Child safety analysis method and system based on monitoring data

Abstract

The present invention discloses a child safety analysis method and system based on monitoring data, and relates to the technical field of electronic digital data processing. The child safety analysis method based on monitoring data includes the following steps: data segmentation; recognition validity analysis; and detection image generation. The present invention performs data segmentation on a feature data set formed by skeleton monitoring data, and then determines whether to perform dangerous posture recognition based on the data segmentation result. If so, recognition validity analysis is performed; otherwise, windowing operation parameter optimization is performed. Finally, based on the recognition validity analysis result, it is determined whether to generate a behavior detection image. If so, the generated behavior detection image is visualized; otherwise, the model learning rate is optimized. This achieves the effect of improving the efficiency of child behavior recognition based on child safety monitoring data, and solves the problem of low posture recognition accuracy of pose posture detection models in the process of child behavior recognition based on child safety monitoring data in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic digital data processing, and in particular to a child safety analysis method and system based on monitoring data. Background Art

[0002] Implementing skeleton-based posture recognition in practical applications is a difficult task because it involves multiple modules, such as person detection and posture estimation. Skeleton-based methods have great robustness advantages in understanding actual human behavior. The increasing maturity of motion target detection and motion target tracking in computer vision technology, as well as the development of artificial intelligence, have laid a solid foundation for the detection of abnormal behavior of moving targets and provided more possibilities for the detection of abnormal behavior of moving targets. However, due to the particularity and uncertainty of young children's behavior and the physical, age, and psychological characteristics of young children themselves, general abnormal behavior detection algorithms often fail to achieve good results. In addition, the recognition efficiency or accuracy of existing recognition algorithms are low, and good results cannot be achieved.

[0003] For example, the invention patent announcement with announcement number: CN111143182B discloses a method, device and storage medium for analyzing process behavior, including: obtaining monitoring data for monitoring multiple application programming interfaces (APIs) called by a preset process; encoding the monitoring data corresponding to the current API according to the relative change in the corresponding values ​​of the attributes associated with the current API called by the preset process and the previous API, and the storage type of the corresponding parameters of the current API, to obtain an API record of the preset process calling the current API, and writing the API record corresponding to the current API into a behavior record file; when performing behavior analysis on the preset process, reading and parsing each API record from the behavior record file, and statistically analyzing the process behavior of the preset process based on the parsed API records to obtain a process behavior analysis result of the preset process.

[0004] For example, the patent application with publication number CN119397445A discloses a method and system for patient behavior monitoring based on big data analysis, including: collecting physiological characteristic data and behavioral characteristic data of patients; preprocessing the collected physiological characteristic data and behavioral characteristic data; calculating physiological characteristic abnormality values ​​based on the preprocessed physiological characteristic data; and performing behavioral pattern analysis based on the preprocessed behavioral characteristic data and physiological characteristic abnormality values.

[0005] In existing technologies, target detection and motion recognition algorithms based on deep learning analyze children's behavior patterns in video streams and automatically identify abnormalities such as crying, pushing, and prolonged stillness. At the same time, by comparing and analyzing monitoring data from different time periods, it can identify high-incidence periods for accidents (such as lunch time and before school) and deploy supervision in advance.

[0006] However, in the process of implementing the technical solutions of the invention in the embodiments of the present application, the present application found that the above technology has at least the following technical problems:

[0007] In the existing technology, for most real-life videos in standard datasets, human posture is not easy to detect (i.e., only partially visible or obscured by other objects), and most existing posture detection models are unable to detect the movements of people during falling motion. There is a problem of low posture recognition accuracy of pose detection models in the process of identifying children's behavior based on child safety monitoring data. Summary of the Invention

[0008] The present invention solves the problem of low posture recognition accuracy of pose posture detection models in the process of identifying children's behavior based on children's safety monitoring data in the prior art by providing a child safety analysis method and system based on child safety monitoring data, thereby improving the efficiency of child behavior recognition based on child safety monitoring data.

