Method for monitoring and identifying dangerous behaviors of smart park

Abstract

The invention discloses a method for monitoring and identifying dangerous behaviors of a smart park, and relates to the field of intelligent security and protection. The method comprises the following steps: acquiring an environment data stream through a sensing device deployed at a key point location of a park and carrying out standardization processing on the environment data stream; the data stream is sent to an edge computing node, a target is recognized through a lightweight target detection model, the detection confidence coefficient is obtained, a motion track is generated through a multi-target tracking algorithm, the rule matching degree is calculated according to a predefined rule base, and dangerous behaviors are preliminarily recognized; calculating a comprehensive model prediction credibility based on the detection credibility and the rule matching degree, and generating a final comprehensive credibility through a multi-dimensional credibility fusion model in combination with the time sequence consistency credibility and the scene matching credibility; and dangerous behavior trend prediction is carried out according to the comprehensive confidence, and an alarm instruction is generated to trigger disposal linkage, so that advanced prevention and quick response of dangerous behaviors are realized, and the intelligent level and operation and maintenance efficiency of park security and protection are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent security and protection, and in particular to a method for monitoring and identifying dangerous behaviors in a smart park. BACKGROUND

[0002] As an important carrier of urban modernization development, the safety management of a smart park is of great importance. The park is densely populated and has complex business formats, covering various scenarios such as production, warehousing, and office work, and harbors various dangerous behaviors such as regional intrusion, falling, work violation, and fire. Timely and accurate monitoring and identification of these behaviors are the core prerequisite for protecting the safety of personnel and property and maintaining the stable operation of the park, and have extremely important social and economic significance. Currently, the monitoring system widely used in traditional parks is mainly based on the traditional mode of "video monitoring + manual research and judgment". The specific process of this mode is as follows: cameras are deployed at key points in the park to collect video data, and the video data is transmitted to the central monitoring room in real time; then, security personnel rely on manual patrol and interpretation of a large number of video images; to assist the manual work, the system usually integrates some primary video analysis functions, such as moving detection or regional intrusion alarm, that is, when there is a pixel-level change or an object enters the preset electronic fence in the monitoring image, the system will trigger an alarm to prompt the security personnel. Although the above method has been widely applied, its inherent technical limitations have led to serious challenges in effectiveness. First, the core link relies on manual work, which is not only inefficient but also causes personnel to be easily fatigued after a long time of monitoring, resulting in a high rate of missed reports of instantaneous dangerous behaviors. Second, the technical recognition ability is weak, and the alarm mechanism based on pixel changes cannot distinguish target types (such as people and vehicles) and behavior semantics (such as fighting and falling), which leads to a high rate of false alarms caused by disturbances such as light and floating objects, and frequently produces the wolf effect, which dilutes the vigilance of security personnel. More importantly, each security sub-system (such as video, access control, and broadcasting) usually operates independently, lacks coordination and linkage, and lacks intelligent decision-making ability, resulting in a long response chain from discovering abnormalities to on-site disposal, so that the entire system is essentially a passive and lagging after-the-fact tracing tool, which is difficult to achieve the active early warning and precise prevention and control required by modern smart parks. SUMMARY

[0003] The purpose of the present application can be achieved by the following technical solution: a method for monitoring and identifying dangerous behaviors in a smart park, comprising the following steps: S1, collecting environmental data streams through perception devices deployed at key points in the park, and standardizing the environmental data streams to obtain standard environmental data streams; S2, sending the standard environmental data streams to an edge computing node, and performing the following operations on the edge computing node: A plurality of preset target categories in the standard environment data stream are identified by a lightweight target detection model, and a detection confidence of the target is obtained; the target categories include personnel, vehicles, safety helmets, flames, and smoke; A unique identity is assigned to the target by a multi-target tracking algorithm, and a continuous motion trajectory is generated; According to a predefined rule library, the motion trajectory of the target is matched with a preset electronic fence area, an aggregation threshold, a violation stay duration, or a safety specification, a rule matching degree is calculated, and a behavior data set is output based on this; S3, based on the detection confidence and the rule matching degree, a comprehensive model prediction confidence is calculated, the stability of the comprehensive model prediction confidence is evaluated in a sliding time window through time sequence consistency analysis, a time sequence consistency confidence is calculated; a scene matching confidence is obtained by matching the output behavior data set with a scene behavior benchmark library, and then the detection confidence is weighted and fused with the time sequence consistency confidence and the scene matching confidence to generate a final comprehensive confidence; S4, a risk aggravation warning signal is generated by linear regression analysis of the comprehensive confidence in the time window; then, based on a preset dangerous behavior type, a comprehensive confidence, and a risk aggravation warning signal, a three-dimensional decision model is used to dynamically determine an alarm level; according to the alarm level, a corresponding hierarchical push mechanism and a preset device linkage operation are started; finally, a digital disposal work order is created for the triggered alarm, and the disposal effect is automatically verified by continuously monitoring whether the same behavior occurs again in the incident area to realize closed-loop management.

[0004] Preferably, S1 specifically includes: A plurality of perception devices deployed at key points in the geographical space of the park continuously collect raw environment data streams, including visual data streams, thermal imaging data streams, and Internet of Things sensor signals; The collected raw environment data streams are uniformly transcoded into a preset standard format, and all standard environment data streams are stamped with a uniform timestamp; The standard environment data stream is enhanced, including denoising, color enhancement, and contrast adjustment; The preprocessed standard environment data stream is sent to S2.

