Campus security abnormal behavior intelligent identification method and system based on machine vision

Abstract

The invention provides an intelligent campus security abnormal behavior identification method and system based on machine vision, and belongs to the technical field of campus security protection, and the method comprises the steps: adjusting a sensitivity parameter and a confidence threshold value of a preset visual behavior identification model based on campus schedule context information; inputting the video data of the monitoring area into the visual behavior recognition model to obtain an abnormal behavior recognition result and confidence; judging whether the confidence is smaller than a confidence threshold and whether the abnormal behavior recognition result is a high-risk behavior; if any condition is met, auxiliary data are acquired, and auxiliary judgment information is generated; inputting the abnormal behavior recognition result, the confidence coefficient and the auxiliary judgment information into a preset decision model, and generating a target recognition result and a danger level; if any condition is not met, determining a danger level by querying a preset behavior type danger level mapping table, and taking an abnormal behavior recognition result as a target recognition result; and determining a corresponding target early warning strategy based on the danger level and the campus schedule context information. According to the invention, the accuracy and response efficiency of safety early warning can be improved.

Description

Technical Field

[0001] This application relates to the field of campus security protection technology, and in particular to a method and system for intelligent identification of abnormal campus security behaviors based on machine vision. Background Technology

[0002] Currently, campus monitoring systems mainly rely on real-time manual monitoring and post-event review, which requires a high investment of manpower and also suffers from inherent problems such as blind spots in monitoring coverage and delayed response to abnormal behavior, making it difficult to meet the actual needs of campuses for 24 / 7, high-precision security.

[0003] To improve management efficiency, some campuses have introduced machine vision-based behavior recognition technology, but the application of this technology still has significant shortcomings. On the one hand, fixed sensitivity parameters and confidence thresholds are used without dynamic adjustment based on campus schedule context information; on the other hand, many models rely solely on video data for single judgments, lacking fusion analysis of auxiliary data such as environment, time, and personnel information, resulting in low accuracy in abnormal behavior recognition, inability to filter invalid alerts, and the possibility of missing high-risk behaviors.

[0004] Therefore, there is an urgent need for a machine vision-based intelligent identification method and system for abnormal campus security behavior to improve the intelligence level and response efficiency of campus security monitoring. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a machine vision-based intelligent identification method and system for abnormal campus security behaviors.

[0006] A first aspect of this application provides a machine vision-based intelligent identification method for abnormal campus security behaviors, including: Based on the contextual information of the campus schedule, the sensitivity parameters and confidence threshold of the preset visual behavior recognition model are adjusted to obtain the target sensitivity parameters and target confidence threshold. The video data of the campus surveillance area during the current time period is input into the visual behavior recognition model for analysis to obtain the first abnormal behavior recognition result and the corresponding first confidence level. Determine whether the first confidence level is less than the target confidence threshold, and whether the first abnormal behavior identification result is a predefined high-risk behavior; If any condition is met, then auxiliary data associated with the video data is obtained, and auxiliary judgment information is generated based on the auxiliary data; The first abnormal behavior identification result, the first confidence level, and the auxiliary judgment information are input into a preset decision model for comprehensive judgment to generate target identification result and danger level; If any condition is not met, the danger level is determined by querying a preset behavior type danger level mapping table based on the first abnormal behavior identification result and the first confidence level, and the first abnormal behavior identification result is used as the target identification result. Based on the hazard level and the campus schedule context information, a corresponding target early warning strategy is determined.

[0007] A second aspect of this application provides a machine vision-based intelligent recognition system for abnormal campus security behaviors, comprising: The target parameter adjustment module is used to adjust the sensitivity parameters and confidence threshold of the preset visual behavior recognition model based on the campus schedule context information, so as to obtain the target sensitivity parameters and target confidence threshold. The surveillance video analysis module is used to input video data of the campus surveillance area during the current time period into the visual behavior recognition model for analysis, and obtain the first abnormal behavior recognition result and the corresponding first confidence level. The dual-condition judgment module is used to determine whether the first confidence level is less than the target confidence threshold and whether the first abnormal behavior identification result is a predefined high-risk behavior. The auxiliary information generation module is used to obtain auxiliary data associated with the video data if any condition is met, and generate auxiliary judgment information based on the auxiliary data. The target result generation module is used to input the first abnormal behavior identification result, the first confidence level, and the auxiliary judgment information into a preset decision model for comprehensive judgment, and generate a target identification result and a danger level; if any condition is not met, the danger level is determined by querying a preset behavior type danger level mapping table based on the first abnormal behavior identification result and the first confidence level, and the first abnormal behavior identification result is used as the target identification result. The target strategy determination module is used to determine the corresponding target early warning strategy based on the hazard level and the campus schedule context information.

[0008] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described intelligent recognition method for abnormal campus security behavior based on machine vision.

[0009] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described intelligent identification method for abnormal campus security behaviors based on machine vision.

[0010] The beneficial effects of the intelligent identification method and system for abnormal campus security behavior based on machine vision provided in this application are as follows: This application intelligently adjusts the recognition threshold of the visual model based on the campus schedule context, matching it with the risk level of different time periods and scenarios, thereby effectively reducing the false alarm rate in routine scenarios; when insufficient confidence or high-risk behavior occurs, cross-validation and weighted fusion are performed using auxiliary data from audio and environmental sensors to overcome the limitations of simple visual analysis in situations of occlusion or poor lighting. Simultaneously, for high-confidence routine abnormal behavior, a mapping table can be quickly queried for immediate response, while for complex or high-risk events, a multimodal fusion decision-making process is initiated to ensure more accurate judgment. Differentiated early warning strategies are generated based on the danger level and specific scenario, achieving precise scheduling and rapid intervention of security resources, and improving the accuracy of campus security early warnings and the efficiency of emergency response. Attached Figure Description

[0011] Figure 1 A flowchart illustrating a machine vision-based intelligent identification method for abnormal campus security behavior provided in an embodiment of this application; Figure 2 A structural block diagram of a machine vision-based intelligent recognition system for abnormal campus security behavior provided in an embodiment of this application; Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0012] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0013] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1-3 The following is an explanation using specific examples.

[0014] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a machine vision-based intelligent recognition method for abnormal campus security behavior, provided in an embodiment of this application. The method includes: S101: Adjust the sensitivity parameters and confidence threshold of the preset visual behavior recognition model based on the campus schedule context information to obtain the target sensitivity parameters and target confidence threshold; In this embodiment, the campus schedule context information is a comprehensive set of information, including: current time period attributes, such as class time, lunch break, evening self-study, and holidays; monitoring area function types, such as teaching building corridors, playground, library reading area, and school gate; and calendar events, such as final exam week, spring sports meet, fire drill, and open day, which are divided into first, second, and third types of risk events. The parameter adjustment process is a multi-level pipeline. First, based on the combination of break time and playground, a preset first mapping table is queried to obtain the basic sensitivity parameters and basic confidence thresholds. The logic for obtaining the basic sensitivity parameters and basic confidence thresholds is as follows: During break time, students concentrate on activities on the playground, with high frequency and large movements such as rope skipping, chasing, and ball games. Compared to class time, the risk of accidents such as collisions and falls among students increases significantly. Therefore, it is necessary to increase the basic sensitivity parameters while appropriately lowering the basic confidence threshold to avoid missing potential risks due to excessively high thresholds, thus achieving comprehensive coverage of safety risks in the break time playground scenario.

[0015] Secondly, based on the detected current calendar event, for example, if the current calendar event is final exam week, this event is classified as a first-risk event type, as students are under great pressure and unexpected events such as disputes may occur. Therefore, a first adjustment factor is used to weight and correct the basic parameters to obtain the target sensitivity parameter and the target confidence threshold.

[0016] This embodiment analyzes feedback data from a past period, specifically the historical recognition results of the teaching building corridor during a certain time period. Feedback statistics show a false alarm rate of up to 15% for misidentifying brisk walking as running, exceeding the first preset threshold of 10%, while the false negative rate is only 1%, less than the second preset threshold of 5%. This indicates that the visual behavior recognition model is overly sensitive. Therefore, it is necessary to fine-tune the corrected parameters based on a third adjustment factor, ultimately targeting the sensitivity parameter. Through this series of threshold adjustments, a precise match between the parameters and the specific scenario's risk level and historical performance is achieved, providing a sensitive yet reliable benchmark for subsequent recognition. The first and second adjustment factors are determined by analyzing the correlation between behavioral data and security events during historical periods; the third and fourth adjustment factors are determined through regression analysis based on parameter adjustment records in historical data and corresponding changes in false alarm and false negative rates.