[0009] The present invention provides a child safety analysis method based on monitoring data, comprising the following steps: step 1, performing data segmentation on a feature data set formed by skeleton monitoring data to obtain a data segmentation result, the skeleton monitoring data being used to visualize the skeleton information of the child in the target monitoring area, and the data segmentation being used to generate a multidimensional data list based on the result of the skeleton monitoring data segmentation; step 2, judging whether to perform dangerous posture recognition based on the data segmentation result at the end of a preset segmentation period, and if so, performing recognition validity analysis based on the dangerous posture recognition result at the end of the preset recognition period to obtain a recognition validity analysis result, and otherwise performing windowing operation parameter optimization, and the recognition validity analysis being used to quantify the validity of a constructed pose posture detection model in performing dangerous posture recognition on the skeleton monitoring data, and the windowing operation parameter optimization means improving the prediction performance of the pose posture detection model by adjusting the windowing operation parameters; step 3, judging whether to generate a behavior detection image based on the recognition validity analysis result, and if so, visualizing the generated behavior detection image, and otherwise performing model learning rate optimization, and the model learning rate optimization means improving the response rate of the pose posture detection model by adjusting the initial learning rate.

[0010] The present invention provides a child safety analysis system based on monitoring data, comprising: a data partitioning module, a recognition effectiveness analysis module, and a detection image generation module; wherein the data partitioning module is used to perform data partitioning on a feature data set formed by skeleton monitoring data to obtain a data partitioning result, the skeleton monitoring data is used to visualize the skeleton information of the child in the target monitoring area, and the data partitioning is used to generate a multidimensional data list based on the result of the skeleton monitoring data segmentation; the recognition effectiveness analysis module is used to determine whether to perform dangerous posture recognition based on the data partitioning result at the end of a preset partitioning period, and if so, perform recognition effectiveness analysis based on the dangerous posture recognition result at the end of the preset recognition period to obtain a recognition effectiveness analysis result; otherwise, perform windowing operation parameter optimization, and the recognition effectiveness analysis is used to quantify the effectiveness of a constructed pose posture detection model in performing dangerous posture recognition on the skeleton monitoring data, and the windowing operation parameter optimization means improving the prediction performance of the pose posture detection model by adjusting the windowing operation parameters; the detection image generation module is used to determine whether to generate a behavior detection image based on the recognition effectiveness analysis result, and if so, visualize the generated behavior detection image, and otherwise perform model learning rate optimization, and the model learning rate optimization means improving the response rate of the pose posture detection model by adjusting the initial learning rate.

[0011] One or more technical solutions provided in the present invention have at least the following technical effects or advantages:

[0012] 1. Data division is performed on the feature data set formed by skeleton monitoring data, and then the data division result at the end of the preset division period is used to determine whether dangerous posture recognition is to be performed. If so, recognition validity analysis is performed, otherwise windowing operation parameters are optimized. Finally, based on the recognition validity analysis result, it is determined whether a behavior detection image is to be generated. If so, the generated behavior detection image is visualized, otherwise the model learning rate is optimized, thereby realizing real-time and efficient processing of child safety monitoring data, and further realizing the improvement of the efficiency of child behavior recognition based on child safety monitoring data, effectively solving the problem of low posture recognition accuracy of pose posture detection model in the process of child behavior recognition based on child safety monitoring data in the existing technology.

[0013] 2. By dynamically adjusting the sliding window length and sliding window step size, the preset posture prediction model can better adapt to the changes in the skeleton monitoring data, improve the prediction performance of the preset posture prediction model for the posture of children in the target monitoring area, and thus improve the real-time performance of children's safety monitoring. At the same time, by sending alarm instructions and prompting preset personnel to intervene, it is ensured that problems arising during the optimization process can be handled in a timely manner, thereby improving the accuracy and efficiency of children's behavior recognition in the target monitoring area.

[0014] 3. When the obtained total input and output duration is within the allowable range of the total input and output duration in the database, the difference between the obtained input response duration and the reference input response duration in the database is compensated by the input response duration compensation value to obtain the input response duration score. At the same time, the obtained input response duration score, output response duration score and input and output response duration score are coupled to obtain the recognition validity interference value, thereby improving the accuracy of obtaining the recognition validity interference value, and further achieving a more accurate evaluation of the recognition validity of the pose posture detection model within the preset recognition period. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flowchart of a child safety analysis method based on monitoring data provided in an embodiment of the present application;

[0016] Figure 2 A framework diagram of infant skeleton monitoring and posture recognition provided in an embodiment of the present application;

[0017] Figure 3 A flowchart of infant skeleton monitoring and posture recognition provided in an embodiment of the present application;