[0005] Preferably, S2 specifically includes: The standard environment data stream that has completed standardization is sent to the edge computing node pre-deployed by the system; in the video image decoded from the standard environment data stream, one or more virtual electronic fence areas are preset; The standard environment data stream is decoded into image frames, and each frame is input into the lightweight target detection model to output the bounding box information, target category, and detection confidence of all targets in the current frame; The bounding box information of the target, the target category and the detection confidence are input into a multi-target tracking algorithm, a unique identity ID is assigned to each target, and a continuous motion trajectory of each target is formed; the motion trajectory at least includes the position coordinates, the motion speed and the motion direction of the target; Based on the output target category, detection confidence, target identity ID and motion trajectory, a preliminary specific dangerous behavior recognition is performed according to a predefined rule library, and a rule matching degree is calculated; All recognized behavior events are encapsulated into a unified behavior data set as the final output of the S2 step.

[0006] Preferably, the comprehensive model prediction confidence is specifically: The product of the detection confidence and the rule matching degree is the comprehensive model prediction confidence, which is used to comprehensively reflect the target detection accuracy and the behavior rule compliance degree.

[0007] Preferably, the scene matching confidence is specifically: The preset scene behavior benchmark library is queried based on the current system time, the event occurrence position and the behavior type to obtain a corresponding benchmark threshold; The real-time recognized data is compared with the benchmark threshold, and different scene matching confidences are respectively given according to different intervals of the comparison result.

[0008] Preferably, the time sequence consistency confidence is specifically: Based on the video frame sequence decoded from the standard environment data stream, a sliding time window with a length of N frames is defined, which covers the current video frame being processed and the previous N-1 frames of video, collectively referred to as historical video frames; In the sliding time window with a preset time length, the number of frames in which the dangerous behavior type preliminarily recognized in all historical video frames is the same as that of the current video frame is counted, and the proportion of the same frame number in the total number of frames N in the window is calculated; The standard deviation of the comprehensive model prediction confidence in the time window is calculated, and the standard deviation is normalized according to a preset confidence standard deviation reference range to obtain a normalized comprehensive model prediction confidence standard deviation; then, based on the normalized comprehensive model prediction confidence standard deviation, a volatility adjustment coefficient is determined through a numerical mapping rule, wherein the volatility adjustment coefficient is negatively correlated with the normalized comprehensive model prediction confidence standard deviation; The proportion of the number of historical frames consistent with the recognition result of the current frame in the window is multiplied by an adjustment coefficient based on the volatility of the comprehensive model prediction confidence, and the product thereof is taken as the time sequence consistency confidence.

[0009] Preferably, the rule matching degree is specifically: Based on the real-time monitored target motion trajectory, state attribute and quantity relationship, comparison and analysis are carried out with the predefined rule condition, and a rule matching degree value between 0 and 1 is mapped according to the degree of compliance with the dangerous behavior condition; when the behavior does not comply with the dangerous condition at all, the rule matching degree is 0; when the behavior fully complies with or exceeds the dangerous condition, the rule matching degree is 1.

[0010] Preferably, the alarm level determination specifically comprises: A behavior type-reference risk mapping table is preset, a reference risk level is set for each dangerous behavior; on the basis of the reference risk level, a level fine-tuning is carried out in combination with the real-time comprehensive confidence, and the alarm level is promoted according to the risk aggravation early warning signal.

[0011] Preferably, the device linkage operation is executed by querying a preset event device linkage rule library, and the rule setting of the rule library follows the principles of proximity, urgency and matching.

[0012] Preferably, the risk aggravation early warning signal generation step comprises: The system maintains a two-dimensional rectangular coordinate system with time as the horizontal axis and comprehensive confidence as the vertical axis, and continuously draws the real-time calculated comprehensive confidence as a data point in the two-dimensional rectangular coordinate system to form a discrete point sequence evolving over time; A backtracking time window is set, and the discrete points in the window are extracted to constitute an initial data point set; data preprocessing is carried out on the initial data point set, which comprises identifying and eliminating abnormal fluctuation points by using the median absolute deviation method, and performing smoothing processing on the data sequence after eliminating abnormal points by using the exponential weighted moving average method to obtain a preprocessed data point set; The preprocessed data point set is input into a linear regression model to fit the time as the independent variable, and the slope of the fitted straight line is taken as the comprehensive confidence change slope at the current time; The comprehensive confidence change slope is compared with a preset positive slope threshold value; if the comprehensive confidence change slope is greater than the preset positive slope threshold value, it is determined that the risk level is rapidly rising, and the system immediately generates and outputs a risk aggravation early warning signal; The positive slope threshold value is selected as the optimal value from the candidate threshold values according to a predetermined detection rate and false alarm rate balance strategy by collecting a sample set of time sequence confidence data and corresponding event real labels generated in the historical operation of the system, and analyzing the early warning detection rate and false alarm rate under different candidate threshold values based on the sample set through a receiver operating characteristic curve.

[0013] The beneficial effects of the present application are: (1) The present application breaks through the limitation of traditional monitoring relying on instantaneous state recognition only, and systematically introduces time series trend analysis; by linear regression to calculate the change slope of the comprehensive confidence in real time, the system can sensitively capture the accumulation trend of risk in the early stage, and send a risk intensification warning signal before the dangerous behavior completely breaks out. This makes the security response move from post-disposal to in-process inhibition or even pre-disposal, realizes real active security through continuous tracking, resource prepositioning and key prompting of high-risk targets, and realizes real active security; (2) The present application introduces three-dimensional fusion verification of model prediction confidence, time series consistency and scene matching degree in the decision center based on real-time recognition on the edge side, effectively filters out false alarms caused by environmental interference, and greatly improves the reliability of alarm information; further, the system constructs dynamic hierarchical response rules based on inherent risk of behavior, real-time comprehensive confidence and risk trend signal, realizes accurate judgment and differentiated disposal from general violation to emergency to risk acceleration, and ensures that limited security resources can be preferentially invested in the most urgent and highest deterioration possibility threat; (3) The present application not only completes the automation from event detection to linkage disposal, but also forms a management closed loop through digital disposal work order and effect automatic verification mechanism; all disposal records and subsequent monitoring results are fed back to the database to drive continuous optimization of the identification model and the decision threshold; this makes the system evolve from a static tool to an intelligent system with self-learning and self-optimization capabilities, and finally builds a smart security ecosystem with sensitive perception, intelligent decision, efficient disposal and continuous evolution. BRIEF DESCRIPTION OF DRAWINGS

[0014] The present application will be further described below in conjunction with the drawings.

