A big data-based intelligent campus traffic management system and method

Abstract

The application discloses a kind of big data-based wisdom campus traffic management system and method, and the application relates to wisdom campus management technical field, the present application is by collecting RFID card reader data, monitoring video data, timetable data, and the calibration number is generated by the monitoring data calibration algorithm driven by timetable, based on decision tree and the split strategy of compound rank is constructed risk grading model, by input real-time calibration number, timetable number, time period type, RFID coverage, historical congestion frequency, weather condition, output corresponding risk level, and according to real-time risk level, corresponding measures are taken, the specific reason of generating this risk can also be analyzed.The present application can dynamically adjust risk level, significantly reduce misjudgment and omission, realize fine risk control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of smart campus management, in particular to a smart campus traffic management system and method based on big data. BACKGROUND

[0002] Affected by the scale of the campus and the flow of teachers and students, traditional campus traffic management faces problems such as inaccurate data collection, extensive risk assessment, and delayed emergency response.

[0003] In the prior art, campus passenger flow statistics mainly rely on single monitoring cameras or RFID technology. Although the monitoring cameras have a wide coverage, they cannot distinguish target groups, and the counting includes a large number of irrelevant personnel, resulting in calibration deviation. Although the RFID technology can accurately identify students with cards, it has coverage blind spots and reading problems, and cannot completely capture the flow of students without cards or those who do not carry campus cards.

[0004] At the same time, the prior art mostly uses fixed formulas to calculate passenger flow density, without combining the campus work and rest rules (such as the flow difference between large and small breaks), weather influence (rainy day indoor route passenger flow surge), and historical congestion data, resulting in extensive risk classification, and the prior art does not analyze the reasons for the situation according to the risk level, so that the best treatment plan cannot be taken when the situation occurs again.

[0005] In summary, the existing campus traffic management system has problems such as one-sided data collection, rigid risk assessment, and delayed response mechanism, and there is an urgent need for a smart management solution that integrates multi-source data, dynamically classifies risks, and accurately responds to emergencies. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a smart campus traffic management system and method based on big data, which solves the problems of large error of single data source, extensive risk assessment, and delayed emergency response.

[0007] To achieve the above purpose, the present application realizes the following technical scheme: a smart campus traffic management system based on big data, comprising:

[0008] A data collection module is configured to collect RFID card reader data, monitoring video data, and course schedule data, and generate a calibrated number of people through a course schedule driven monitoring data calibration algorithm;

[0009] A dynamic analysis module is configured to construct a risk classification model based on a decision tree and a composite rank splitting strategy, input real-time calibrated number of people, course schedule number of people, time period type, RFID coverage rate, historical congestion frequency, and weather conditions, and output a risk level;

[0010] The response analysis module takes corresponding measures according to the real-time risk level and analyzes the specific reasons for the risk.

[0011] As a further scheme of the present application, the calibration number is calculated according to the formula , wherein R represents the number of students with cards, M represents the number of students in the timetable, and K represents the total number of students captured by the monitoring camera, is the cardless rate, is the recognition rate of the monitoring on the students.

[0012] As a further scheme of the present application, the specific steps for constructing the risk level module are as follows:

[0013] A plurality of groups of historical data are obtained, and the risk level of each group of data is manually labeled;

[0014] Features of each group of historical data are extracted, and the features include the calibration number C, the number of students in the timetable M, the time period type T, the RFID coverage rate RC, the historical congestion frequency H, and the weather condition W.

[0015] The extracted features are subjected to data cleaning and conversion;

[0016] The composite rank of each feature is calculated, and the feature list is sorted according to the composite rank;

[0017] The training set and the test set are divided in a ratio of 8:2;

[0018] The features are selected according to the sorted feature table for splitting, the average Gini index drop value before and after splitting is calculated, if the drop value is greater than or equal to the drop threshold value, the splitting is accepted, otherwise the next feature is selected for splitting;

[0019] The above process is repeated for each sub-node, and the high-rank features are preferentially tried until all features are tried or the average Gini index drop value meets the standard, thereby obtaining the trained decision tree model;

[0020] The test set is substituted into the trained decision tree model, the accuracy, recall rate and F1 value are calculated, and the final performance of the model is evaluated.