[0017] S102: Input the video data of the campus monitoring area during the current time period into the visual behavior recognition model for analysis, and obtain the first abnormal behavior recognition result and the corresponding first confidence level; In this embodiment, the behavior recognition model is trained using a large-scale campus security dataset. Its backbone network employs a 3D convolutional neural network or Transformer architecture, enabling it to analyze the spatiotemporal features of video sequences. The behavior recognition model takes real-time video stream segments of the campus playground area, current time period attributes, and the functional type of the monitored area as input. During the recognition process, the model first performs frame-by-frame analysis and temporal correlation processing on the video segments, extracting core features such as human key points, movement trajectories, and group interactions from the images. Then, based on these features, it determines abnormal behavior.

[0018] The model outputs two core results: the first abnormal behavior identification result and the first confidence score. If the identification result shows multiple people gathering with pushing or shoving, this behavior belongs to a predefined potential conflict behavior. For example, the first confidence score for this behavior is 0.72. This value integrates factors such as target detection accuracy, behavior classification clarity, and temporal consistency, representing the model's degree of confidence in its current judgment. If the behavior recognition model can clearly capture the complete bounding boxes of three or more people, and the sequence of human key points shows obvious unfriendly physical contact, and this action pattern persists for multiple frames, the confidence score will be high. Conversely, if there is severe occlusion of people in the image, or the action is blurry, the confidence score will decrease. The aforementioned confidence score of 0.72 indicates that the model determines the event is likely abnormal behavior, but due to factors such as shooting angle, lighting conditions, or partial occlusion, there is still some uncertainty. The first confidence score will serve as a key trigger for subsequent decision-making processes, directly affecting whether auxiliary data is needed for further verification.

[0019] S103: Determine whether the first confidence level is less than the target confidence level threshold, and whether the first abnormal behavior identification result is a predefined high-risk behavior; In this embodiment, the results output by the behavior recognition model are judged, specifically by using dual-condition constraints, including judging whether the first confidence level output by the behavior recognition model is less than the target confidence level threshold; and judging whether the first abnormal behavior recognition result output by the behavior recognition model is a predefined high-risk behavior.

[0020] The judgment process specifically includes comparing the first confidence level output by the behavior recognition model with the target confidence threshold, and determining whether the first confidence level is less than the target confidence threshold. Simultaneously, it determines whether the first abnormal behavior identification result output by the behavior recognition model is a predefined high-risk behavior. Specifically, the pre-constructed list of high-risk behaviors includes behavior types that could lead to serious safety consequences in a campus setting, such as multiple-person pushing and shoving conflicts, throwing objects from heights, running and colliding with crowds, and climbing over walls. This judgment step does not rely on probability values ​​but directly defines whether the currently identified behavior belongs to the high-risk category requiring priority response by precisely matching the behavior type in the list, ensuring priority control of major safety hazards.

[0021] These two judgments together constitute the system's dual screening mechanism. If either condition is met, the subsequent process is triggered. The first confidence level is compared with the target confidence level threshold. If the first confidence level is less than the target confidence level threshold, it indicates that the identification result is unreliable, or the identification result belongs to high-risk behavior, indicating a high risk level. If either condition is met, the auxiliary data verification process is initiated. If neither condition is met, that is, the identification result has high credibility but does not belong to high-risk behavior, it directly enters the regular risk level matching stage.

[0022] S104: If any condition is met, obtain the auxiliary data associated with the video data, and generate auxiliary judgment information based on the auxiliary data; In this embodiment, the auxiliary process is triggered if either the first confidence level is less than the target confidence level threshold or the first abnormal behavior identification result is a predefined high-risk behavior. For example, in a certain scenario, if the first confidence level of the behavior identification result is 0.60, which is less than the target threshold of 0.65, the auxiliary process is triggered. Then, the collection of auxiliary data is initiated. The acquired auxiliary data includes audio data: an audio stream synchronously collected from the built-in microphone of the monitoring camera; and environmental sensor data: data from IoT nodes near the playground, including a heat map of personnel density, ambient noise levels in decibels, and real-time light intensity.

[0023] Secondly, the acquired auxiliary data underwent data quality assessment and preprocessing. Specifically, for the audio data, an acoustic quality assessment was performed, and its signal-to-noise ratio was calculated. Background wind noise interference was found, but the human voice frequency range was clear, and the data was assessed as usable. Subsequently, a noise reduction filter was used for preprocessing to obtain the first effective data, such as purified audio signals. For the environmental sensor data, its completeness and rationality were checked. All data fields were found to be valid, with high personnel density and a noise level of 75 dB, which is higher than that of the library area. This data was assessed as high quality and used as the second effective data.

[0024] The first valid data is used to extract features to obtain time-frequency domain feature vectors, which are then input into a pre-trained sound event detection model. After analysis, the behavior recognition model outputs the recognition probability of abnormal sounds such as loud shouting and arguing. The obtained recognition probabilities are then weighted and fused to finally obtain the comprehensive audio anomaly probability.

[0025] For the second set of valid data, personnel density, noise level, and light intensity are fused to form a multi-dimensional feature vector, which is then input into an anomaly detection model based on the isolated forest algorithm to obtain the environmental risk index of the current environment. The higher the environmental risk index value, the more abnormal the environment.

[0026] Finally, weighted fusion is performed to generate auxiliary judgment information. Specifically, based on the weighting strategy table, the benchmark audio weighting coefficient and the benchmark environment weighting coefficient are determined. Based on the determined benchmark audio weighting coefficient and benchmark environment weighting coefficient, the audio anomaly probability and environmental risk index are weighted and fused to generate auxiliary judgment information. S105: Input the first abnormal behavior identification result, the first confidence level, and the auxiliary judgment information into the preset decision model for comprehensive judgment, and generate the target identification result and the danger level; if any condition is not met, based on the first abnormal behavior identification result and the first confidence level, determine the danger level by querying the preset behavior type danger level mapping table, and use the first abnormal behavior identification result as the target identification result; In this embodiment, if the first confidence level is less than the target threshold or the identification result is a high-risk behavior, the comprehensive judgment process of the decision model is initiated. Specifically, this includes: firstly, inputting the first abnormal behavior identification result, the first confidence level, and auxiliary judgment information into a preset decision model. This model evaluates the reliability of the current information by analyzing the deviation between the first confidence level and the auxiliary judgment information, and obtains an uncertainty index. Secondly, a target decision path is selected based on the uncertainty index and the behavior type. For example, when the uncertainty index is less than 0.3 and the behavior is a potential conflict, a path with weighted enhancement of auxiliary information is adopted, increasing the weight of the auxiliary judgment information to 0.6, and merging it with the first confidence level to calculate the comprehensive credibility. Finally, based on the comprehensive credibility and the behavior risk benchmark, the target identification result and the corresponding danger level are generated. If the first confidence level is greater than the target threshold and the identification result is not a high-risk behavior, then the fast matching process is directly entered: the first abnormal behavior identification result and the first confidence level are matched with the preset behavior type danger level mapping table. For example, the risk level corresponding to a person running fast on the playground or during lunch break is determined. Since the confidence level is greater than 0.8, no additional verification is required. The person running fast is directly taken as the target identification result, and the danger level is determined to be low risk level.

[0027] S106: Determine the corresponding target early warning strategy based on the hazard level and campus schedule context information.

[0028] In this embodiment, the determination of the target early warning strategy is a multi-level decision-making process that comprehensively considers the severity of the danger level and the specificity of the current campus schedule context information.

[0029] This embodiment pre-sets a four-level early warning mechanism, each corresponding to different response procedures and handling methods. Initial strategy matching is performed based on the danger level output by the decision-making model. If the danger level is Level 1, such as the detection of armed fighting or large-scale conflict, a red alert strategy is immediately triggered, sending a real-time alarm signal to the campus security center, automatically activating the emergency broadcast system to provide on-site evacuation instructions, and simultaneously pushing the event information to the local police station's linkage platform. If the danger level is Level 2, such as the detection of heated arguments or horseplay, an orange alert strategy is activated, sending handling instructions to security personnel in the area and simultaneously notifying the class teacher or counselor to go to the scene. If the danger level is Level 3, such as the detection of a single person climbing over a wall, a yellow alert strategy is adopted: recording the behavior trajectory and notifying the nearest patrol post to verify.