[0015] Figure 1 is a principle block diagram of a dangerous behavior monitoring and identification method of a smart park of the present application.

[0016] Figure 2 is a flowchart of a dangerous behavior monitoring and identification method of a smart park of the present application. DETAILED DESCRIPTION

[0017] To solve the above technical problems, the present application provides a dangerous behavior monitoring and identification method of a smart park, which will be described in detail below in conjunction with Figure 1 and Figure 2 specific embodiments; A dangerous behavior monitoring and identification method of a smart park, the specific steps comprising: S1, the monitoring camera equipment continuously collects video data in the park in real time, and transmits the video data to the corresponding edge node through the network; the video data is standardized before being input into the identification module, including resolution unification, color mapping calibration and encoding format alignment, so that the input data remains consistent, facilitating subsequent model identification; The S1 includes the following steps: S11, by deploying a variety of sensing devices at key points in the geographical space of the park, continuously collecting raw environmental data streams in a synchronous or asynchronous manner; the key points include entrances and exits, main roads, parking lots, production workshops, warehouses, and fire passages.

[0018] The raw environmental data stream specifically includes: Visual data stream: continuous video stream collected by high-definition visible light camera; Thermal imaging data stream: infrared video stream reflecting the temperature distribution of the scene collected by thermal imaging camera; Internet of Things sensor signals: including trigger signals from smoke detectors, temperature detectors, and door magnetic switches; S12, uniformly transcode the collected raw environmental data streams of different encoding and resolution into a preset standard format, such as H.264 encoding, to obtain standard environmental data streams; Based on the Network Time Protocol (NTP), all standard environmental data streams are stamped with uniform and accurate time stamps to ensure the timing consistency of subsequent data analysis; S13, enhance the standard environmental data stream (especially the visual data) to obtain the preprocessed standard environmental data stream, to improve the robustness of subsequent image recognition and protect privacy; the enhancement includes: Denoising, color enhancement, and contrast adjustment of standard environmental data stream, etc. image processing algorithm to optimize image quality in insufficient light or bad weather; S14, send the preprocessed standard environmental data stream to S2; In this embodiment, by constructing a multi-source heterogeneous sensing network and a standardized data stream pipeline, high-quality collection and preprocessing of park environmental data are realized; the compatibility problem of data from different sources is solved, providing uniform and pure data input for subsequent analysis; S2, based on the edge computing node deployed at the key nodes of the park, the preprocessed standard environmental data stream is analyzed in real time, and the preliminary behavior recognition result and related feature data without credibility verification are output, providing original input for subsequent data analysis; The S2 includes the following steps: S21, send the standardized standard environment data stream to the edge computing node pre-deployed by the system; the edge computing node is deployed in the key area or important monitoring intensive area of the park, including the perimeter, entrance, road node, parking area and operation risk area of the smart park; and in the obtained video picture, one or more virtual electronic fence areas are preset according to the safety management rules of the smart park; S22, decode the standard environment data stream transmitted in S1 step, and input frame by frame to the lightweight target detection model (such as YOLOv5s); the lightweight target detection model is configured to be able to identify a plurality of preset target categories, including personnel, vehicles, safety helmets, flames and smoke; the model outputs the boundary box information, target category and detection confidence C1 of all targets in the current frame; It should be noted that the lightweight target detection model (such as YOLOv5s) needs to go through a supervised training process before use. The training process uses a park scene video dataset containing a large number of target boundary boxes and categories, and optimizes the loss function (such as mean square error loss and cross entropy loss) to make the model learn the mapping from the input image to the target position and category. The detection confidence C1 is the probability estimate value calculated by the output layer of the trained lightweight target detection model based on the learned parameters when performing forward inference, which is the direct output of the lightweight target detection model internal calculation without secondary calculation by external modules. S23, input the boundary box information, target category and detection confidence of the target into the multi-target tracking algorithm (such as DeepSORT); the algorithm assigns a unique identity ID to each detected target, and forms a continuous motion trajectory for each target through cross-frame association technology. The motion trajectory at least includes the position coordinates, motion speed and motion direction of the target in each frame; S24, rule matching degree calculation based on multi-dimensional features: Firstly, the system extracts the core features for rule matching by analyzing the target data, including: target category, detection confidence C1, identity ID, motion trajectory, position coordinates, motion speed and direction. The core features for rule matching include target state attributes and the number of relationships between targets; the target state attribute refers to the static or dynamic condition presented by the target individual at a certain time, such as whether to wear a safety helmet, whether to be in a smoking state, and the vehicle parking time; the number of relationships between targets refers to the correlation features formed by statistics among multiple targets in a certain area, such as the number of personnel and the number of vehicles in the area; Subsequently, by comparing and analyzing the real-time monitored target motion trajectory, target state attribute and quantity relationship between targets with predefined rule conditions, a rule matching degree M between 0 and 1 is calculated to quantify the degree of behavior conforming to the dangerous condition; wherein M = 0 represents that the behavior completely does not conform to the dangerous condition, and M = 1 represents that the behavior completely conforms to or exceeds the dangerous condition; The target state attribute includes whether a safety helmet is worn, whether in a smoking state, and vehicle parking time length; and the quantity relationship between targets includes the number of personnel in the region and the number of vehicles; The rule matching degree is 0 when the behavior completely does not conform to the dangerous condition, and the rule matching degree is 1 when the behavior completely conforms to or exceeds the dangerous condition. Based on this, the identification of specific dangerous behavior and the calculation of the rule matching degree are as follows: Region intrusion behavior identification: according to the motion trajectory of personnel or vehicles and the predefined electronic fence region coordinates, it is judged whether the bottom center point of the target trajectory enters the electronic fence region, and if the target continuously stays in the region for more than a preset number of frames, it is determined as a region intrusion behavior; When the target enters the electronic fence region, the rule is completely matched, so the rule matching degree M = 1.0; when the target does not enter the electronic fence region, the rule is completely unmatched, so the rule matching degree M = 0.0; when the target is continuously determined to enter (i.e. M continuously is 1.0) in N frames (for example, N = 5), a region intrusion behavior event is triggered; Personnel gathering behavior identification: according to the positions of all personnel and the predefined gathering detection region, the number of personnel in the region is counted, and if the number of personnel exceeds the threshold value dynamically set according to the region type and time period, it is determined as a personnel gathering behavior; At this time, the calculation of the rule matching degree M is as follows: The system presets a gathering threshold J1 (for example, 10 people) of the region, and the calculation formula of the rule matching degree M is: M = min (N / J, 1.0), which means that when the number of personnel reaches or exceeds the threshold, the rule matching degree M is the maximum value 1.0; when the gathering threshold J1 is not reached, M is the ratio of the current number of personnel to the threshold; N is the current number of personnel; when the current number of personnel continuously exceeds the gathering threshold, a personnel gathering behavior event is triggered; Vehicle behavior identification: according to the motion trajectory, motion speed, direction and stay time of the vehicle target, a comprehensive judgment is made; including: judging whether the low-speed stay of the vehicle in the no-parking area exceeds the allowed time length; calculating the vehicle speed by laying a virtual detection line and comparing it with the speed limit value of the region to judge whether it is speeding; analyzing the angle between the vehicle motion direction and the specified direction vector to judge whether it is reverse driving; At this time, the calculation of the rule matching degree M is as follows: The system presets a maximum allowed parking time J2 (for example, 60 seconds), and calculates the stay time T of the vehicle in the no-parking area in real time. The calculation formula of the rule matching degree M is: M = min(T / J2, 1.0). This formula indicates that the longer the vehicle stay time, the higher the rule matching degree, and the maximum value 1.0 is reached after exceeding the maximum time. When the stay time of the vehicle in the no-parking area exceeds the maximum allowed parking time, the illegal parking behavior event is triggered. Construction behavior recognition: according to the bounding box position of the personnel target and the pre-defined work area, the head region image of the personnel located in the work area is intercepted for secondary image classification to determine whether the personnel wears a safety helmet according to the regulation; At this time, the value of the rule matching degree M is determined according to the rule condition: The rule condition is not wearing a safety helmet according to the regulation. When the personnel wears a safety helmet according to the regulation, the rule matching degree M is 0. When the personnel does not wear a safety helmet, the rule matching degree M is 1. S25, encapsulate all the recognized behavior events into a unified behavior data set as the final output of the S2 step; The behavior data set includes the following fields: event unique identifier ID, dangerous behavior type (such as area intrusion, personnel gathering), associated target identity identifier ID, behavior occurrence timestamp, behavior occurrence geographic location, detection confidence, and rule matching degree; The behavior data set also includes auxiliary data for transmission to S3 for deep analysis, such as the number of personnel and the target motion speed. In this embodiment, by deploying a lightweight target detection model and a pre-defined rule library on the edge node, real-time and preliminary recognition of typical dangerous behaviors is realized. A serial processing framework of target detection-multi-target tracking-rule matching is adopted to realize real-time matching of the motion trajectory, state of the target, and rules such as electronic fence and safety specifications, and to quantitatively output the rule matching degree. The low-delay advantage of edge computing is fully utilized to ensure millisecond-level response to sudden behaviors such as intrusion and gathering. At the same time, the complex behavior recognition task is decomposed into an interpretable rule judgment to provide structured feature data for subsequent deep decision-making.