[0021] As a further scheme of the present application, the specific steps for converting the extracted features are as follows:

[0022] For C and M, binning processing is adopted, if C

[0023] For T, the time length is combined with the traffic stage to obtain a composite feature, and the composite feature is converted into a one-hot encoding vector, and the time length specifically includes a large interval, a small interval and a non-interval, and the traffic stage specifically includes an end-of-class peak, a non-peak and a pre-class peak;

[0024] For W, one-hot encoding is adopted, rain encoding is [1, 0], snow encoding is [0, 1], and sunny encoding is [0, 0];

[0025] For RC, a min-max normalization method is adopted to normalize it to the interval [0, 1];

[0026] For H, binaryzation is adopted, if H >= H1, it is encoded as 1, and if H < H1, it is encoded as 0.

[0027] As a further scheme of the application, for T, the specific steps of converting the composite feature into a one-hot encoding vector are:

[0028] The time length and the traffic stage are independently encoded, the first two bits represent the time length, that is, a large interval = 10, a small interval = 01, and a non-interval = 00, and the last three bits represent the traffic stage, that is, an end-of-class peak = 100, a non-peak = 010, and a pre-class peak = 001;

[0029] The time length and the traffic stage are combined to obtain a composite feature, and the composite feature is converted into a one-hot encoding vector, specifically, {a large interval, an end-of-class peak} = [1, 0, 1, 0, 0], {a large interval, a non-peak} = [1, 0, 0, 1, 0], {a large interval, a pre-class peak} = [1, 0, 0, 0, 1], {a small interval, a non-peak} = [0, 1, 0, 1, 0], {a small interval, an end-of-class peak} = [0, 1, 1, 0, 0], {a small interval, a pre-class peak} = [0, 1, 0, 0, 1], {a non-interval, a non-peak} = [0, 0, 0, 1, 0], {a non-interval, an end-of-class peak} = [0, 0, 1, 0, 0], and {a non-interval, a pre-class peak} = [0, 0, 0, 0, 1].

[0030] As a further scheme of the application, the specific method for calculating the composite rank of each feature is:

[0031] Invite campus safety experts to score the importance of each feature, and assign ranks in descending order of scores to obtain expert score ranks R1;

[0032] Calculate the Spearman correlation coefficient of each feature and the risk level, and assign ranks in descending order of absolute values to obtain correlation coefficient ranks R2;

[0033] According to the formula R = sqrt(R1 * R2), the composite rank is calculated.

[0034] As a further scheme of the present application, the parameters for initializing the decision tree model specifically include:

[0035] the maximum number of layers is 5;

[0036] the minimum number of samples required for node splitting is 10;

[0037] the minimum number of samples required for leaf nodes is 5;

[0038] select the Gini index drop value as the feature splitting criterion.

[0039] As a further scheme of the present application, the specific steps for taking corresponding measures according to the real-time risk level are:

[0040] if the decision tree model output is low risk, a strategy of regular round patrol combined with intelligent compression is adopted; if the decision tree model output is medium risk, a strategy of key attention combined with dynamic optimization is adopted; if the decision tree model output is high risk, a strategy of real-time overview combined with three-dimensional prevention and control is adopted;

[0041] the strategy of regular round patrol combined with intelligent compression includes reducing the monitoring video resolution to 720P, extending the round patrol interval to 10 seconds; the security patrol frequency is once every hour; the data storage period is compressed to 7 days;

[0042] the strategy of key attention combined with dynamic optimization includes adjusting the monitoring video resolution to 1080P, shortening the round patrol interval to 5 seconds; the security patrol frequency is increased to once every 20 minutes;

[0043] the strategy of real-time overview combined with three-dimensional prevention and control includes adjusting the monitoring video resolution to 4K, starting real-time AI people counting and abnormal behavior recognition; the security personnel arrive at the scene within 5 minutes and stay in real time; the data storage period is permanently retained, and cloud backup is triggered.

[0044] As a further scheme of the present application, the step of analyzing the risk causes is:

[0045] all splitting features from the root node to the target risk level leaf node are recorded in the trigger order;

[0046] for the extracted features, the composite rank R calculated in the training phase is used for screening: the top 20% high-priority features are retained, and arranged in ascending order of composite rank;

[0047] generate attribution text according to the priority order: high-priority features lead to low-priority features, and finally trigger the risk level.