[0030] Secondly, the warning strategy is adjusted based on the context of the campus schedule. For example, when events of the same risk level are detected: during final exam week, the system will shorten the warning response time by 30% and add a notification linking to the psychological counselor; during large-scale events such as the spring sports meet, the system will automatically lower the warning level for identifying gathering behavior in the playground area to avoid over-responding; and during nighttime self-study hours, the system will raise the warning level for detecting abnormal sounds in the teaching building area and prioritize non-intrusive methods such as light warnings.

[0031] This dual decision-making mechanism, based on both hazard level and contextual information, enables rapid response to real security threats and avoids unnecessary interference in special campus activity scenarios, achieving precise and intelligent security early warning management.

[0032] As can be seen from the above, this application intelligently adjusts the recognition threshold of the visual model based on the campus schedule context, matching it with the risk level of different time periods and scenarios, thereby effectively reducing the false alarm rate in routine scenarios. When insufficient confidence or high-risk behavior occurs, cross-validation and weighted fusion are performed using auxiliary data from audio and environmental sensors to overcome the limitations of pure visual analysis in situations of occlusion or poor lighting. Simultaneously, for high-confidence routine abnormal behaviors, a mapping table can be quickly queried for immediate response, while for complex or high-risk events, a multimodal fusion decision-making process is initiated to ensure more accurate judgment. Differentiated early warning strategies are generated based on the level of danger and specific scenarios, achieving precise scheduling and rapid intervention of security resources, and improving the accuracy of campus security early warnings and the efficiency of emergency response.

[0033] In one embodiment of this application, adjusting the sensitivity parameters and confidence threshold of a preset visual behavior recognition model based on campus schedule context information to obtain target sensitivity parameters and target confidence threshold includes: Analyze the campus schedule context information to determine the current time period attributes, monitoring area function type, and calendar events; Based on the current time period attributes and the functional type of the monitored area, query the preset first mapping table to obtain the basic sensitivity parameters and basic confidence threshold. Based on calendar event types, the basic sensitivity parameters and basic confidence thresholds are weighted and corrected to obtain the corrected sensitivity parameters and confidence thresholds. Based on the feedback statistics of historical identification results within a preset period, the corrected sensitivity parameters and confidence thresholds are adjusted to obtain the target sensitivity parameters and target confidence thresholds.

[0034] In this embodiment, by parsing the campus context information, the current time period attributes, the monitoring area function type, and calendar events are determined. The current time period attributes include time characteristics such as class time, break time, lunch break, evening self-study, and nighttime. The monitoring area function types include different functional areas such as teaching buildings, playgrounds, laboratories, dormitories, and school gates. The calendar event types include a first risk event type, a second risk event type, and a third risk event type, with the first risk event type having a higher risk level than the second risk event type, and the second risk event type having a higher risk level than the third risk event type. Based on the current time period attributes and the functional type of the monitored area, the preset first mapping table is queried to obtain the basic sensitivity parameters and basic confidence thresholds. The first mapping table represents the parameter baseline values ​​corresponding to different time period attributes and area functional types. For example, since the activity frequency and accident risk frequency of the playground during recess are significantly higher than those during class time in the teaching building, the basic sensitivity parameters set in the system configuration are also higher than those during class time in the teaching building; the basic confidence threshold for the laboratory during experimental classes is lower than that during non-experimental periods, as shown in Table 1.

[0035] Table 1 First Mapping Table

[0036] Calendar events, as special scenario factors, are adjusted using a weighted approach to optimize the baseline sensitivity parameters and baseline confidence thresholds. For high-risk calendar events, a first adjustment factor is used to increase the baseline sensitivity parameter and decrease the baseline confidence threshold to enhance recognition sensitivity. For low-risk calendar events, a second adjustment factor is used to decrease the baseline sensitivity parameter and increase the baseline confidence threshold to reduce false alarms. For medium-risk calendar events, the baseline parameters remain unchanged, maintaining the original settings. Specifically, taking the adjustment of campus monitoring parameters at a middle school on Thursday as an example, 2 PM is a class time. Initially, the laboratory area is set with a baseline sensitivity parameter of 0.9 and a baseline confidence threshold of 0.65 according to the first mapping table. However, due to a high-risk calendar event—a chemistry experiment competition to be held in the laboratory that afternoon—the first adjustment factor is used. In this first adjustment factor, the sensitivity coefficient is 1.2 and the threshold coefficient is 0.8, increasing the baseline sensitivity parameter to 1.08 and decreasing the baseline confidence threshold to 0.52. This ensures more sensitive detection of abnormal behaviors such as violations and reagent spills, avoiding missed detections. If the day is a low-risk calendar event (a weekend), the teaching building corridor should, according to the first mapping table, have a basic sensitivity parameter of 0.8 and a basic confidence threshold of 0.75 during breaks. In this case, the second adjustment factor is used, where the sensitivity coefficient is 0.7 and the threshold coefficient is 1.1. This lowers the sensitivity parameter to 0.56 and raises the confidence threshold to 0.825, preventing the misjudgment of normal behaviors such as students taking books or cleaning as abnormal. However, if the day is a medium-risk calendar event (no special events), such as Monday morning class time, the classrooms in the teaching building directly use the basic parameters of the first mapping table, i.e., a sensitivity of 0.6 and a confidence threshold of 0.85, without any additional adjustments.

[0037] By analyzing feedback data within a preset period, including false alarm rate and false negative rate, the corrected sensitivity parameters and confidence thresholds are adjusted to obtain the target sensitivity parameters and target confidence thresholds.

[0038] As can be seen from the above, this embodiment determines key information by parsing the campus schedule context, and then adjusts parameters based on the first mapping table, calendar event weighting correction, and historical feedback statistics. Through precise adjustments at multiple stages, the sensitivity parameters and confidence thresholds of the visual behavior recognition model can dynamically adapt to different time periods, areas, and event scenarios on campus, reducing false alarms or missed alarms caused by fixed parameters, and further improving the model's adaptability and recognition reliability in the complex campus environment.

[0039] In one embodiment of this application, the calendar event types include: a first risk event type, a second risk event type, and a third risk event type, wherein the risk level of the first risk event type is greater than that of the second risk event type, and the risk level of the second risk event type is greater than that of the third risk event type; Based on calendar event types, the basic sensitivity parameters and basic confidence thresholds are weighted and adjusted to obtain the adjusted sensitivity parameters and confidence thresholds, including: For the first risk event type, the basic sensitivity parameter is increased based on the first adjustment factor and the basic confidence threshold is decreased based on the first adjustment factor. For the second risk event type, the basic sensitivity parameter and basic confidence threshold remain unchanged; For the third risk event type, the basic sensitivity parameter is reduced based on the second adjustment factor, and the basic confidence threshold is increased based on the second adjustment factor.

[0040] In this embodiment, the risk levels of calendar events are classified according to various calendar events on campus, resulting in a clear definition of the specific scope of the first, second, and third risk event types. For example, large-scale sports meets, campus open days, and other activities involving large numbers of people are classified as first-risk event types; daily classroom teaching and regular club activities are classified as second-risk event types; and the campus vacancy periods during winter and summer vacations and nighttime rest periods are classified as third-risk event types.

[0041] Secondly, for the first risk event type, a specific value for the first adjustment factor is determined (e.g., 1.2). For example, the specific value of the first adjustment factor is determined based on the characteristics of the first risk event type, such as the number of participants, the intensity of the activity, and the number of potential hazards. Specifically, for large gatherings with more than 500 participants, the first adjustment factor is set to 1.3, based on safety records and risk assessment reports of similar historical events; for activities involving dangerous experimental operations, the first adjustment factor is set to 1.5 based on the experimental risk level assessment. After determining the factor, the increased sensitivity parameter is obtained by multiplying the basic sensitivity parameter by the first adjustment factor, and the decreased confidence threshold is obtained by dividing the basic confidence threshold by the first adjustment factor. This enhances the model's ability to capture abnormal behavior and ensures more sensitive safety monitoring in high-risk scenarios.

[0042] The second risk event type is at a normal risk level. The basic sensitivity parameters and basic confidence thresholds are sufficient to meet the normal needs of abnormal behavior identification. Keeping the parameters stable can avoid unnecessary interference. Therefore, there is no need to adjust the parameters. We can directly use the basic sensitivity parameters and basic confidence thresholds obtained by querying the first mapping table based on the current time period attributes and the function type of the monitoring area.