[0019] S3, based on the output behavior data set, cross-verify and evaluate the confidence of the behavior data set through a multi-dimensional confidence fusion model; The S3 includes the following steps: S31, based on the detection confidence C1 output in S2 and the rule matching degree M, calculate the comprehensive model prediction confidence C2: C2=C1*M, the comprehensive model prediction confidence C2 comprehensively reflects the initial confidence of the entire dangerous behavior recognition; it should be noted that the design of the rule matching degree M makes the value of M 0 when the behavior does not conform to the predefined rule, so that the comprehensive model prediction confidence C2 is 0, ensuring that such cases are filtered out, which effectively prevents false alarms, for example, when personnel do not enter the electronic fence area or the work personnel correctly wear safety helmets, the system will not produce dangerous behavior alarm; For evaluating the stability of the recognition result in time dimension, the system maintains a sliding time window with length N frames (for example, N=15, corresponding to about 0.5 seconds), stores the continuous recognition results of the same target and the same behavior in the window and their corresponding comprehensive model prediction confidence C2, and the time sequence consistency confidence C3 is calculated by the following formula: C3=(F / N)*(1-σ), where F is the number of frames in which the same dangerous behavior category is identified in the preliminary behavior recognition category in all frames in the sliding time window; and σ is the standard deviation of the comprehensive model prediction confidence C2 in the sliding window, used to measure the volatility of the output confidence; The value range of the time sequence consistency confidence C3 is [0, 1], and the value tends to 1, indicating that the result is stable and consistent in time, and tends to 0, indicating that the result is volatile or inconsistent; S32, the processor acquires the context information of the current recognition event in real time, including the system time, the geographic location where the event occurs, and the preliminary recognized behavior type; according to the three elements, a preset scene behavior reference library is queried; The scene behavior reference library stores the reference threshold or reference state of each type of behavior in various typical scenes; the real-time recognized behavior feature data is compared with the reference threshold or reference state obtained by querying, and a scene matching confidence between 0 and 1 is calculated according to the degree of deviation from the normal range; the higher the confidence value, the more serious the abnormality of the current behavior in the scene; For the personnel gathering behavior: the scene matching confidence C4 is calculated by the following method: Query the reference library to obtain the reference threshold T (for example: 5 people) of the current scene (such as the lunch period in the office area); Statistical the number of personnel N in the current real-time recognized area, and calculate the scene matching confidence C4: if N≤T, then C4=0.2 (the situation is normal, and the scene dimension does not provide confidence support); If T If N>2*T, then C4=0.9, the situation is seriously abnormal, and the scene dimension strongly supports this as a dangerous behavior; For regional intrusion behavior, the scene matching credibility is calculated by the following way: Query the reference library to obtain the reference intrusion frequency of the current scene (such as the core R&D area-night), combine historical data to calculate the actual intrusion frequency of the current period of the region, and output the corresponding scene matching credibility according to the comparison and analysis of the actual intrusion frequency of the current period of the region and the reference intrusion frequency; For vehicle illegal parking behavior, the scene matching credibility is calculated by the following way: Query the reference library to obtain the reference allowed parking time length (for example: 30 minutes) of the current scene (such as the logistics loading and unloading area-working time), obtain the parking time length of the current vehicle, and output the corresponding scene matching credibility according to the comparison and analysis of the parking time length of the current vehicle and the reference allowed parking time length; The scene behavior reference library ensures that its data source is clear and achievable by the following way: In the early stage of system deployment, the park administrator directly configures the initial reference threshold for different scenes according to the park management regulations and historical experience, which ensures that the system can run without historical data; Automatic learning update: after the system runs, the long-term average value of each type of behavior in each scene is continuously calculated; the average value is used to smooth the values in the reference library periodically (such as every week) to make the system reference adapt to the actual operation changes of the park; This step corrects the credibility by introducing scene prior knowledge. The system is pre-installed with a scene rule library, which at least includes time period, region type and other dimensions. The scene rule library is initialized by the administrator according to the park management regulations in the early stage of system deployment, and is automatically updated by the cloud platform through statistical historical data after the system runs; S33, the weighted sum model is adopted to fuse the credibility of the above detection confidence C1, time sequence consistency credibility C3, and scene matching credibility C4, and calculate the final comprehensive confidence C 综合 , the specific calculation is: C 综合 =α*C1+β*C3+γ*C4; Wherein, α, β, γ are weight coefficients, and α+β+γ=1, the recommended value optimized according to a large amount of experimental data is: α=0.6, β=0.3, γ=0.1, which shows that the model prediction itself is the most important basis, the time sequence stability is the second, and the scene correction is auxiliary; The comprehensive confidence C 综合 is compared with a preset global confirmation threshold (for example 0.75), and if the comprehensive confidence C 综合≥ global confirmation threshold, the dangerous behavior recognition result is confirmed to be valid, and the dangerous behavior recognition result is output and enters a subsequent alarm and disposal process.

