity categories based on the campus electronic map to generate a preliminary set of routes. ; right Clustering of continuous coordinate points in the data, identifying key stopping points and turning points, and constructing a simplified path trajectory set P1; Behavioral features are extracted from the monitoring image segments at each node, and a four-dimensional feature vector containing walking speed, dwell time, direction change rate and interaction frequency is calculated to form the node behavioral feature set F1. The generated path trajectory set P1 is bound to the feature vector set F1 over time to construct a behavior reference template that can be dynamically updated with visitors.

4. A campus security control system integrating security monitoring and visitor management according to claim 1, characterized in that: The method of automatically generating abnormal behavior alarm instructions by combining the visitor's current actual location information L2(t) with the offset segment of the path trajectory set P1 includes: Compare the visitor's location L2(t) at the current time point t with the behavioral reference path trajectory set P1 using geographic coordinates, calculate the shortest spatial distance d(t), and determine whether it falls within the offset segment of P1. If d(t) is greater than the set spatial tolerance threshold r, and the deviation point does not belong to its expected dwell area L1, it is determined to be a trajectory deviation event; By combining the current deviation index D(t) with the trajectory offset status, the risk level triggering rules are matched to determine whether the conditions for generating an alarm are met. If the triggering conditions are met, an abnormal behavior alarm command containing the visitor identifier, abnormal type, location of occurrence, and timestamp will be generated.

5. A campus security control system integrating security monitoring and visitor management according to claim 4, characterized in that: in, The visitor was simultaneously labeled as high-risk. ,include: Receive the generated abnormal behavior alarm command and parse the visitor identifier, abnormal type, location L2(t) and timestamp t information contained therein; Issue early warning information to the teaching building or floor area corresponding to the visitor's current location to alert on-site personnel to abnormal behavior; The trajectory data during the period of abnormal behavior is cut from the path trajectory set P1 and marked as abnormal trajectory segment P1M, and uploaded to the campus security terminal platform with a risk level label attached. Update the visitor's identity tag from the original visitor tag V to a high-risk tag. And store it in the visitor management database.

6. A campus security control system integrating security monitoring and visitor management according to claim 5, characterized in that: The updated feedback records for the corresponding model M1 in the visitor behavior model library include: Receiving high-risk labels And the corresponding abnormal trajectory segment P1M, as the feedback input sample for the visitor behavior reference model M1; Perform trajectory difference analysis between P1M and the original behavior reference path trajectory set P1, extract the difference feature vector ΔF, and accumulate the behavior offset weight according to its occurrence frequency. Adjust the weights of the corresponding feature dimensions in the node behavior feature vector set F1 in the visitor behavior reference model M1; When the cumulative feedback samples reach a preset threshold, the local retraining process of the visitor behavior reference model M1 is triggered. The incremental learning algorithm is used to update the trajectory pattern and feature weight distribution, and the results are saved to the visitor behavior model library.

Images