[0043] This embodiment comprehensively analyzes the characteristics of the second risk event type to determine the specific value of the second adjustment factor. This includes personnel density, activity intensity, and historical security records during this type of event, while also considering the allocation of campus security resources during this period. The specific value of the second adjustment factor is determined by quantifying the impact of these factors on monitoring and identification accuracy. For example, it can be set to 0.7 during summer and winter vacations when the campus is vacant, and to 0.8 during nighttime rest periods. After determining the second adjustment factor, the reduced sensitivity parameter is obtained by multiplying the basic sensitivity parameter by the second adjustment factor, and the increased confidence threshold is obtained by dividing the basic confidence threshold by the second adjustment factor. This reduces the model's sensitivity, decreases false alarms in low-risk scenarios, and ensures the efficient operation of the monitoring system.

[0044] As can be seen from the above, this embodiment classifies calendar events according to risk levels and adopts differentiated parameter adjustment strategies for different types. Increasing model sensitivity and lowering the confidence threshold for high-risk events enables more accurate anomaly detection; decreasing sensitivity and increasing the confidence threshold for low-risk events reduces unnecessary warnings. This precise parameter differentiation allows the model to flexibly adapt to event risk, ensuring effective identification in high-risk scenarios while avoiding resource waste in low-risk scenarios.

[0045] In one embodiment of this application, the feedback statistics include false alarm rate and false negative rate; Based on the feedback statistics of historical identification results within a preset period, the corrected sensitivity parameters and confidence thresholds are adjusted to obtain the target sensitivity parameters and target confidence thresholds, including: If the false alarm rate is greater than the first preset threshold and the false negative rate is less than or equal to the second preset threshold, then the corrected confidence threshold is increased based on the third adjustment factor or the corrected sensitivity parameter is decreased based on the third adjustment factor to obtain the target sensitivity parameter and the target confidence threshold. If the false negative rate is greater than the second preset threshold and the false positive rate is less than or equal to the first preset threshold, then the corrected confidence threshold is lowered based on the fourth adjustment factor or the corrected sensitivity parameter is increased based on the fourth adjustment factor to obtain the target sensitivity parameter and the target confidence threshold.

[0046] In this embodiment, a feedback statistics period is preset, for example, the feedback statistics period is set to one month, and the historical recognition results of the visual behavior recognition model within one month are collected. Then, feedback statistics of false alarm rate and false negative rate are extracted from the historical recognition results. Based on the feedback statistics, a first preset threshold for the upper limit standard of false alarm rate and a second preset threshold for the upper limit standard of false negative rate are determined, providing a basis for judgment for subsequent parameter adjustment.

[0047] In this embodiment, when the false positive rate is greater than the first preset threshold and the false negative rate is less than or equal to the second preset threshold, it indicates that the model's current judgment of abnormal behavior is too lenient, leading to too many unnecessary alarms. In this case, the parameters are adjusted based on the third adjustment factor. Either the corrected confidence threshold is multiplied by the third adjustment factor to increase the confidence threshold, or the corrected sensitivity parameter is divided by the third adjustment factor to decrease the sensitivity. Either method yields the final target sensitivity parameter and target confidence threshold, thereby reducing false positives. When the false negative rate is greater than the second preset threshold and the false positive rate is less than or equal to the first preset threshold, it indicates that the model's current judgment of abnormal behavior is too strict, resulting in some genuine anomalies not being identified. In this case, the parameters are adjusted using the fourth adjustment factor. Either the corrected confidence threshold is divided by the fourth adjustment factor to decrease the threshold, or the corrected sensitivity parameter is multiplied by the fourth adjustment factor to increase the sensitivity. Either method yields the target sensitivity parameter and target confidence threshold, thereby reducing the false negative rate and improving the model's ability to detect abnormal behavior.

[0048] For example, the campus sets a first preset threshold of 5%, a second preset threshold of 2%, a third adjustment factor of 1.1, and a fourth adjustment factor of 1.2. A preset period is one month. During this month, the model's nighttime recognition feedback in the dormitory area shows a false alarm rate of 7% (greater than 5%) and a false negative rate of 1.5% (less than 2%). In this case, one of two operations can be chosen: first, multiply the corrected confidence threshold for this scenario by 1.1 to increase it to 0.66; second, divide the corrected sensitivity parameter by 1.1 to decrease it to 0.73. Either operation yields the target parameter to reduce false alarms. If, in another period, the model's feedback during the sports meet shows that the false negative rate is 3% and greater than 2%, and the false positive rate is 3% and less than 5%, then one of two operations can be chosen: First, divide the original 0.4 confidence threshold by 1.2 to reduce it to 0.33; second, multiply the original 1.2 sensitivity parameter by 1.2 to increase it to 1.44. Either operation can be used to obtain the target parameter to reduce false negatives and ensure that the recognition accuracy adapts to the needs of different scenarios.

[0049] As can be seen from the above, this embodiment adjusts parameters based on different false positive and false negative rates. When the false positive rate is too high, the confidence threshold is increased or the sensitivity is decreased to reduce invalid warnings; when the false negative rate is too high, the confidence threshold is decreased or the sensitivity is increased to avoid missing anomalies. By dynamically optimizing parameters in conjunction with historical feedback statistics, the model performance can be continuously improved, enabling the model to maintain a low false positive and false negative rate during long-term use on campus, thereby enhancing recognition stability.

[0050] In one embodiment of this application, auxiliary judgment information is generated based on auxiliary data, including: Based on the data type of auxiliary data, data quality assessment and preprocessing are performed to obtain effective data from multiple sources. Determine the target features corresponding to the multi-source effective data based on the data type of the multi-source effective data; Auxiliary judgment information is determined based on the target features corresponding to multi-source effective data.

[0051] In this embodiment, the auxiliary data types include audio data and environmental sensor data. For audio data, acoustic quality assessment methods are used to detect issues such as noise interference, abnormal volume, and missing data. Preprocessing techniques such as noise reduction, redundancy removal, and missing segment completion are then employed to optimize the data. For environmental sensor data, numerical data assessment methods are used to check whether the data exceeds a reasonable range, whether there are abnormal fluctuations, or acquisition errors. Furthermore, outlier removal, data smoothing, and standardization are used to purify the data, resulting in multi-source valid data. The environmental sensor data includes temperature, humidity, light intensity, and personnel density data; the multi-source valid data consists of the processed audio data and the environmental sensor data.

[0052] This embodiment first performs specialized processing based on the data type: for audio data, it analyzes its deep features to construct feature vectors that can characterize sound patterns; for environmental sensor data, it filters and integrates its core indicators to construct multi-dimensional feature vectors representing the overall environmental state. Next, these feature vectors are input into corresponding analysis models, which perform in-depth interpretation and risk assessment to obtain output results. The risk quantification results with clear semantics in the output results are summarized and integrated to collectively constitute auxiliary judgment information used to assist in behavioral determination.

[0053] As can be seen from the above, this embodiment first evaluates and preprocesses the auxiliary data to obtain effective data, then determines the target features and generates auxiliary judgment information. Through a standardized auxiliary data processing procedure, the value of the auxiliary data can be fully explored, providing more multi-dimensional reference for abnormal behavior identification, making up for the limitations of single video data identification, helping to more comprehensively and accurately judge campus safety anomalies, and improving the credibility of the overall identification results.

[0054] In one embodiment of this application, the auxiliary data includes audio data and environmental sensor data; Based on the data type of the auxiliary data, data quality assessment and preprocessing are performed to obtain multi-source valid data, including: The acoustic data quality is obtained by evaluating the audio data based on acoustic quality assessment methods. Environmental sensor data is evaluated using numerical data assessment methods to obtain environmental data quality. Determine the target acoustic preprocessing strategy based on acoustic data quality; Determine the target numerical preprocessing strategy based on the quality of environmental data; The acoustic data is processed based on the target acoustic preprocessing strategy to obtain the first effective data; Numerical data is processed based on the target numerical preprocessing strategy to obtain the second effective data.

[0055] In this embodiment, the audio data quality level is comprehensively determined by analyzing indicators such as signal-to-noise ratio, signal integrity, and frequency range. For example, the audio data quality level is divided into three levels: excellent, medium, and poor. Among them, the signal-to-noise ratio is used to determine whether there is excessive background noise, the signal integrity is used to check whether there are audio frame breaks or distortion, and the frequency range is used to confirm whether it covers the common frequency range of abnormal sounds, such as the frequencies of screams and breaking glass.