[0020] If the comprehensive confidence C 综合〈 If the comprehensive confidence C It should be noted that the global confirmation threshold (for example, 0.75) is an optimal balance point based on a large amount of experimental data, and the system can control the false positive rate to be less than 5% while maintaining a high detection rate of greater than 95% at this threshold. A 0.85 threshold is also set to distinguish between ordinary risks and high-risk events. When the confidence exceeds this threshold, the system determines that it is a high-risk behavior and triggers a higher level of emergency response. The global confirmation threshold is dynamically adjusted according to a preset period or the number of cumulative recognitions, or the global confirmation threshold is automatically increased when the system false positive rate exceeds the preset target. When the system false negative rate exceeds the preset target, the global confirmation threshold is automatically decreased. The final output comprehensive confidence and dangerous behavior type are sent to S4. In this embodiment, a three-dimensional confidence evaluation model that integrates model prediction, time sequence consistency and scene matching is constructed to realize cross-validation and accurate decision-making of the preliminary recognition result. The core technical measure is to introduce a weighted summation algorithm to quantitatively integrate the model prediction confidence, time sequence stability evaluation and scene matching evaluation, and make a final judgment based on a global threshold. False alarms caused by environmental interference and temporary misrecognition are effectively filtered out, and the system false positive rate is significantly reduced. At the same time, the quantitative confidence objectively reflects the confidence level of the recognition result, providing a scientific basis for graded alarm and accurate disposal, and is the intelligent core of the entire system to achieve high reliability.