[0018] Figure 4 This is a schematic diagram of the structure of a child safety analysis system based on monitoring data provided in an embodiment of the present application. DETAILED DESCRIPTION

[0019] The embodiments of the present application solve the problem of low posture recognition accuracy of pose posture detection models in the process of identifying child behavior based on child safety monitoring data in the prior art by providing a child safety analysis method and system based on monitoring data. The method performs data division on the feature data set formed by skeleton monitoring data to obtain a data division result, and then determines the model prediction performance evaluation value based on the data division result at the end of a preset division period. At the same time, based on the obtained model prediction evaluation value, it is determined whether dangerous posture recognition is performed through the constructed pose posture detection model. If so, a recognition validity analysis is performed based on the dangerous posture recognition result at the end of the preset recognition period to obtain a recognition validity analysis result. Otherwise, windowing operation parameters are optimized. Finally, based on the recognition validity analysis result, it is determined whether a behavior detection image is generated. If so, the generated behavior detection image is visualized. Otherwise, the model learning rate is optimized, thereby improving the efficiency of child behavior recognition based on child safety monitoring data.

[0020] The technical solution in the embodiment of the present application is to solve the problem of low posture recognition accuracy of the pose posture detection model in the process of identifying child behavior based on child safety monitoring data. The overall idea is as follows:

[0021] The feature data set formed by skeleton monitoring data is divided into data, and then the data division results at the end of the preset division period are used to determine whether dangerous posture recognition is to be performed. If so, the recognition validity analysis is performed, otherwise the windowing operation parameters are optimized. Finally, based on the recognition validity analysis results, it is determined whether a behavior detection image is to be generated. If so, the generated behavior detection image is visualized, otherwise the model learning rate is optimized, thereby achieving the effect of improving the efficiency of child behavior recognition based on child safety monitoring data.

[0022] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0023] like Figure 1 FIG. 1 is a flowchart of a method for analyzing child safety based on monitoring data according to an embodiment of the present application. The method for analyzing child safety based on monitoring data according to an embodiment of the present application includes the following steps:

[0024] Step one, perform data division on the feature data set formed by the skeleton monitoring data to obtain the data division result. The skeleton monitoring data is used to visualize the skeleton information of the children in the target monitoring area. The skeleton information of the children includes the position of the bones, motion trajectory and posture characteristics of the children. The skeleton monitoring data represents the result of denoising the data points in the kindergarten monitoring image collected in the target monitoring area, and serves as the input of the preset posture prediction model. The preset posture prediction model represents a prediction model that performs posture estimation and motion trajectory analysis based on the acquired child skeleton information after pre-training. The preset posture prediction model in this application refers to the YOLOv8 model, which is the latest version of the YOLO (You Only Look Once) series, which can more quickly and accurately detect multiple data points in the kindergarten monitoring image. The feature data set represents the result of feature fusion and is combined with time series information to form a data set with spatiotemporal characteristics. The result of feature fusion represents the coordinates corresponding to the data feature points output by the preset posture prediction model.

[0025] Among them, data partitioning is used to generate a multidimensional data list from the results of skeleton monitoring data segmentation. The multidimensional data list stores the dimensional information corresponding to the child's skeleton information in a structured form, such as the child's bone position coordinates, posture category (such as standing, sitting), and movement speed; the following five behaviors are customized as dangerous behaviors: fighting, choking while eating, on the edge of stairs, crying, and falling. The training data and test data after data partitioning are both multidimensional data lists. Each piece of data needs to be segmented, that is, windowed. The five dangerous behaviors are usually presented in the multidimensional data list using binary. For example, if a fight is detected, the position corresponding to the fight in the multidimensional data list is recorded as 1, otherwise it is recorded as 0.