[0048] A smart campus traffic management method based on big data, comprising:

[0049] S1, collect RFID card reader data, monitoring video data, and timetable data, and generate calibration number through a calibration algorithm driven by the timetable;

[0050] S2, construct a risk grading model based on a decision tree and a split strategy of composite rank, input real-time calibration number, timetable number, time period type, RFID coverage rate, historical congestion frequency, and weather condition, and output risk grade;

[0051] S3, take corresponding measures according to the real-time risk grade, and analyze the specific reason for the risk.

[0052] The present application provides a big data-based intelligent campus traffic management system and method, which has the following advantages compared with the prior art:

[0053] (1) The present application fuses three source data of timetable theoretical number, monitoring actual count, and RFID accurate reading, constructs a dynamic calibration algorithm, solves the problems of incomplete single monitoring coverage, RFID missed reading, and missed detection of non-card crowd, uses the timetable as a reference value to detect data deviation, combines monitoring coverage area and RFID target filtering, realizes accurate calibration of target group flow number, effectively reduces the counting error of the traditional single technical solution, and provides reliable data basis for risk assessment;

[0054] (2) The risk grading model based on the decision tree of the present application fuses time period type, weather, historical congestion and other multi-dimensional characteristics, combines the composite rank priority strategy, automatically learns the campus work and rest rules and complex scene risk mode, breaks through the single threshold limitation of the traditional fixed formula, dynamically adjusts the risk grade for scenes such as high peak during a large class, indoor flow surge in rainy days, significantly reduces the false and missed judgments, and realizes fine risk control. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 is a principle block diagram of the present application;

[0056] Figure 2 is a step flowchart of the present application. DETAILED DESCRIPTION

[0057] The technical solutions 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 only 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.

[0058] As Figure 1 , the present application provides a big data-based intelligent campus traffic management system, which comprises:

[0059] Data acquisition module for collecting RFID card reader data, monitoring video data, and course schedule data, and generating calibration number through course schedule driven monitoring data calibration algorithm;

[0060] The RFID card reader is mainly installed at the main entrance, such as the main passageway of a teaching building, a laboratory, or a dormitory building, or a high-frequency flow node, such as the main stairway of each floor of a teaching building (connecting the core passageway of each floor), and the exit of an equipment room of an experimental building.

[0061] The campus card is usually bound to the corresponding class, and the RFID card reader can count the number of students with the card in real time, naturally excluding other classes, staff, and other non-target groups. Since the RFID technology is mature, the reading accuracy can reach more than 95% in non-congestion scenarios, providing a benchmark for the calibration model as strong verification data. Especially when the monitoring number is misjudged due to insufficient light, the stable count of RFID can pull back the calibration result, avoiding the amplification of the deviation of single monitoring data.

[0062] The installation positions of the monitoring device mainly include the main entrance and side door of the teaching building, such as the main door, side door, and fire passage, which are easily missed by RFID coverage. Key path nodes, such as corridor intersections, stairway entrances, and corridor connections. Cross-regional connection points, such as the necessary passageway from the laboratory to the teaching building and the outdoor corridor from the cafeteria to the teaching building.

[0063] The specific reasons for RFID combined with monitoring equipment are as follows:

[0064] RFID is usually only deployed at the main entrance, but students may pass through the side door or fire passage, especially during peak hours.

[0065] Monitoring can record the student's stop or detour from area A to area B, such as passing through the library corridor when going from the classroom to the playground, which supplements the defect that RFID can only record the start and end points.

[0066] Students who do not carry a campus card cannot be identified by RFID, such as new students who have not received their cards or damaged cards, but monitoring can capture their presence through face recognition or group counting in combination with the campus card photo library.

[0067] The theoretical number specified by the course schedule is the rigid demand for student flow. For example, a certain university detected 700 students flowing through a certain path before class through RFID and monitoring, while according to the course schedule of the entire school, there should be 710 students attending class at that time. Due to the problem of missed reading, 10 people were not detected, triggering the compensation of the missed detection of the calibration model.