[0021] S4, a risk aggravation early warning signal is generated by linear regression analysis of the comprehensive confidence in the time window. Then, based on the preset dangerous behavior type, comprehensive confidence and risk aggravation early warning signal, a three-dimensional decision model is used to dynamically determine the alarm level. According to the alarm level, the corresponding graded push mechanism and preset device linkage operation are started. Finally, a digital disposal work order is created for the triggered alarm, and the disposal effect is automatically verified by continuously monitoring whether the same behavior occurs again in the incident area to realize closed-loop management. The S4 includes the following steps: S41, the system maintains a two-dimensional rectangular coordinate system with time as the horizontal axis and comprehensive confidence as the vertical axis, and continuously plots the comprehensive confidence output by S3 as a data point in the coordinate system to form a discrete point sequence evolving over time. A backtracking time window (e.g. 30 seconds) is set for trend analysis, all data points in the discrete point sequence within the window are extracted to form an initial data point set; in order to reduce environmental noise interference, the initial data point set is preprocessed: first, the median absolute deviation method is used to identify and eliminate abnormal fluctuation points; then, the exponential weighted moving average method is used to smooth the data point set after eliminating abnormal points to retain the trend while suppressing high-frequency noise, and a preprocessed data point set is obtained; The preprocessed data point set is input into a linear regression model, with time as the independent variable, and an optimal straight line equation is fitted by the least squares method, and the slope k of the fitted straight line equation is the current comprehensive confidence change slope K; wherein K>0 indicates that the comprehensive confidence is rising, indicating that the potential risk is accumulating; K<0 indicates that the comprehensive confidence is declining, indicating that the risk is dissipating; The calculated comprehensive confidence change slope K is compared with a preset positive slope threshold Kmax: if the comprehensive confidence change slope K is greater than the preset positive slope threshold Kmax, it is determined that the risk level of the currently identified dangerous behavior is rapidly rising, and the system immediately generates a risk intensification warning signal; if the comprehensive confidence change slope K is greater than zero and the comprehensive confidence change slope K is less than or equal to the preset positive slope threshold Kmax, it is determined that the risk trend of the currently identified dangerous behavior is stable or slowly rising, and the system does not trigger a risk intensification warning and continues to perform a monitoring task; if the comprehensive confidence change slope K is less than zero, it is considered that the current risk trend is stable or declining, and the system does not trigger a risk intensification warning and continues to perform a monitoring task; The warning signal is used to trigger proactive prevention and control measures, for example: starting continuous key tracking on the target, sending a "target behavior upgrade, please pay attention" prompt to security personnel, or preloading emergency resources to the area; It should be noted that the value of the positive slope threshold Kmax is determined through a systematic optimization process based on historical data; specifically: collect time series confidence data and their corresponding event true labels generated during the system trial run; analyze the warning detection rate and false alarm rate under different thresholds through ROC curve analysis; finally, select the optimal balance point according to the park safety strategy; for example, in a specific experimental environment, the Kmax=0.02 [confidence / second] used in this embodiment achieves a detection rate of >90% and a false alarm rate of <15% on the test set, and the physical meaning of this value is to identify the risk acceleration state where the confidence increases by more than 0.02 per second; S42, according to the three-dimensional decision model of the dangerous behavior type, the comprehensive confidence and the risk aggravation early warning signal, a unified decision rule base is used to dynamically determine the final alarm level, and the three core dimensions are comprehensively considered: the dangerous behavior type: the inherent risk level of the behavior itself is classified preliminarily; the comprehensive confidence: the reliability of the current recognition result is reflected; the risk aggravation early warning signal: whether the risk situation is rapidly deteriorating is indicated; The system maintains a behavior type-benchmark risk mapping table, and presets a benchmark risk level for each identifiable dangerous behavior as the basis for determining the alarm level; for example: extremely high-risk behavior (benchmark level: level 1): flame recognition, personnel falling, violent fighting and other behaviors directly threatening personal safety; high-risk behavior (benchmark level: level 2): perimeter intrusion, illegal fire operation, personnel gathering in safety area and other behaviors that may rapidly develop into serious events; general risk behavior (benchmark level: level 3): not wearing a safety helmet, vehicle illegal parking, smoking in a no-smoking area and other daily violations; On the basis of the behavior benchmark risk level, the final alarm level is obtained by fine-tuning according to the comprehensive confidence and upgrading according to the risk aggravation early warning signal; the specific judgment rules are as follows: Level 1 alarm (emergency): any dangerous behavior type, as long as the comprehensive confidence C 综合 ≥ 0.85, then directly trigger level 1 alarm; or for the behavior with a benchmark risk level of level 2 (high risk), if 0.75≤ comprehensive confidence C 综合 < 0.85, and the system generates a risk aggravation early warning signal, then it is upgraded to level 1 alarm; Level 2 alarm (important): for the behavior with a benchmark risk level of level 2 (high risk), if 0.75≤ comprehensive confidence C 综合 < 0.85, and no risk aggravation early warning signal is triggered, then it is determined as level 2 alarm; or for the behavior with a benchmark risk level of level 3 (general risk), if the comprehensive confidence C 综合 ≤ 0.75, and the system generates a risk aggravation early warning signal, then it is upgraded to level 2 alarm; Level 3 alarm (general): for the behavior with a benchmark risk level of level 3 (general risk), if the comprehensive confidence C 综合 ≤ 0.75, but no risk aggravation early warning signal is triggered, then it is determined as level 3 alarm; According to the final determined alarm level, the system generates a standardized alarm message, and the alarm message at least includes: event unique identifier ID, dangerous behavior type, final alarm level, geographic location, timestamp, comprehensive confidence, whether there is a risk aggravation early warning signal, associated video stream address and target feature description; S43, a hierarchical push mechanism is established based on the alarm level, specifically: Primary alarm: synchronously pushed to the park central command center large screen (highest priority pop-up window), relevant responsible person mobile terminal (strong reminder push) and on-site broadcast system (voice alarm); Secondary alarm: pushed to the park central command center large screen (pop-up window) and the mobile terminal of the relevant area responsible person; Tertiary alarm: pushed to the park central command center background management system (to-do work order list) and the mobile terminal of the on-site security personnel; S44, the system pre-stores an event device linkage rule library, after receiving an alarm event, automatically triggers the preset device linkage operation according to the identified event type, specifically: Perimeter intrusion event: automatically control the pan-tilt camera closest to the event point to turn to the preset monitoring position and start the target tracking algorithm; Fire access occupation event: automatically pop up the multi-angle video screen of the central command center in the area and send disposal instructions to the area security personnel; Flame recognition event: immediately start the on-site sound and light alarm and link the access control system to open the emergency evacuation passage; Personnel gathering event: automatically adjust the on-site lighting intensity and activate the people flow density monitoring function; It should be noted that the event device linkage rule library is based on the park safety plan and device layout and is configured by the administrator during system initialization; for example, according to the emergency plan, opening all evacuation passages is required in case of fire, so the rule of flame recognition event-linkage access controller-execute open door instruction is set; the setting of the rule follows the principles of proximity, urgency and matching; the proximity principle is to automatically select the closest device (such as the closest camera); the urgency principle is to determine the response range according to the event level (such as triggering all devices in the field for primary alarm); the matching principle is that the device function matches the disposal demand (such as linking the tracking camera for intrusion event and linking the alarm and access control for fire event); after the system receives the alarm, it automatically queries this rule library, generates control instructions and sends them to the corresponding devices, thereby realizing intelligent and rapid linkage response; S45, the system automatically creates a disposal work order for each confirmed alarm, and the work order status includes: to be accepted, in disposal, disposed, and verified closed; the disposal personnel update the work order status in real time through the mobile terminal and can upload text description or on-site pictures as disposal credentials; When the work order status is marked as disposed, the system automatically starts the verification mechanism: Call the analysis model of S3 to continuously monitor the incident area (the monitoring time can be configured, such as 60-300 seconds); If the same dangerous behavior is not detected again within the monitoring period, the work order status is automatically updated to verified closed; If the same behavior is detected again during the monitoring period, the alarm is regenerated and the alarm level is raised; All verification results are recorded in the database for subsequent system optimization analysis. In this embodiment, by early warning of the trend of dangerous behavior and establishing a hierarchical alarm-device linkage-closed loop verification disposal mechanism, the whole process automation management from risk identification to elimination is realized; specifically, by linear regression analysis of the comprehensive confidence in the time window, a risk aggravation warning signal is generated; then, based on the preset dangerous behavior type, comprehensive confidence and risk aggravation warning signal, a three-dimensional decision model is used to dynamically determine the alarm level; according to the alarm level, the corresponding hierarchical push mechanism and the preset device linkage operation are started; finally, a digital disposal work order is created for the triggered alarm, and the disposal effect is automatically verified by continuously monitoring whether the same behavior occurs again in the incident area to realize closed loop management; this measure realizes the trend warning, rapid response and standardized disposal of safety events, greatly improves the emergency efficiency; and through continuous monitoring and state verification after disposal, a management closed loop of self-checking and optimization is formed, and finally the traditional passive monitoring is upgraded to an active, precise and traceable modern security system.