[0026] Step 2: Based on the data segmentation results at the end of the preset segmentation period, determine whether to perform dangerous posture recognition. If so, perform recognition validity analysis based on the dangerous posture recognition results at the end of the preset recognition period to obtain the recognition validity analysis result. Otherwise, perform windowing operation parameter optimization. Recognition validity analysis is used to quantify the effectiveness of the constructed pose posture detection model in dangerous posture recognition of skeleton monitoring data. Windowing operation parameter optimization means adjusting the windowing operation parameters to improve the prediction performance of the pose posture detection model. Pose posture detection (human key point detection) models are commonly used in the field of computer vision. The dangerous posture recognition result represents the recognition validity data obtained by the pose posture detection model at the end of the preset recognition period. The recognition validity data includes input response time, output response time, and total input and output time. The input response time is used to quantify the length of time the pose posture detection model receives training data and successfully inputs it. The output response time is used to quantify the length of time from the pose posture detection model starting to output posture estimation results to the end of successful output. The total input and output time is used to quantify the length of time the pose posture detection model actually works within the preset recognition period.

[0027] Step 3: Based on the results of the recognition effectiveness analysis, determine whether to generate a behavior detection image. If so, visualize the generated behavior detection image. Otherwise, optimize the model learning rate. Model learning rate optimization means adjusting the initial learning rate to improve the response rate of the pose detection model.

[0028] In this embodiment, if Figure 2 As shown in FIG, a framework diagram of the child skeleton monitoring and posture recognition provided by the embodiment of the present application is shown. Figure 3 As shown, the flowchart of the child skeleton monitoring and posture recognition provided in the embodiment of the present application can ensure the recognition effectiveness of the pose posture detection model in the child safety analysis method based on monitoring data by obtaining the total input and output time, determining whether the convergence is qualified, and taking corresponding adjustment measures. This method not only improves the recognition efficiency of the model, but also enhances the generalization ability of the model, providing a more reliable and efficient monitoring method for child safety.

[0029] Furthermore, whether to perform dangerous posture recognition is determined based on the data division result at the end of the preset division period. The specific process includes: when the obtained model prediction performance evaluation value is not greater than the model prediction performance evaluation value preset in the database, the data division result at the end of the preset division period is recorded as qualified data division and dangerous posture recognition is performed. Dangerous posture recognition is performed by a pose posture detection model trained with training data in the classification data. The classification data includes training data and test data obtained from the data division results corresponding to qualified data division. The pose posture detection model represents a classification model with behavior recognition function obtained by inputting the skeleton image of the child in the kindergarten surveillance image into the ST-GCN network. During the training of the ST-GCN network: the training data of the classifier mainly comes from the surveillance video data provided by a kindergarten institution, and customized dangerous behavior videos are shot by themselves as abnormal data to enhance data diversity. The data source is legal and meets the project requirements. For the collected data, the useless parts with messy, blurred, and obscured content are deleted to filter out clear and valid data.

[0030] When the obtained model prediction performance evaluation value is greater than the model prediction performance evaluation value preset in the database, the data division result at the end of the preset division period is recorded as unqualified data division and the windowing operation parameters are optimized. The model prediction performance evaluation value represents the harmonic average result of the multidimensional data list generation time and the skeleton monitoring data segmentation time within the preset division period. The harmonic average is used to eliminate the interference of extreme values ​​in the multidimensional data list generation time and the skeleton monitoring data segmentation time on the preset posture prediction model prediction performance evaluation process. The windowing operation parameters include the sliding window length and the sliding window step size.

[0031] The aforementioned database is a database for storing various types of set data established before the design of the child safety analysis method based on monitoring data. The database includes but is not limited to preset model prediction performance evaluation values, preset recognition validity interference values, preset division time periods, and preset recognition time periods. The various numerical values ​​therein are directly set by technical personnel. Among them, the setting basis of the preset recognition validity interference value can be determined according to the actual application scenario of the pose posture detection model. For example, the preset recognition validity interference value is represented by the result of summing and averaging the historical recognition validity interference values ​​of the pose posture detection model in the database within the historical recognition time period. In addition, the various numerical values ​​in the database can be set and fine-tuned by technical personnel according to actual debugging.

[0032] In this embodiment, dangerous posture recognition is performed when the data division is qualified, which can ensure the reliability of the recognition results. When the data division is unqualified, optimization operations are performed, which can continuously improve the prediction performance of the model and lay a solid foundation for subsequent monitoring and recognition work. At the same time, the use of harmonic mean helps to reduce the interference of extreme values ​​on model evaluation, making the evaluation results more objective and accurate, thereby effectively improving the accuracy and stability of the pose posture detection model.