[0056] Secondly, for the environmental sensor data in the auxiliary data, a quality assessment is conducted using numerical data evaluation methods to determine the environmental data quality. Specifically, the numerical reasonableness, data continuity, and data accuracy of the sensor data are checked, and the environmental data quality is classified into different levels. Numerical reasonableness requires determining whether the temperature is within the normal physical range and whether the light intensity matches the characteristics of the actual time period; data continuity requires checking for long periods of missing or abrupt data changes; and data accuracy requires confirming whether it meets the sensor's preset accuracy standards, such as whether humidity data is accurate to one decimal place.

[0057] This embodiment determines the corresponding target acoustic preprocessing strategy based on the evaluated acoustic data quality. For example, if the acoustic data quality is high, simple noise reduction and format standardization are used; if the acoustic data quality is medium, echo cancellation and audio enhancement steps are added; if the acoustic data quality is poor, data repair is required first, such as filling in missing frames, before performing noise reduction, enhancement, and other processing to ensure that the preprocessed data can effectively extract features.

[0058] Secondly, based on the quality of environmental data, determine the appropriate target numerical preprocessing strategy. Specifically, when the environmental data quality is high, only data cleaning and unit standardization are required; when the quality is medium, missing data needs to be supplemented, such as by using interpolation, and minor jump values ​​need to be corrected; when the quality is poor, valid data segments need to be screened first, and then data calibration, such as comparing and correcting with data from other sensors in the same area, needs to be performed to improve data accuracy.

[0059] In this embodiment, the original acoustic data is processed to obtain the first valid data. Taking medium-quality audio data as an example, wavelet denoising algorithm is first used to remove background noise, then echo cancellation technology is used to eliminate echoes generated in an open indoor environment, and finally the processed audio is sampled and normalized to convert it into a standard format suitable for subsequent feature extraction, thus obtaining the first valid data.

[0060] The second set of effective data is obtained by processing the numerical data from environmental sensors. For example, for environmental data of medium quality, linear interpolation is first used to complete the data missing in a short period of time. Then, the abnormal jump values ​​in temperature and humidity are removed by the 3σ principle. Finally, the processed data such as temperature, humidity, light intensity, and personnel density are integrated into a structured numerical matrix to form the second set of effective data.

[0061] As can be seen from the above, this embodiment uses corresponding evaluation methods and preprocessing strategies to obtain effective data for audio and environmental sensor data, respectively. Developing differentiated processing schemes for the characteristics of different types of auxiliary data effectively removes data noise and invalid information, ensuring the quality of effective data. This lays a solid foundation for subsequent extraction of target features and generation of auxiliary judgment information, ensuring that the auxiliary data can reliably support the identification of abnormal behavior on campus.

[0062] In one embodiment of this application, determining the target features corresponding to the multi-source valid data based on the data type of the multi-source valid data includes: The first valid data is preprocessed to obtain the time-frequency domain feature vector; The second set of valid data is fused to obtain a multidimensional feature vector, which includes temperature, humidity, light intensity and population density.

[0063] In this embodiment, the first valid data undergoes preprocessing, specifically frame segmentation, where the continuous audio signal is divided into a fixed-length frame sequence and a window function is added to prevent signal discontinuity between frames. Next, a Fast Fourier Transform (FFT) is used to convert each frame's time-domain audio signal into a frequency-domain signal to obtain frequency distribution information. Finally, key features such as the Mel-frequency cepstral coefficients, spectral centroid, and spectral bandwidth of each frame are calculated, and these features are combined in frame order to form a time-frequency domain feature vector that characterizes the audio signal. Next, the second valid data is time-synchronized to ensure a one-to-one correspondence between different types of data in the time dimension. The synchronized single-type data is then standardized, and the standardized data is fused to obtain a multi-dimensional feature vector. The second valid data includes temperature, humidity, light intensity, and personnel density data.

[0064] As can be seen from the above, this embodiment extracts time-frequency domain feature vectors from valid audio data and fuses valid environmental data to obtain multi-dimensional feature vectors. By extracting key features that match the data type, it can accurately capture abnormal information in audio and environmental data, enabling these features to more accurately represent potential safety anomalies on campus. This provides high-quality feature support for the subsequent generation of auxiliary judgment information, helping to improve the accuracy of anomaly identification.

[0065] In one embodiment of this application, determining auxiliary judgment information based on target features corresponding to multi-source valid data includes: Input the time-frequency domain feature vector into the preset sound event detection model to identify abnormal sound events and output the corresponding audio abnormality probability. The environmental risk index is calculated by inputting the multidimensional feature vector into the anomaly detection model based on the isolated forest algorithm. The probability of audio anomalies is weighted and fused with the environmental risk index to generate auxiliary judgment information.

[0066] In this embodiment, historical data of common abnormal and normal sounds collected on campus are used as samples. Their respective time-frequency domain feature vectors are extracted to construct a training set. A sound event detection model capable of distinguishing between abnormal and normal sounds is trained using a deep learning algorithm. The time-frequency domain feature vector extracted from the first valid data is input into the sound event detection model. The model analyzes the input feature vector, determines whether abnormal sounds exist, and outputs the corresponding audio anomaly probability. The audio anomaly probability is a value between 0 and 1, with a higher value indicating a greater likelihood of abnormal sounds.

[0067] Historical environmental data such as temperature, humidity, light intensity, and population density in various areas of the campus under normal conditions were collected to construct normal training samples. The isolated forest algorithm was used to train the samples, enabling the model to learn the distribution patterns of normal environmental data and form an anomaly detection model. The multidimensional feature vector of the second effective data was input into the anomaly detection model. The model evaluated the anomaly degree of the input multidimensional feature vector and converted the evaluation result into a quantitative environmental risk index. The environmental risk index is a value between 0 and 1, with a higher value indicating a higher risk of anomalies in the environmental condition.

[0068] Based on the characteristics of the campus setting and historical data, weights are assigned to the audio anomaly probability and the environmental risk index. For example, in densely populated areas, the audio weight can be set to 0.6 and the environmental weight to 0.4. Then, the auxiliary judgment information is calculated using the weighted summation formula: Auxiliary judgment information = Audio anomaly probability × Audio weight + Environmental risk index × Environmental weight. This result is the auxiliary judgment information used to help determine abnormal behavior.

[0069] As can be seen from the above, this embodiment utilizes a professional model to obtain the anomaly probability and risk index from audio and environmental features respectively, and then weightedly fuses them to generate auxiliary judgment information. By introducing a professional model to analyze features of different dimensions, the degree of anomaly in audio and the environment can be accurately quantified, and weighted fusion can integrate the information from both to generate a more comprehensive and objective auxiliary judgment basis, further improving the accuracy of identifying abnormal behavior on campus.

[0070] In one embodiment of this application, the probability of audio anomalies is weighted and fused with an environmental risk index to generate auxiliary judgment information, including: Based on the campus schedule context information and the preset weighting strategy table, determine the benchmark audio weighting coefficient corresponding to the audio anomaly probability and the benchmark environmental weighting coefficient corresponding to the environmental risk index. Based on historical auxiliary data anomaly records from the same period and region, the baseline audio weighting coefficients and baseline environmental weighting coefficients are adjusted to obtain the target audio weighting coefficients and target environmental weighting coefficients; among which... If the frequency of historical audio anomaly records is greater than the first frequency threshold, the baseline audio weighting coefficient is increased by the first preset amplitude, and the baseline environment weighting coefficient is decreased by the first preset amplitude. If the frequency of abnormal records in historical environmental data is high, the baseline environmental weighting coefficient will be increased based on the second preset amplitude, and the baseline audio weighting coefficient will be decreased based on the second preset amplitude. Based on the target audio weighting coefficient and the target environment weighting coefficient, the audio anomaly probability and the environmental risk index are weighted and fused to calculate auxiliary judgment information.