[0022] The present application aims to overcome the inherent defects of traditional park security systems, such as response lag, high false alarm rate, system isolation and lack of prediction ability, and follows the intelligent closed loop logic of perception-cognition-decision-action, and realizes the fundamental change from passive monitoring to active prevention and control through the construction of a hierarchical and collaborative technical architecture. Firstly, a standardized multi-source data acquisition and preprocessing process is established to provide high-quality and consistent input data for the system; the most innovative design is that a complete multi-dimensional confidence fusion decision mechanism is integrated in the edge computing layer; this system does not simply rely on the results of a single algorithm, but innovatively calculates the model prediction confidence, time consistency confidence and scene matching confidence on the edge side in parallel, and makes a comprehensive judgment through a weighted fusion model; this breaks through the limitation of traditional edge computing which only completes simple identification, so that the edge node not only guarantees real-time response capability, but also has cloud-level analysis and judgment accuracy; further, by early warning of the trend of dangerous behavior and establishing an automatic disposal mechanism of alarm grading and device linkage, as well as a complete data closed loop from alarm to verification, the continuous tracking, resource preposition and key prompt of high-risk targets are formed, and the real active security is realized; this concept effectively solves the industry problem that response speed and recognition accuracy are difficult to balance, and realizes a major breakthrough in completing complex cognitive decision on the edge side.

[0023] The above merely illustrates and describes the structure of the present application, and those skilled in the art can make various modifications or supplements to the described specific embodiments or adopt similar ways to replace, as long as the modifications or supplements do not deviate from the structure of the present application or exceed the scope defined by the present claims, and should belong to the protection scope of the present application.

Claims

1. A method for monitoring and identifying dangerous behaviors in a smart park, characterized in that, Includes the following steps; S1. Collect environmental data streams by sensing devices deployed at key locations in the park, and standardize the environmental data streams to obtain standard environmental data streams; S2. Send the standard environment data stream to the edge computing node, and perform the following operations on the edge computing node: A lightweight target detection model is used to identify multiple preset target categories in a standard environmental data stream and obtain the target detection confidence. The target categories include people, vehicles, safety helmets, flames, and smoke. A multi-target tracking algorithm is used to assign a unique identifier to the target and generate a continuous motion trajectory. Based on a predefined rule base, the movement trajectory of the target is matched with preset electronic fence areas, gathering thresholds, illegal stay durations or safety regulations, the rule matching degree is calculated, and a behavior dataset is output based on this. S3. Based on the detection confidence and rule matching degree, the comprehensive model prediction confidence is calculated. Through temporal consistency analysis, the stability of the comprehensive model prediction confidence is evaluated within the sliding time window, and the temporal consistency confidence is calculated. The scene matching confidence is obtained by matching the output behavior dataset with the scene behavior benchmark library. Then, the detection confidence, temporal consistency confidence, and scene matching confidence are weighted and fused to generate the final comprehensive confidence. S4. By performing linear regression analysis on the comprehensive confidence level within the time window, an early warning signal of increased risk is generated and an alarm is triggered.

2. The method for monitoring and identifying dangerous behaviors in a smart park according to claim 1, characterized in that, S1 specifically includes: By deploying various sensing devices at key locations in the park's geospatial area, raw environmental data streams are continuously collected. These raw environmental data streams include visual data streams, thermal imaging data streams, and IoT sensor signals. The collected raw environmental data streams are uniformly transcoded into a preset standard format, and a uniform timestamp is added to all standard environmental data streams; Enhancement processing is applied to standard environment data streams, including noise reduction, color enhancement, and contrast adjustment; The preprocessed standard environment data stream is sent to S2.