[0033] Furthermore, the specific steps of optimizing the windowing operation parameters include: inputting the obtained model prediction performance evaluation value deviation into the offline reinforcement learning algorithm of the preset posture prediction model to output the actual increase in the sliding window length, the model prediction performance evaluation value deviation is used to quantify the degree of difference between the obtained model prediction performance evaluation value and the preset model prediction performance evaluation value, that is, the difference between the obtained model prediction performance evaluation value and the preset model prediction performance evaluation value, the preset model prediction performance evaluation value represents the difference between the average value of the multidimensional data list generation time and the skeleton monitoring data segmentation time within the preset division time period and the reference average value in the corresponding database, the reference average value represents the historical average value of the historical multidimensional data list generation time and the historical skeleton monitoring data segmentation time within the historical division time period in the database.

[0034] After the sliding window length is increased once, the first model prediction performance evaluation value deviation obtained is re-input into the offline reinforcement learning algorithm of the preset posture prediction model to output the actual reduction in the sliding window step size. The first model prediction performance evaluation value deviation represents the difference between the model prediction performance evaluation value obtained again after the sliding window length is increased once and the preset model prediction performance evaluation value.

[0035] If the reduction amplitude corresponding to the deviation of the second model prediction performance evaluation value obtained after one windowing operation parameter optimization is not greater than the reference reduction amplitude in the corresponding database, an alarm instruction is sent, otherwise the windowing operation parameter optimization is continued. If the deviation of the second model prediction performance evaluation value obtained by the windowing operation parameter optimization within the preset optimization times is not greater than 0, the windowing operation parameter optimization is completed, otherwise the preset personnel are prompted to intervene. The second model prediction performance evaluation value deviation represents the difference between the model prediction performance evaluation value obtained again after one windowing operation parameter optimization and the preset model prediction performance evaluation value. One windowing operation parameter optimization includes one increase in the sliding window length and one decrease in the sliding window step size.

[0036] In this embodiment, the sliding window length determines the prediction accuracy of the preset posture prediction model for the skeleton monitoring data, but the increase in the sliding window length may increase the amount of calculation. Therefore, when the sliding window length increases, it must be accompanied by a decrease in the sliding window step size, because the sliding window step size determines the update frequency of the preset posture prediction model for the skeleton monitoring data. The reduction in the sliding window step size means that the preset posture prediction model will update the prediction results more frequently, thereby improving real-time performance.

[0037] Therefore, by adjusting the sliding window step size, a balance can be found between prediction accuracy and real-time performance, so that the pose detection model can better adapt to different monitoring scenarios and data characteristics, thereby improving prediction performance. The optimized model can more accurately identify dangerous postures of young children, reduce false alarms and missed alarms, improve the reliability of the monitoring system, and thus improve the operating efficiency of the entire monitoring system.

[0038] Furthermore, an identification validity analysis is performed, which also includes determining whether to perform an identification validity analysis based on the obtained total input and output time. The specific process is: when the obtained total input and output time is within the allowable range of the total input and output time in the database, it is determined that the pose posture detection model has converged successfully and an identification validity analysis is performed; when the obtained total input and output time is not within the allowable range of the total input and output time in the database, it is determined that the pose posture detection model has failed to converge successfully and the regularization item is adjusted. The allowable range of the total input and output time represents the range corresponding to the maximum and minimum values ​​of the historical input and output response time of the pose posture detection model in the database during the historical recognition period.

[0039] Among them, the specific process of adjusting the regularization term is: when the acquired total input and output time is greater than the maximum allowable total input and output time in the database (that is, the maximum value of the historical input and output response time), it is determined that the pose posture detection model is overfitting and the regularization term of the pose posture detection model is increased based on the acquired regularization term increase amplitude until the re-acquired total input and output time is within the allowable range of the total input and output time in the database. The increase amplitude of the regularization term represents the result obtained by mapping the difference between the acquired total input and output time and the maximum allowable total input and output time in the database.

[0040] When the acquired total input-output duration is less than the minimum allowable total input-output duration in the database (i.e., the minimum value of the historical input-output response duration), the pose posture detection model is determined to be underfitting and the regularization term of the pose posture detection model is reduced based on the obtained regularization term reduction amplitude until the re-acquired total input-output duration is within the allowable range of the total input-output duration in the database. The reduction amplitude of the regularization term represents the result obtained by mapping the difference between the minimum allowable total input-output duration and the acquired total input-output duration in the database.