[0068] According to the formula The number of students to be calibrated is calculated, where R represents the number of students with cards, representing the basic reliability data; M is the theoretical number of students, i.e., the number of students who need to attend classes according to the timetable; and K is the total number of students captured by the surveillance cameras. The cardless rate can be obtained based on historical statistics. To monitor the student recognition rate, it can be obtained from historical statistics;

[0069] R, as the measured value of the main channel, solves the problem of missed readings on the main path for students with cards. The reliability of R can be improved by deploying multiple card readers.

[0070] To compensate for the undercounting of students without cards at non-main passages, and to avoid underestimating the overall number of students due to the lack of RFID at side gates;

[0071] K may include a large number of irrelevant personnel, such as teachers and visitors. The theoretical number of students without cards is eliminated using K-(MR), and then... Filtering retains only valid target students, thus solving the problem of false positives in monitoring.

[0072] M serves as a rigid baseline to prevent abnormal data fluctuations from interfering with data. For example, if temporary activities cause a sudden increase in K, MR is used to constrain the compensation range and prevent the number of calibrators from deviating excessively from the theoretical value.

[0073] For example, during the long break on Friday morning, the theoretical number of students moving between classes is M=1200, the actual number of students measured at the main gate using RFID is R=1000, the total number of students monitored is K=300, and the historical cardless rate... =10%, monitoring recognition rate =85%;

[0074] The theoretical number of students not carrying their campus cards is calculated as follows: MR = 1200 - 1000 = 200 students, meaning the timetable shows that 200 students did not carry their campus cards.

[0075] Calculate the unidentified portion of students without cards: =10%×200=20 people, meaning that 10% of the 200 people were not identified by the monitoring.

[0076] The number of students monitored exceeds the theoretical number of students without cards: K-(MR)=300-200=100 people, meaning that there are actually 100 students who exceed the theoretical number of students without cards, possibly because students with cards passed through the side door.

[0077] The number of students effectively identified as target students is calculated to be 100 × 85% = 85, after which 15 other irrelevant individuals are excluded.

[0078] The number of calibrators, Pcal, is 1000 + 20 + 85 = 1105.

[0079] The actual number of people flowing is 1100, if only using RFID statistics, it will get 1000, 85 students with cards and 20 students without cards passing through the side door are missed, the error reaches 18.75%, if only using monitoring equipment statistics, it will get 300, which is obviously lower than the actual number of people flowing, the error reaches 75%, and according to the calibration formula, it gets 1105, which is close to the actual number of people flowing, the error reaches 0.45%.

[0080] The dynamic analysis module constructs a risk classification model based on the decision tree and the composite rank splitting strategy, inputs the calibration number, the course table number, the time period type, the RFID coverage rate, the historical congestion frequency and the weather condition, and outputs the real-time risk level;

[0081] A plurality of sets of historical data are acquired, and each set of data is manually marked with a risk level, the risk level being high, medium or low;

[0082] High risk means that actual congestion occurs and needs manual intervention to dredge, which can be judged by the following conditions: the flow density > Pnum1 person / ㎡ and the average flow speed < Vnum1 m / s and the duration > Tnum minutes;

[0083] Medium risk means that the flow is close to the congestion threshold and has potential risk: the flow density ∈ [Pnum2, Pnum1] and the average flow speed ∈ [Vnum2, Vnum1] and the early warning system is triggered but there is no actual congestion;

[0084] Low risk means that the flow is normal and has no abnormal fluctuation: the flow density < Pnum2 person / ㎡ and the average flow speed > Vnum1 m / s;

[0085] Wherein, Pnum1 and Pnum2 are flow density thresholds, Vnum1 and Vnum2 are average flow speed thresholds, and Tnum is a duration threshold, all of which need to be set according to the actual situation of the campus;

[0086] The key features of each set of historical data are extracted, the key features including calibration number C, course table number M, time period type T, RFID coverage rate RC, historical congestion frequency H and weather condition W;

[0087] The specific reasons for selecting these key features are as follows:

[0088] C represents the target number of people output by the monitoring, course table and RFID calibration model, reflects the actual flow size, is directly related to the congestion possibility, and the more the number, the greater the space bearing pressure;