3. The method for monitoring and identifying dangerous behaviors in a smart park according to claim 1, characterized in that, S2 specifically includes: The standardized environmental data stream is sent to the pre-deployed edge computing nodes of the system; one or more virtual electronic fence areas are preset in the video image obtained by decoding the standardized environmental data stream; The standard environment data stream is decoded into image frames and input into the lightweight object detection model frame by frame, outputting the bounding box information, object category and detection confidence of all objects in the current frame; The bounding box information, target category, and detection confidence of the target are input into the multi-target tracking algorithm. A unique identifier (ID) is assigned to each target, and a continuous motion trajectory is formed for each target. The motion trajectory includes at least the target's position coordinates, motion speed, and motion direction. Based on the output target category, detection confidence, target identification ID and movement trajectory, preliminary identification of specific dangerous behaviors is performed according to a predefined rule base, and the rule matching degree is calculated; All identified behavioral events are encapsulated into a unified behavioral dataset, which serves as the final output of step S2.

4. The method for monitoring and identifying dangerous behaviors in a smart park according to claim 1, characterized in that, The reliability of the predictions made by the integrated model is specifically as follows: The product of detection confidence and rule matching degree is the comprehensive model prediction confidence, which is used to comprehensively reflect the accuracy of target detection and the degree of compliance with behavioral rules.

5. The method for monitoring and identifying dangerous behaviors in a smart park according to claim 1, characterized in that, The credibility of the scene matching is specifically as follows: Based on the current system time, event location, and behavior type, query the preset scenario behavior benchmark library to obtain the corresponding benchmark threshold; The real-time identified data is compared with a benchmark threshold, and different scene matching credibility is assigned according to the different intervals in which the comparison results are located.

6. The method for monitoring and identifying dangerous behaviors in a smart park according to claim 1, characterized in that, The reliability of the timing consistency is specifically as follows: Based on the video frame sequence obtained by decoding the standard environment data stream, a sliding time window of length N frames is defined. This window covers the video frame currently being processed, collectively referred to as the current video frame, and the previous N-1 consecutive video frames, collectively referred to as the historical video frames. Within a sliding time window of a preset duration, count the number of frames in all historical video frames whose preliminarily identified dangerous behavior type is the same as the dangerous behavior type in the current video frame, and calculate the proportion of such frames in the total number of frames N in the window. Calculate the standard deviation of the confidence level of the comprehensive model prediction within the time window, and normalize the standard deviation according to the preset confidence level standard deviation reference range to obtain the normalized comprehensive model prediction confidence level standard deviation; then determine the volatility adjustment coefficient based on the normalized comprehensive model prediction confidence level standard deviation through numerical mapping rules, wherein the volatility adjustment coefficient is negatively correlated with the normalized comprehensive model prediction confidence level standard deviation. The percentage of historical frames within the window that match the current frame's recognition result is multiplied by an adjustment coefficient based on the volatility of prediction confidence using a comprehensive model. The product of these factors is then used as the temporal consistency confidence.

7. The method for monitoring and identifying dangerous behaviors in a smart park according to claim 4, characterized in that, The rule matching degree is specifically as follows: Based on the real-time monitored target motion trajectory, target state attributes, and quantitative relationships between targets, the data is compared and analyzed with predefined rule conditions, and mapped to a rule matching degree value between 0 and 1 according to the degree to which it meets the dangerous behavior conditions. The target status attributes include whether a helmet is being worn, whether someone is smoking, and the duration the vehicle has been parked; the quantitative relationships between targets include the number of people and vehicles in the area. Specifically, when a behavior does not meet the dangerous conditions at all, the rule matching degree is 0; when a behavior fully meets or exceeds the dangerous conditions, the rule matching degree is 1.

8. The method for monitoring and identifying dangerous behaviors in a smart park according to claim 1, characterized in that, After the risk escalation warning signal is generated, the alarm level is dynamically determined through a three-dimensional decision model based on the preset dangerous behavior type, comprehensive confidence level and risk escalation warning signal; according to the alarm level, the corresponding graded push mechanism and preset equipment linkage operation are activated; finally, a digital handling work order is created for the triggered alarm, and the handling effect is automatically verified by continuously monitoring whether the same behavior occurs again in the incident area to achieve closed-loop management. The alarm level is determined as follows: A behavior type-benchmark risk mapping table is pre-set, and a benchmark risk level is set for each dangerous behavior; based on the benchmark risk level, the level is fine-tuned in combination with the real-time comprehensive confidence level, and the alarm level is upgraded according to the risk aggravation warning signal.

9. The method for monitoring and identifying dangerous behaviors in a smart park according to claim 8, characterized in that, The device linkage operation is executed by querying a pre-set event device linkage rule base. The rule base is set according to the principles of proximity, urgency, and matching.

10. The method for monitoring and identifying dangerous behaviors in a smart park according to claim 9, characterized in that, The risk escalation warning signal is generated through the following steps: The system maintains a two-dimensional rectangular coordinate system with time as the horizontal axis and comprehensive confidence as the vertical axis, and continuously plots the comprehensive confidence as data points in the two-dimensional rectangular coordinate system to form a discrete point sequence that evolves over time. Set a backtracking time window, extract discrete points within the window to form an initial data point set; perform data preprocessing on the initial data point set, the data preprocessing includes identifying and removing abnormal fluctuation points using the median absolute deviation method, and smoothing the data sequence after removing abnormal points using the exponential weighted moving average method to obtain the preprocessed data point set; The preprocessed set of data points is input into a linear regression model, and time is used as the independent variable for fitting. The slope of the fitted line is used as the slope of the change in the overall confidence level at the current moment. The slope of the overall confidence level change is compared with a preset positive slope threshold. If the slope of the overall confidence level change is greater than the preset positive slope threshold, the risk level is determined to be rising rapidly, and the system immediately generates and outputs a risk aggravation warning signal. The positive slope threshold is determined by collecting a sample set of time-series confidence data and corresponding event real labels generated during the historical operation of the system. Based on the sample set, the detection rate and false alarm rate of the warning under different candidate thresholds are analyzed by receiver operation characteristic curve. According to a predetermined detection rate and false alarm rate balancing strategy, the optimal value is selected from the candidate thresholds as the positive slope threshold.

Images