[0071] In this embodiment, a baseline weighting coefficient is determined based on the campus schedule context information and a preset weighting strategy table. The weighting strategy table is constructed by decomposing the campus schedule context into three dimensions: time period, monitoring area function type, and ongoing campus activity type. For each specific scenario formed by different combinations of these three dimensions, a corresponding baseline audio weighting coefficient and a baseline environment weighting coefficient are set. It defines the baseline audio weighting coefficient and the baseline environment weighting coefficient under different campus schedule context scenarios. Specifically, when the campus schedule context is: time period: Wednesday afternoon during lab class; monitoring area function type: chemistry lab; ongoing campus activity type: Grade 11 chemistry lab operation assessment, by querying the preset weighting strategy table, the baseline audio weighting coefficient for this scenario can be determined to be 0.6 and the baseline environment weighting coefficient to be 0.4. When the scenario changes to: time period: Sunday morning; monitoring area function type: first floor corridor of the teaching building; ongoing campus activity type: occasional students returning to school to retrieve personal items, querying the weighting strategy table yields a baseline audio weighting coefficient of 0.3 and a baseline environment weighting coefficient of 0.7. If the scenario is: time period: Monday morning during recess; monitoring area function type: school playground; ongoing campus activity type: school-wide recess exercise, then the baseline audio weighting coefficient and baseline environment weighting coefficient can be obtained from the weighting strategy table as 0.5, thus achieving accurate matching of the baseline weighting coefficients under different scenarios.

[0072] Secondly, based on historical auxiliary data anomaly records from the same period and region, the baseline weighting coefficient is adjusted to obtain the target weighting coefficient. The first and second preset ranges are determined according to the impact of historical anomaly records on actual security events. Specifically, historical data from the same period and monitoring area as the current time period are retrieved, and the frequency of audio anomaly records and environmental data anomaly records is statistically analyzed to determine the first and second frequency thresholds. If the frequency of historical audio anomaly records is greater than the first frequency threshold, it indicates that audio anomalies in that region are more valuable for reference. The baseline audio weighting coefficient is increased based on the first preset range, while the baseline environmental weighting coefficient is decreased based on the first preset range to obtain the target audio weighting coefficient and the target environmental weighting coefficient. If the frequency of historical environmental data anomaly records is greater than the second frequency threshold, it indicates that environmental anomalies in that region are more critical for judgment. The baseline environmental weighting coefficient is increased based on the second preset range, while the baseline audio weighting coefficient is decreased based on the second frequency threshold to obtain the target audio weighting coefficient and the target environmental weighting coefficient. For example, if over 80% of historical audio anomaly records are ultimately confirmed to be related to real security incidents, the first preset amplitude can be set to 0.15; if the accuracy rate of environmental anomaly records in providing early warnings of security incidents reaches 70%, the second preset amplitude can be set to 0.1. When the baseline audio weighting coefficient is 0.6 and the baseline environmental weighting coefficient is 0.4, the audio adjustment condition is triggered, then the target audio weighting coefficient is 0.6 + 0.15 = 0.75, and the target environmental weighting coefficient is 0.4 - 0.15 = 0.25.

[0073] This embodiment utilizes a target weighting coefficient to weight and fuse the audio anomaly probability and the environmental risk index to generate auxiliary judgment information. The audio anomaly probability is multiplied by the target audio weighting coefficient to obtain the audio anomaly contribution value; the environmental risk index is multiplied by the target environmental weighting coefficient to obtain the environmental risk contribution value; the two contribution values ​​are then added together, and the result is used as auxiliary judgment information for subsequent comprehensive judgment of abnormal behavior.

[0074] As can be seen from the above, this embodiment first determines the baseline weighting coefficients based on the campus schedule and weight table, then adjusts the coefficients according to historical anomaly records and integrates them to generate auxiliary judgment information. Considering the impact of campus schedules and historical data on the importance of different auxiliary information, dynamically adjusting the weighting coefficients allows the integrated auxiliary judgment information to better fit the actual campus scenario, more accurately reflect abnormal situations, and provide a more targeted reference for the comprehensive judgment of abnormal behavior.

[0075] In one embodiment of this application, the first abnormal behavior identification result, the first confidence level, and auxiliary judgment information are input into a preset decision model for comprehensive judgment to generate a target identification result and a danger level, including: The decision uncertainty index is calculated based on the first confidence level and auxiliary judgment information; Based on the decision uncertainty index and the behavior type of the first abnormal behavior identification result, a target decision path is selected from the preset set of decision paths; Based on the target decision path, the first abnormal behavior identification result, the first confidence level, and auxiliary judgment information are processed to obtain the target identification result and the danger level.

[0076] In this embodiment, the decision uncertainty index is first calculated based on the first confidence level and auxiliary judgment information. Specifically, a preset weighted difference formula is used for calculation, where the formula is: Decision uncertainty index = |First confidence level -Auxiliary judgment information| ×Adjustment coefficient The adjustment coefficient is dynamically set according to the specific scenario; for example, it is set to 1.2 for high-risk scenarios and 1.0 for normal scenarios. This formula quantifies the difference between the visual recognition result and the auxiliary information into a value between [0,1]. The higher the value, the greater the discrepancy between the initial recognition result and the auxiliary information, and the higher the uncertainty of the system in the decision-making process.

[0077] Secondly, based on the decision uncertainty index and the behavior type identified by the first abnormal behavior, a target decision path is selected from a pre-defined set of decision paths. This set of decision paths is pre-constructed based on historical campus security data, expert experience rules, and typical scenario characteristics. Essentially, the set of decision paths is a mapping table containing multiple differentiated judgment logics. Its core design dimensions include the threshold range of the first confidence level, specific abnormal behavior types, and combination patterns of auxiliary judgment information. By configuring corresponding risk assessment logic and evidence verification mechanisms for different combinations of confidence level, behavior type, and scenario context, multiple decision paths pointing to different levels of judgment rigor and handling procedures are formed, collectively constituting the decision path set.

[0078] This embodiment selects a target decision path from a pre-defined set of decision paths by using the calculated decision uncertainty index and the behavior type identified by the first abnormal behavior. For example, the decision path set needs to pre-define three types of paths: Path 1 is suitable for low-uncertainty and non-high-risk behaviors; Path 2 is suitable for medium-uncertainty or high-risk behaviors, or behaviors such as climbing a wall; Path 3 is suitable for high-uncertainty or high-risk behaviors. In actual operation, it is first determined whether the current decision uncertainty index range and behavior type belong to high-risk, and then the corresponding path is matched as the target decision path.

[0079] The relevant information is processed according to the target decision path to obtain the target identification result and the danger level. If path one is selected, the first abnormal behavior identification result is directly used as the target identification result, and the danger level is determined by querying the preset confidence-danger level mapping table based on the first confidence level. If path two is selected, a comprehensive score is calculated by fusing the first confidence level and auxiliary judgment information, and the initial result is verified or corrected based on whether the score exceeds the threshold. Finally, the danger level is determined by combining the behavior type of the corrected result. If path three is selected, a deep neural network model is called to perform multi-dimensional fusion analysis on the first abnormal behavior characteristics, the first confidence level, and auxiliary judgment information, and the model directly outputs the target identification result and the corresponding danger level.

[0080] As can be seen from the above, this embodiment first calculates the decision uncertainty index, then selects a decision path based on the index and behavior type, and finally processes the results according to the path to obtain the target result and the danger level. By flexibly selecting the decision path based on uncertainty and behavior type, more appropriate processing methods can be adopted for different identification situations, avoiding the limitations of a single decision-making model, improving the flexibility and accuracy of abnormal behavior judgment, and ensuring accurate identification of campus safety anomalies and determination of danger levels.

[0081] In one embodiment of this application, a target decision path is selected from a preset set of decision paths based on the decision uncertainty index and the behavior type of the first abnormal behavior identification result, including: If the decision uncertainty index is less than the preset first uncertainty threshold, and the first abnormal behavior identification result does not belong to the predefined high-risk behavior type, the first decision path is selected as the target decision path. If the decision uncertainty index is greater than or equal to the preset first uncertainty threshold and less than or equal to the preset second uncertainty threshold, or if the first abnormal behavior identification result belongs to the predefined high-risk behavior type, then the second decision path is selected as the target decision path. If the decision uncertainty index is greater than the preset second uncertainty threshold, or if the first abnormal behavior identification result belongs to a predefined high-risk behavior type, the third decision path is selected as the target decision path. The preset second uncertainty threshold is greater than the preset first uncertainty threshold.

[0082] In this embodiment, preset decision path determination criteria and thresholds are clearly defined. A first uncertainty threshold and a second uncertainty threshold are preset, with the second uncertainty threshold being greater than the first uncertainty threshold. At the same time, predefined high-risk behavior types are defined, and corresponding triggering conditions are set for the first decision path, the second decision path, and the third decision path, respectively, as the basis for path selection.