[0089] M represents the number of students required to flow between two places in a time period according to the course table, that is, the theoretical flow size, embodies the rigid demand, and the flow driven by the course table is inevitable, which has higher risk than random flow;

[0090] T can be divided into large break (20 minutes), small break (10 minutes), peak (the first and last 5 minutes of break), non-peak. The length of the period affects the degree of aggregation. Large breaks allow longer stays and are prone to local congestion;

[0091] RC represents the proportion of students with cards to the number of students on the schedule, reflecting data reliability. Low coverage means that a large number of target students are not identified by RFID and may flow through unmonitored paths, resulting in underestimation of actual traffic;

[0092] H represents the number of times a route has been labeled as high-risk or medium-risk in the past week. Frequent areas in the past may have design flaws, such as narrow stairs or narrow intersections, and need to be monitored to verify the effectiveness of improvements;

[0093] W represents sunny, rainy, and snowy, which will affect indoor and outdoor route selection. Rain and snow cause indoor routes to surge in traffic, while outdoor routes such as outdoor stairs see a decrease in usage;

[0094] Data cleaning and conversion are performed on the above extracted features:

[0095] For C and M, binning is used. For example, the original value range of C is [0, 300], and it is divided into 3 intervals. Low traffic, C < 80, is coded as 0, medium traffic, 80 <= C < 150, is coded as 1, and high traffic, C >= 150, is coded as 2;

[0096] Binning converts continuous numerical values into threshold intervals for safety, warning, and danger, which aligns with the experience of security personnel. After binning, all belong to medium or high traffic, reducing the sensitivity of the model to noise;

[0097] For T, large break and small break represent the duration of the break, reflecting the adequacy of time for students to freely move, affecting the size of the gathering. Peak and non-peak represent the peak of traffic during the break, capturing the instantaneous burst and stable state of traffic, affecting the probability of congestion;

[0098] Large break, small break, peak, and non-peak are different dimensional features that can be combined to describe the attributes of the period;

[0099] The specific operation is as follows:

[0100] Two-dimensional classification is combined into a four-tuple feature, period type = {time length, traffic stage}, time length value set = {large break, small break, non-break}, traffic stage value set = {end of class peak, non-peak, class preparation peak};

[0101] The time length and the flow stage are independently coded, the first two bits represent the time length, i.e. long break = 10, short break = 01, and non-break = 00, and the last three bits represent the flow stage, i.e. end-of-class peak = 100, non-peak = 010, and pre-class peak = 001;

[0102] The time length and the flow stage are combined to obtain a composite feature, and the composite feature is converted into a one-hot encoding vector, for example, {long break, end-of-class peak} = [1, 0, 1, 0, 0], {short break, non-peak} = [0, 1, 0, 1, 0], and {non-break, non-peak} = [0, 0, 0, 1, 0];

[0103] For W, one-hot encoding is adopted, rain can be encoded as [1, 0], snow can be encoded as [0, 1], and sunny can be encoded as [0, 0];

[0104] The decision tree cannot directly process text labels, and one-hot encoding is converted into a numerical vector, which can preserve the independence of categorical variables;

[0105] For RC, the min-max normalization method is adopted to normalize it to the interval [0, 1];

[0106] In order to unify the scale of other features and avoid model bias caused by the difference in numerical range;

[0107] For H, binary is adopted, for example, H≥2 times can be encoded as 1, and H<2 times can be encoded as 0;

[0108] When C and W have missing values, the following methods can be used to complete the missing values:

[0109] If C has missing values, the corresponding value of M is used to replace it, and if W has missing values, it is defaulted to sunny;

[0110] The specific method for calculating the composite rank to reflect the importance of the feature is:

[0111] Invite campus safety experts to score the importance of each feature, such as 1-5 points, and assign ranks in descending order according to the scores to obtain expert score rank R1;

[0112] Calculate the Spearman correlation coefficient of each feature and the risk level, and assign ranks in descending order according to the absolute values to obtain correlation coefficient rank R2;

[0113] According to the formula R = sqrt(R1 x R2), the composite rank is calculated;

[0114] Divide the training set and test set in the ratio of 8:2 to ensure that the proportion of each risk level in the training set is consistent with the overall proportion;