[0041] In this embodiment, by introducing a judgment process based on the total input and output time and a regularization term adjustment mechanism, the convergence of the pose posture detection model can be more effectively evaluated, and targeted adjustments can be made according to the fitting state of the pose posture detection model. This can not only improve the recognition accuracy of the pose posture detection model, but also optimize the response speed of the pose posture detection model, thereby further improving the overall performance of child skeleton monitoring and dangerous posture recognition.

[0042] Furthermore, an analysis of recognition effectiveness is performed based on the dangerous posture recognition results at the end of a preset recognition period. The specific process includes: when the obtained total input and output time is within the allowable range of the total input and output time in the database, first, the difference between the obtained input response time and the reference input response time in the database is compensated by the input response time compensation value to obtain the input response time score. The specific restriction expression of the input response time score SYX1 is: SYX1=y1×Y1 / Y10, where SYX1 represents the input response time score of the pose posture detection model in the preset recognition period, y1 represents the input response time compensation value, Y1 represents the input response time of the pose posture detection model in the preset recognition period, and Y10 represents the reference input response time. The reference input response time is represented by the result of summing and averaging the historical input response times of the pose posture detection model in the historical recognition period in the database.

[0043] Then, the output response time compensation value is used to compensate for the difference between the obtained output response time and the reference output response time in the database to obtain the output response time score. The specific restricted expression of the output response time score SYX2 is: SYX2=y2×Y2 / Y20, where SYX2 represents the output response time score of the pose posture detection model within the preset recognition period, y2 represents the output response time compensation value, Y2 represents the output response time of the pose posture detection model within the preset recognition period, and Y20 represents the reference output response time. The reference output response time is represented by the sum and average of the historical output response times of the pose posture detection model in the database within the historical recognition period.

[0044] Next, the difference between the obtained input-output response time and the reference input-output response time in the database is compensated by the input-output response time compensation value to obtain the input-output response time score. The specific restriction expression of the input-output response time score SYX3 is: SYX3=y3×Y3 / Y30, where SYX3 represents the input-output response time score of the pose posture detection model within the preset recognition period, y3 represents the input-output response time compensation value, Y3 represents the input-output response time of the pose posture detection model within the preset recognition period, and Y30 represents the reference input-output response time. The reference input-output response time is represented by the sum and average of the historical input-output response times of the pose posture detection model in the historical recognition period in the database.

[0045] Finally, the obtained input response time score, output response time score and input-output response time score are coupled to obtain the recognition effectiveness interference value. The recognition effectiveness interference value represents the quantitative data of the degree of influence of the input response time, output response time and input-output response time on the recognition effectiveness of the pose posture detection model. The specific restriction expression of the recognition effectiveness interference value SYX is: SYX=SYX1+SYX2+SYX3, where SYX represents the recognition effectiveness interference value of the pose posture detection model within the preset recognition period.

[0046] Among them, the input response time, output response time, and input-output response time are all recorded in real time by time sensors, and the units are the same as the reference input response time, reference output response time, and reference input-output response time, which are all milliseconds (ms).

[0047] The input response time compensation value, output response time compensation value, and input / output response time compensation value represent the degree of influence of the input response time, output response time, and input / output response time, respectively, as pre-set in the database, on the pose detection model recognition process. Specifically, the database stores preset compensation values ​​corresponding to the input response time, output response time, and input / output response time. These compensation values ​​are mapped to the input response time, output response time, and input / output response time, respectively. This mapping can be one-to-one or many-to-one. For example, in practical applications, the real-time input response time, output response time, and input / output response time can be input into this mapping to quickly obtain the corresponding compensation value. In this example, the input response time compensation value, output response time compensation value, and input / output response time compensation value typically range from 0 to 1, and the sum of the three is 1.

[0048] In this embodiment, the recognition validity interference value increases with the increase of input response time, output response time, and input and output response time. Among them, the increase in input response time may cause the system to accumulate more delays when processing initial data, thereby indirectly affecting the timeliness of the output response. The extension of the output response time may have a counter-effect on the efficiency of the input response. For example, the accumulation of input data is aggravated due to the hysteresis of the feedback mechanism. The synchronous growth of the input and output response times may form a compound delay effect, causing the overall response capability of the system to show a nonlinear attenuation trend.