[0083] The setting of the first and second uncertainty thresholds follows a four-step logic: data-driven assessment, scenario adaptation, expert verification, and dynamic iteration. First, based on historical data, the system analyzes historical identification data from the campus security system, focusing on the correlation between the decision uncertainty index and the false positive / false negative rate. A baseline value balancing accuracy and efficiency is selected. Second, adjustments are made based on scenario risk sensitivity, with the baseline threshold fluctuating upwards and downwards according to the different security needs of different scenarios. Third, an expert review process is conducted, with campus security leaders and algorithm engineers jointly verifying the threshold's rationality and optimizing it based on typical historical cases. Finally, a quarterly iteration mechanism is established, continuously fine-tuning the threshold based on actual handling efficiency and false positive / false negative data each quarter, while temporarily adjusting the threshold for extremely high-risk behaviors to ensure it always aligns with the actual needs of campus security management. Fourth, the calculated decision uncertainty index is compared with the first uncertainty threshold to determine if the first threshold is lower than the first uncertainty threshold. Simultaneously, it is determined whether the first abnormal behavior identification result belongs to a predefined high-risk behavior type.

[0084] This embodiment selects the first decision path as the target decision path based on the decision uncertainty index being less than the first uncertainty threshold and the first abnormal behavior identification result not belonging to a predefined high-risk behavior type. This is suitable for scenarios where the identification result is reliable and the risk is low. Based on the decision uncertainty index being greater than or equal to the first uncertainty threshold and less than or equal to the second uncertainty threshold, the second path is triggered regardless of whether the behavior type is high-risk. Alternatively, if the decision uncertainty index is not in this range, but the first abnormal behavior identification result belongs to a predefined high-risk behavior type, the second decision path is also selected as the target decision path. This is suitable for scenarios with medium uncertainty or involving high-risk behaviors. Based on the decision uncertainty index being greater than the second uncertainty threshold, the third path is triggered regardless of the behavior type. Alternatively, if the decision uncertainty index does not exceed the second threshold, but the first abnormal behavior identification result belongs to a predefined high-risk behavior type, the third decision path is selected as the target decision path. This is suitable for scenarios with high uncertainty or where high-risk behaviors require in-depth verification.

[0085] As can be seen from the above, this embodiment clarifies the corresponding decision path based on different combinations of decision uncertainty index and behavior type. Clear path selection rules enable the model to quickly and accurately determine the appropriate decision path when facing different recognition scenarios, avoiding arbitrariness in path selection, ensuring the standardization and consistency of the decision-making process, and thus improving the stability and reliability of campus abnormal behavior recognition results.

[0086] In one embodiment of this application, the first abnormal behavior identification result, the first confidence level, and auxiliary judgment information are processed based on the target decision path to obtain the target identification result and the danger level, including: If the first decision path is selected, the first abnormal behavior identification result is directly used as the target identification result, and the danger level is determined by querying the behavior type danger level mapping table based on the first confidence level. If the second decision path is chosen, the comprehensive confidence score is calculated based on the first confidence level and auxiliary judgment information; Based on the comprehensive confidence score and the preset decision rules, the first abnormal behavior identification result is verified or corrected to generate the target identification result; Based on the type of target identification results and the comprehensive confidence score, the hazard level is determined through a hazard level mapping table. If the third decision path is chosen, the initial risk level is determined based on the behavior type and decision uncertainty index of the first abnormal behavior identification results; Based on the initial risk level, feature fusion is performed on video data, audio data, and associated environmental sensor data to generate a multimodal feature vector; Multimodal feature vectors will be generated and input into a preset neural network decision model to generate target recognition results and hazard levels.

[0087] In this embodiment, different processing logics are used to generate target identification results and risk levels for different decision paths, ensuring that the judgment process conforms to the risk characteristics of the scenario and the reliability of information.

[0088] When the first decision path is selected, it directly determines the result because it is suitable for scenarios with high credibility of the identification results and low behavioral risk. Specifically, there is no need to revise the initial identification conclusion; the first abnormal behavior identification result is directly determined as the target identification result. Based on the specific value of the first confidence level, a preset three-dimensional mapping table of behavior type-confidence level-risk level is queried to determine the risk level of different behaviors at different confidence levels. For example, a student running alone corresponds to low risk at a confidence level of 0.8 and remains low risk at a confidence level of 0.9. This allows for rapid matching and determination of the corresponding risk level. The entire process requires no additional data verification, balancing efficiency and accuracy.

[0089] When the second decision path is selected, due to the uncertainty of the identification result or the scenario where the behavior belongs to a high-risk type, the judgment is completed through a three-step comprehensive analysis: First, calculate the comprehensive confidence score using a weighted fusion formula. The weights can be dynamically adjusted according to the scenario. In a typical scenario, the default formula is used: Comprehensive Score = First Confidence × 0.6 + Auxiliary Judgment Information × 0.4. For example, when the first confidence is 0.7 and the auxiliary judgment information is 0.8, the comprehensive score is 0.7 × 0.6 + 0.8 × 0.4 = 0.74. Second, based on the calculated comprehensive score and the preset decision rules, identify the first abnormal behavior. The results are then verified or corrected. Specifically, when the comprehensive score is greater than or equal to 0.7, the identification result is considered valid. When the comprehensive score is greater than or equal to 0.5 but less than 0.7, the first abnormal behavior identification result needs to be verified or corrected. When the comprehensive score is less than 0.5, it is judged as normal behavior. If the original identification result is suspected pushing and the comprehensive score is 0.65 (in the 0.5-0.7 range), the behavior is corrected to slight physical contact, generating a more accurate target identification result. The third step is to consult the danger level mapping table based on the specific type of the target identification result and the comprehensive confidence score to determine the final danger level. The danger level mapping table is constructed based on historical campus incident data, analysis of the severity of behavioral consequences, and expert experience, and is used to quantify and correlate different behavior types and their confidence intervals.

[0090] When the third decision path is selected, due to the high uncertainty of the identification results or the extremely high-risk nature of the behavior, a deep fusion of multi-source data is required to complete the judgment to ensure no omissions. Specifically, based on the behavior type and decision uncertainty index of the first abnormal behavior identification result, the risk range is initially defined. Secondly, a matching feature fusion strategy is invoked according to the initial risk level. For example, in high-risk scenarios, the temporal features of human key points in video data, the time-frequency domain features of audio data, and the multi-dimensional features of environmental sensor data need to be integrated. Through feature splicing and normalization, a unified feature vector including multimodal information is generated. Finally, this multimodal feature vector is input into a preset neural network decision model to obtain the final accurate target identification result and the corresponding risk level. The neural network decision model is trained and optimized using historical campus multimodal abnormal behavior samples.

[0091] As can be seen from the above, this embodiment formulates specific processing methods for different decision-making paths: low-uncertainty and non-high-risk behaviors directly determine the result; medium-uncertainty or high-risk behaviors calculate a comprehensive score for verification and correction; and high-uncertainty or high-risk behaviors fuse multimodal features and are judged using a neural network model. This differentiated and precise processing can adopt an adaptive solution based on the complexity and risk level of the identification scenario, maximizing the accuracy of target identification results and hazard level determination, thus ensuring campus safety.

[0092] Corresponding to the machine vision-based intelligent recognition method for abnormal campus security behavior in the above embodiments, Figure 2 This is a structural block diagram of a machine vision-based intelligent recognition system for abnormal campus security behavior, provided as an embodiment of this application. For ease of explanation, only the parts relevant to this embodiment are shown. References Figure 2 The machine vision-based intelligent identification system for abnormal campus security behavior 20 includes: a target parameter adjustment module 21, a monitoring video analysis module 22, a dual condition judgment module 23, an auxiliary information generation module 24, a target strategy determination module 25, and a target strategy determination module 26.

[0093] in, The target parameter adjustment module 21 is used to: adjust the sensitivity parameters and confidence threshold of the preset visual behavior recognition model based on the campus schedule context information, so as to obtain the target sensitivity parameters and target confidence threshold; The surveillance video analysis module 22 is used to input the video data of the campus surveillance area in the current time period into the visual behavior recognition model for analysis, and obtain the first abnormal behavior recognition result and the corresponding first confidence level. The dual-condition judgment module 23 is used to determine whether the first confidence level is less than the target confidence level threshold and whether the first abnormal behavior identification result is a predefined high-risk behavior. The auxiliary information generation module 24 is used to obtain auxiliary data associated with the video data if any condition is met, and generate auxiliary judgment information based on the auxiliary data. The target result generation module 25 is used to: input the first abnormal behavior identification result, the first confidence level, and auxiliary judgment information into a preset decision model for comprehensive judgment, and generate a target identification result and a risk level; if any condition is not met, the risk level is determined by querying a preset behavior type risk level mapping table based on the first abnormal behavior identification result and the first confidence level, and the first abnormal behavior identification result is used as the target identification result. The target strategy determination module 26 is used to determine the corresponding target early warning strategy based on the hazard level and campus schedule context information.