[0115] Initialize the parameters of the decision tree model:

[0116] The maximum number of layers is set to 5, and the logical hierarchy of campus safety data is usually clear, such as time period to regional traffic, and then to historical risk level, and 5 layers of depth can cover most of the feature combinations of the scene while maintaining model interpretability;

[0117] The minimum number of samples required for node splitting is 10, and in campus safety data, low sample splitting may introduce noise, such as abnormal traffic in a certain area, and setting it to 10 can ensure that the splitting is based on statistically significant rules;

[0118] The minimum number of samples required for leaf nodes is 5, and in the campus scene, if a leaf node has too few samples, its risk level judgment may lack reliability;

[0119] The Gini index is selected as the feature splitting criterion, and by calculating the drop in Gini index before and after splitting, the feature with the largest risk level discrimination is preferred, such as time period, which usually has higher discrimination than weather and other secondary factors;

[0120] Sort the feature list by composite rank;

[0121] Try to select features for splitting according to the sorted feature list, calculate the average Gini index drop value before and after splitting, if the drop value ≥ drop threshold, accept the split, otherwise try to select the next feature to split;

[0122] Repeat the above process for each sub-node, preferentially try high-rank features, until all features have been tried or the drop value meets the standard, and obtain the trained decision tree model;

[0123] For example, sort the feature list by composite rank to get: time period type (R=1) > calibration number (R=2) > course table number (R=3) > weather (R=4), try to split by time period type, calculate the average Gini index drop value before and after splitting, if the drop ≥ 0.15, then accept the split, otherwise try to split by calibration number;

[0124] Substitute the test set into the trained decision tree model to calculate the accuracy, recall rate, and F1 value, which are used to evaluate the final performance of the model.

[0125] The response analysis module takes appropriate measures according to the real-time risk level and analyzes the specific reasons;

[0126] If the decision tree model outputs a low risk, the strategy of regular round patrol combined with intelligent compression is adopted, and the specific measures are as follows:

[0127] Reduce the resolution of the monitoring video to 720P and extend the round patrol interval to 10 seconds;

[0128] The security patrol frequency is once every hour;

[0129] Data storage period is compressed to 7 days;

[0130] The low-risk area has stable traffic, such as ordinary corridors outside class breaks, and the probability of sudden security incidents is low. High-frequency monitoring consumes resources, so taking the above measures can save most of the monitoring computing resources, reduce the invalid labor of security personnel, and effectively reduce storage costs.

[0131] If the decision tree model output is medium risk, the strategy of focusing on dynamic optimization is adopted, and the specific measures are as follows:

[0132] The monitoring video resolution is adjusted to 1080P, and the patrol interval is shortened to 5 seconds;

[0133] The security patrol frequency is increased to once every 20 minutes;

[0134] The data storage period is extended to 30 days;

[0135] The medium-risk area has local congestion, such as stairways outside small class breaks, and needs to be moderately enhanced to prevent risk escalation. Taking the above measures can capture the trend of crowd gathering in time, provide early warning, and provide continuous flow data for later channel reconstruction;

[0136] If the decision tree model output is high risk, the strategy of real-time overview combined with three-dimensional prevention and control is adopted, and the specific measures are as follows:

[0137] The monitoring video resolution is adjusted to 4K, and real-time AI people counting and abnormal behavior recognition are enabled;

[0138] The security personnel arrive at the scene within 5 minutes and are stationed in real time;

[0139] The data storage period is permanently retained, and cloud backup is triggered;

[0140] The high-risk area has a high probability of congestion, such as the main entrance and exit during large class breaks, which may cause stampedes, conflicts, and other vicious incidents, and needs the highest level of resource protection. Taking AI abnormal behavior recognition can reduce the delay and discover pushing, retention, and other risk behaviors in advance compared to traditional monitoring. Real-time security personnel stationed can also effectively ensure the rapid unblocking of congestion and other risks;

[0141] After determining the risk level of a certain section through the decision tree model, the specific reasons for causing the risk need to be traced in detail, and the attribution results need to be weighted according to the composite priority of the features. The influence of high business value features is prioritized to form a hierarchical explanation of key reasons and secondary reasons, avoiding the traditional analysis of feature equal listing. The specific operation is as follows:

[0142] All split features from the root node to the target risk level leaf node are recorded in the order of triggering;

[0143] For the extracted features, screening is performed according to the composite rank R calculated in the training stage: the high-priority features in the top 20% of the composite rank are retained, and arranged in ascending order of the composite rank, and the smaller the rank, the higher the importance;

[0144] The attribution text is generated in the order of priority: [high-priority feature] leads to [low-priority feature], and finally triggers [risk level];

[0145] For example, the risk level output by the model is high, and through analysis, the path features are: period type = long break, corresponding composite rank R = 1; the number of people calibrated is greater than or equal to 200, corresponding composite rank R = 2; channel state = smooth, corresponding composite rank R = 4, which is excluded after screening; only the top 2 high-rank features are retained; the specific reasons are analyzed, the current is a long break period (key reason, accounting for 60%), and the number of people flowing is more than 200 (secondary reason, accounting for 40%), and the comprehensive judgment is high risk; the information automatically pushed to the security personnel is: the core reason for [high-risk warning] is the overstaffing of the long break period, and it is suggested to focus on guiding the stairway 1-3.

[0146] For example, Figure 2 A big data-based intelligent campus traffic management method, comprising:

[0147] S1, collecting RFID card reader data, monitoring video data and course table data, and generating calibrated people number through a course table-driven monitoring data calibration algorithm;

[0148] S2, constructing a risk classification model based on a decision tree and a composite rank splitting strategy, inputting real-time calibrated people number, course table people number, period type, RFID coverage rate, historical congestion frequency and weather condition, and outputting risk level;

[0149] S3, taking corresponding measures according to the real-time risk level, and analyzing the specific reasons for generating the risk.

[0150] Some of the data in the above formula are dimensionless numerical calculations, and the contents not described in detail in the specification all belong to the prior art known to those skilled in the art.

[0151] The above embodiments are only used to illustrate the technical method of the present application and are not limiting, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical method of the present application.

Claims

1. A smart campus traffic management system based on big data, characterized in that, include: The data acquisition module is used to collect RFID reader data, surveillance video data, and timetable data, and generates the number of calibrated students through a timetable-driven monitoring data calibration algorithm. The dynamic analysis module constructs a risk level model based on a decision tree and a composite rank splitting strategy. Inputs include real-time calibrated attendance, class schedule attendance, time period type, RFID coverage, historical congestion frequency, and weather conditions. Output is the risk level. The specific operations for constructing the risk level model are as follows: Acquire multiple sets of historical data and manually label the risk level of each set of data; Extract features from each set of historical data, including the number of calibrated participants (C), the number of participants on the timetable (M), the time period type (T), the RFID coverage rate (RC), the historical congestion frequency (H), and the weather conditions (W). The extracted features undergo data cleaning and transformation, specifically including: For C and M, binning is used; if C < C1 and M < M1, they are encoded as 0; if C1 ≤ C < C2 and M1 ≤ M < M2, they are encoded as 1; if C ≥ C2 and M ≥ M2, they are encoded as 2. For T, the time length and traffic phase are combined to obtain composite features, and the composite features are converted into one-hot encoded vectors. Specifically, the time length and traffic phase are encoded independently, with the first two digits representing the time length (e.g., long break = 10, short break = 01, non-break = 00) and the last three digits representing the traffic phase (e.g., peak at the end of class = 100, non-peak = 010, peak before class = 001). The time length and traffic phase are combined to obtain composite features, and the composite features are converted into one-hot encoded vectors, specifically {long break, peak at the end of class} = [1, 0, 1, 0, 0], ... {Peak period} = [1,0,0,1,0], {Long break, pre-class peak} = [1,0,0,0,1], {Short break, off-peak} = [0,1,0,1,0], {Short break, end-of-class peak} = [0,1,1,0,0], {Short break, pre-class peak} = [0,1,0,0,1], {Non-break, off-peak} = [0,0,0,1,0], {Non-break, end-of-class peak} = [0,0,1,0,0], {Peak period before class} = [0,0,0,0,1]; For W, one-hot encoding is used, rain is encoded as [1,0], snow is encoded as [0,1], and clear weather is encoded as [0,0]; For RC, the min-max normalization method is used to normalize it to the interval [0,1]; For H, binarization is used, if H≥H1, it is encoded as 1, if H<H1, it is encoded as 0; Calculate the composite rank of each feature and sort the feature list by composite rank; The training and test sets are divided in an 8:2 ratio, and the parameters of the decision tree model are initialized. Select a feature from the feature list to split, calculate the average Gini index decrease before and after splitting, and if the decrease is greater than or equal to the decrease threshold, accept the split; otherwise, select the next feature to continue splitting. Repeat the above process for each child node, prioritizing high-rank features, until all features have been tried or the average Gini index decreases to the target value, and obtain the trained decision tree model. The test set is fed into the trained decision tree model to calculate the accuracy, recall, and F1 score, and to evaluate the final performance of the model. The response analysis module takes appropriate measures based on the real-time risk level and analyzes the specific reasons for the risk.