[0049] By considering the above-mentioned mutual influence mechanism, it is helpful to establish a multi-dimensional temporal response model, quantify the coupling strength of each response duration parameter, and optimize the data flow scheduling strategy through a dynamic weight distribution algorithm, thereby improving the efficiency of child behavior recognition based on child safety monitoring data, and effectively solving the problem of low posture recognition accuracy of pose posture detection models in the process of child behavior recognition based on child safety monitoring data in the existing technology.

[0050] Furthermore, the specific process of model learning rate optimization is as follows: mapping the obtained recognition validity interference value deviation in the database to obtain the initial learning rate reduction amplitude of the pose posture detection model, the recognition validity interference value deviation is used to quantify the degree of difference between the obtained recognition validity interference value and the recognition validity interference value preset in the database, that is, the difference between the obtained recognition validity interference value and the recognition validity interference value preset in the database; the obtained initial learning rate reduction amplitude and the preset initial learning rate reduction amplitude are jointly input into the linear regression algorithm of the pose posture detection model to output the actual initial learning rate reduction amplitude, the actual initial learning rate reduction amplitude is used to ensure that the pose posture detection model improves the response rate of the pose posture detection model to the feature data set when the convergence is qualified; if the recognition validity interference value deviation obtained again after the preset number of initial learning rate reductions is not greater than 0, the model learning rate optimization is completed, otherwise the preset personnel are prompted to intervene.

[0051] In this embodiment, through the fusion of dynamic mapping and linear regression, the learning rate adjustment strategy can adapt to the changes in interference values ​​in different scenarios, avoid over-adjustment or under-adjustment problems caused by fixed step size adjustment, and improve convergence efficiency. On the premise of ensuring model convergence, by optimizing the learning rate reduction range and reducing unnecessary iterations, the model training time is significantly shortened to adapt to the needs of different hardware environments and task complexity.

[0052] like Figure 4As shown, it is a structural diagram of the child safety analysis system based on monitoring data provided by an embodiment of the present application. The child safety analysis system based on monitoring data provided by an embodiment of the present application includes: a data division module, a recognition validity analysis module and a detection image generation module; wherein the data division module is used to perform data division on the feature data set formed by the skeleton monitoring data to obtain a data division result, the skeleton monitoring data is used to visualize the child skeleton information in the target monitoring area, and the data division is used to generate a multidimensional data list based on the result of the skeleton monitoring data segmentation; the recognition validity analysis module is used to determine whether to perform dangerous posture recognition based on the data division result at the end of the preset division period, and if so, to determine whether to perform dangerous posture recognition based on the preset recognition The recognition effectiveness analysis is performed on the dangerous posture recognition results at the end of different time periods to obtain the recognition effectiveness analysis results. Otherwise, the windowing operation parameters are optimized. The recognition effectiveness analysis is used to quantify the effectiveness of the constructed pose posture detection model in recognizing dangerous postures on skeleton monitoring data. The windowing operation parameter optimization means adjusting the windowing operation parameters to improve the prediction performance of the pose posture detection model. The detection image generation module is used to determine whether to generate a behavior detection image based on the recognition effectiveness analysis results. If so, the generated behavior detection image is visualized. Otherwise, the model learning rate is optimized. The model learning rate optimization means adjusting the initial learning rate to improve the response rate of the pose posture detection model.

[0053] In this embodiment, through the collaborative work between the data partitioning module, the recognition validity analysis module and the detection image generation module, in posture detection, the YOLOv8 pose detection model trained with the Microsoft General Object Scene Dataset is used to generate the corresponding motion skeleton of the child; in behavior classification, the Spatial Temporal Graph Convolutional Networks (ST-GCN) network is used for real-time behavior classification; at the same time, multimodal recognition is introduced, mainly through video and audio for early childhood safety monitoring.

[0054] Among them, multimodal recognition: not only will the image data of the dataset be used as the training basis, but also image tags and voice tags will be used to achieve multimodal fusion and interaction. Through the collaboration of sound and video, the reliability of detection and adaptability to the scene are significantly improved, achieving more natural human-computer interaction and effectively dealing with the failure of a single modality.