[0094] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 2 The functions of the target parameter adjustment module 21, monitoring video analysis module 22, dual condition judgment module 23, auxiliary information generation module 24, target strategy determination module 25, and target strategy determination module 26 are shown.

[0095] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0096] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0097] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0098] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the intelligent identification method for abnormal campus security behavior based on machine vision provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0099] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0100] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0101] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0102] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0103] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0104] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0105] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0106] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent identification of abnormal behaviors in campus security based on machine vision, characterized in that, The method comprises the following steps: adjusting sensitivity parameters and confidence thresholds of a preset visual behavior recognition model based on campus schedule context information to obtain target sensitivity parameters and target confidence thresholds; inputting video data of a current period of a campus monitoring area into the visual behavior recognition model for analysis to obtain a first abnormal behavior recognition result and a corresponding first confidence; determining whether the first confidence is less than the target confidence threshold and whether the first abnormal behavior recognition result is a predefined high-risk behavior; if any condition is met, obtaining auxiliary data associated with the video data, and generating auxiliary judgment information based on the auxiliary data; inputting the first abnormal behavior recognition result, the first confidence, and the auxiliary judgment information into a preset decision model for comprehensive judgment to generate a target recognition result and a danger level; if any condition is not met, determining a danger level by querying a preset behavior type-danger level mapping table based on the first abnormal behavior recognition result and the first confidence, and taking the first abnormal behavior recognition result as the target recognition result; determining a corresponding target early warning strategy based on the danger level and the campus schedule context information. 2.The method of claim 1, wherein, The method of adjusting sensitivity parameters and confidence thresholds of a preset visual behavior recognition model based on campus schedule context information to obtain target sensitivity parameters and target confidence thresholds comprises the following steps: analyzing the campus schedule context information to determine a current period attribute, a monitoring area function type, and a calendar event; querying a preset first mapping table based on the current period attribute and the monitoring area function type to obtain a basic sensitivity parameter and a basic confidence threshold; weighting and correcting the basic sensitivity parameter and the basic confidence threshold based on the calendar event type to obtain a corrected sensitivity parameter and a confidence threshold; adjusting the corrected sensitivity parameter and the confidence threshold based on feedback statistical information of historical recognition results in a preset period to obtain the target sensitivity parameter and the target confidence threshold. 3.The method of claim 2, wherein, The calendar event type comprises a first risk event type, a second risk event type, and a third risk event type, the risk level of the first risk event type is greater than that of the second risk event type, and the risk level of the second risk event type is greater than that of the third risk event type; The method of weighting and correcting the basic sensitivity parameter and the basic confidence threshold based on the calendar event type to obtain a corrected sensitivity parameter and a confidence threshold comprises the following steps: for the first risk event type, increasing the basic sensitivity parameter based on a first adjustment factor and decreasing the basic confidence threshold based on the first adjustment factor; for the second risk event type, keeping the basic sensitivity parameter and the basic confidence threshold unchanged; for the third risk event type, decreasing the basic sensitivity parameter based on a second adjustment factor and increasing the basic confidence threshold based on the second adjustment factor.

4. The method of claim 3, wherein the method further comprises: The feedback statistical information comprises a false positive rate and a false negative rate. The feedback statistical information based on the historical identification result in the preset period is used to adjust the modified sensitivity parameter and the confidence threshold to obtain the target sensitivity parameter and the target confidence threshold, including: If the false positive rate is greater than a first preset threshold and the false negative rate is less than or equal to a second preset threshold, the modified confidence threshold is increased based on a third adjustment factor or the modified sensitivity parameter is reduced based on a third adjustment factor to obtain the target sensitivity parameter and the target confidence threshold; If the false negative rate is greater than a second preset threshold and the false positive rate is less than or equal to a first preset threshold, the modified confidence threshold is reduced based on a fourth adjustment factor or the modified sensitivity parameter is increased based on a fourth adjustment factor to obtain the target sensitivity parameter and the target confidence threshold. 5.The method of claim 1, wherein, The auxiliary data is used to generate auxiliary judgment information, including: Based on the data type of the auxiliary data, data quality evaluation and preprocessing are respectively performed to obtain multi-source effective data; Based on the data type of the multi-source effective data, the target features corresponding to the multi-source effective data are determined; Based on the target features corresponding to the multi-source effective data, the auxiliary judgment information is determined. 6.The method of claim 5, wherein, The auxiliary data includes audio data and environmental sensor data; Based on the data type of the auxiliary data, data quality evaluation and preprocessing are respectively performed to obtain multi-source effective data, including: The audio data is evaluated based on an acoustic quality evaluation method to obtain acoustic data quality; The environmental sensor data is evaluated based on a numerical data evaluation method to obtain environmental data quality; A target acoustic preprocessing strategy is determined based on the acoustic data quality; A target numerical preprocessing strategy is determined based on the environmental data quality; The acoustic data is processed based on the target acoustic preprocessing strategy to obtain first effective data; The numerical data is processed based on the target numerical preprocessing strategy to obtain second effective data.

7. The method of claim 6, wherein the method further comprises: Based on the data type of the multi-source effective data, the target features corresponding to the multi-source effective data are determined, including: The first effective data is preprocessed to obtain a time-frequency domain feature vector; The second effective data is subjected to data fusion processing to obtain a multi-dimensional feature vector, and the multi-dimensional feature vector includes temperature, humidity, light intensity, and personnel density. 8.The method of claim 1, wherein, The first abnormal behavior identification result, the first confidence, and the auxiliary judgment information are input into a preset decision model for comprehensive determination to generate a target identification result and a danger level, including: Based on the first confidence and the auxiliary judgment information, a decision uncertainty index is calculated; According to the decision uncertainty index and the behavior type of the first abnormal behavior identification result, a target decision path is selected from a preset decision path set; The first abnormal behavior identification result, the first confidence, and the auxiliary judgment information are processed based on the target decision path to obtain the target identification result and the danger level. 9.The method of claim 8, wherein, The behavior type according to the decision uncertainty index and the first abnormal behavior recognition result is used for selecting a target decision path from a preset decision path set, and the target decision path comprises: If the decision uncertainty index is less than a preset first uncertainty threshold, and the first abnormal behavior recognition result does not belong to a predefined high-risk behavior type, a first decision path is selected as the target decision path; If the decision uncertainty index is greater than or equal to the preset first uncertainty threshold and less than or equal to a preset second uncertainty threshold, or the first abnormal behavior recognition result belongs to the predefined high-risk behavior type, a second decision path is selected as the target decision path; If the decision uncertainty index is greater than the preset second uncertainty threshold, or the first abnormal behavior recognition result belongs to the predefined high-risk behavior type, a third decision path is selected as the target decision path; The preset second uncertainty threshold is greater than the preset first uncertainty threshold.

10. The machine vision-based campus security abnormal behavior intelligent recognition system according to claim 1, wherein a target parameter adjustment module is configured to adjust sensitivity parameters and a confidence threshold of a preset visual behavior recognition model based on campus schedule context information to obtain target sensitivity parameters and a target confidence threshold; a monitoring video analysis module is configured to input video data of a monitoring area of the campus in a current period into the visual behavior recognition model for analysis to obtain a first abnormal behavior recognition result and a corresponding first confidence; a dual condition judgment module is configured to judge whether the first confidence is less than the target confidence threshold and whether the first abnormal behavior recognition result is a predefined high-risk behavior; an auxiliary information generation module is configured to, if any condition is met, acquire auxiliary data associated with the video data and generate auxiliary judgment information based on the auxiliary data; a target result generation module is configured to input the first abnormal behavior recognition result, the first confidence and the auxiliary judgment information into a preset decision model for comprehensive judgment to generate a target recognition result and a danger level; if any condition is not met, to determine a danger level by querying a preset behavior type-danger level mapping table based on the first abnormal behavior recognition result and the first confidence, and to use the first abnormal behavior recognition result as the target recognition result; and a target strategy determination module is configured to determine a corresponding target warning strategy based on the danger level and the campus schedule context information. ​

Images