2. The smart campus traffic management system based on big data according to claim 1, characterized in that, According to the formula Calculate the number of students calibrated, where R represents the number of students with cards, M is the number of students on the class schedule, and K is the total number of students captured by the surveillance cameras. For cardless rate, To monitor the student recognition rate.

3. The smart campus traffic management system based on big data according to claim 1, characterized in that, The specific method for calculating the composite rank of each feature is as follows: Campus safety experts were invited to rate the importance of each feature, and the scores were assigned ranks in descending order to obtain the expert rating rank R1. Calculate the Spearman correlation coefficient between each feature and the risk level, assign ranks to them in descending order of absolute value, and obtain the rank R2 of the correlation coefficient. The composite rank is calculated using the formula R = sqrt(R1 × R2).

4. The smart campus traffic management system based on big data according to claim 1, characterized in that, The parameters for initializing the decision tree model specifically include: The maximum number of layers is 5; The minimum number of samples required for node splitting is 10; The minimum number of samples required for a leaf node is 5; The Gini index decrease value was selected as the characteristic splitting criterion.

5. The smart campus traffic management system based on big data according to claim 1, characterized in that, The specific steps for taking corresponding measures based on the real-time risk level are as follows: If the decision tree model outputs low risk, a strategy combining conventional round-robin and intelligent compression is adopted; if the decision tree model outputs medium risk, a strategy combining focused attention and dynamic optimization is adopted. If the decision tree model outputs a high-risk signal, a strategy combining real-time overview and multi-dimensional prevention and control will be adopted. The strategy of combining conventional patrol with intelligent compression includes reducing the resolution of surveillance video to 720P, extending the patrol interval to 10 seconds, increasing the security patrol frequency to once per hour, and compressing the data storage period to 7 days. The key focus and dynamic optimization strategy includes adjusting the surveillance video resolution to 1080P, shortening the patrol interval to 5 seconds, and increasing the security patrol frequency to once every 20 minutes. The strategy of combining real-time overview with three-dimensional prevention and control includes adjusting the resolution of surveillance video to 4K, enabling real-time AI people counting and abnormal behavior recognition; security personnel arriving on-site within 5 minutes and stationing themselves there in real time; and data storage being permanently retained and triggering cloud backup.

6. The smart campus traffic management system based on big data according to claim 1, characterized in that, The steps to analyze the causes of risks are as follows: All splitting features from the root node to the leaf node of the target risk level are recorded in the order of triggering. The extracted features are filtered based on the composite rank R calculated during the training phase: the top 20% of high-priority features with the highest composite rank are retained and arranged in ascending order of composite rank. Attribution texts are generated in priority order: high-priority features lead to lower-priority features, which ultimately trigger the risk level.

7. A smart campus traffic management method based on big data, used to execute the smart campus traffic management system based on big data as described in claims 1-6, characterized in that, include: S1. Collect RFID reader data, surveillance video data, and timetable data, and generate the number of calibrated individuals through a timetable-driven surveillance data calibration algorithm; S2. Construct a risk level model based on the splitting strategy of decision tree and composite rank. Input the real-time number of calibrated students, the number of students on the timetable, the time period type, RFID coverage, historical congestion frequency, and weather conditions, and output the risk level. S3. Take appropriate measures based on the real-time risk level and analyze the specific reasons for the risk.

Images