[0055] At the same time, the "Smart Eyes-Campus Guardian" video analysis system was developed in HTML language to provide users with a simple and convenient interactive platform, and visualize the generated behavior detection images, which helps to overcome the inaccuracy of detection caused by the current detection model due to the limitations of children's body shape, and realize intelligent monitoring and early warning of the safety status of children in the target monitoring area, providing more effective protection for children's safety.

[0056] To sum up, the embodiment of the present application divides the feature data set formed by skeleton monitoring data, and then determines whether to perform dangerous posture recognition based on the data division result at the end of the preset division period. If so, the recognition effectiveness analysis is performed, otherwise the windowing operation parameters are optimized. Finally, based on the recognition effectiveness analysis result, it is determined whether to generate a behavior detection image. If so, the generated behavior detection image is visualized, otherwise the model learning rate is optimized, thereby realizing real-time and efficient processing of child safety monitoring data, and then realizing the improvement of the efficiency of child behavior recognition based on child safety monitoring data, effectively solving the problem of low posture recognition accuracy of pose posture detection model in the process of child behavior recognition based on child safety monitoring data in the prior art.

[0057] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0058] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A child safety analysis method based on monitoring data, characterized in that: The following steps are involved: Step one, performing data division on the feature data set formed by the skeleton monitoring data to obtain a data division result, wherein the skeleton monitoring data is used to visualize the skeleton information of the child in the target monitoring area, and the data division is used to generate a multidimensional data list based on the result of the segmentation of the skeleton monitoring data, wherein the feature data set represents the result of feature fusion and forms a data set with spatiotemporal characteristics after combining with time series information, and the result of the feature fusion represents the coordinates corresponding to the data feature points output by the preset posture prediction model, and the skeleton monitoring data represents the result of denoising the data points in the kindergarten monitoring image collected in the target monitoring area, and serves as the input of the preset posture prediction model, and the preset posture prediction model represents a prediction model for performing posture estimation and motion trajectory analysis based on the acquired child skeleton information after pre-training, wherein the child skeleton information includes the child's bone position, motion trajectory and posture features, and the multidimensional data list stores the dimensional information corresponding to the child's bone position coordinates, posture category and motion speed in a structured form; Step 2: Determine whether to perform dangerous posture recognition based on the data segmentation result at the end of the preset segmentation period. If so, perform recognition effectiveness analysis based on the dangerous posture recognition process data at the end of the preset recognition period to obtain a recognition effectiveness analysis result. Otherwise, perform windowing operation parameter optimization. The recognition effectiveness analysis is used to quantify the effectiveness of the constructed pose posture detection model in recognizing dangerous postures on the skeleton monitoring data. The windowing operation parameter optimization means adjusting the windowing operation parameters to improve the prediction performance of the pose posture detection model. Step 3: Based on the recognition validity analysis results, determine whether to generate a behavior detection image. If so, visualize the generated behavior detection image. Otherwise, perform model learning rate optimization. The model learning rate optimization means adjusting the initial learning rate to improve the response rate of the pose posture detection model. The specific process of determining whether to perform dangerous posture recognition based on the data division result at the end of the preset division period includes: When the obtained model prediction performance evaluation value is not greater than the model prediction performance evaluation value preset in the database, the data division result at the end of the preset division period is recorded as qualified data division and dangerous posture recognition is performed; When the obtained model prediction performance evaluation value is greater than the model prediction performance evaluation value preset in the database, the data partitioning result at the end of the preset partitioning period is recorded as unqualified data partitioning and the windowing operation parameters are optimized; The model prediction performance evaluation value represents the harmonic average result of the multidimensional data list generation time and the skeleton monitoring data segmentation time within a preset division period. The harmonic average is used to eliminate the interference of extreme values ​​in the multidimensional data list generation time and the skeleton monitoring data segmentation time on the preset posture prediction model prediction performance evaluation process. The windowing operation parameters include the sliding window length and the sliding window step size.

2. The method for analyzing child safety based on monitoring data according to claim 1, wherein: The dangerous posture recognition is performed by a pose posture detection model trained with training data in the classification data; The classification data includes training data and test data obtained from the data division results corresponding to the qualified data division; The pose detection model represents a classification model with behavior recognition function obtained by inputting the skeleton image of the child in the kindergarten monitoring image into the ST-DCN network.

3. The method for analyzing child safety based on monitoring data according to claim 2, wherein: The specific steps of optimizing the windowing operation parameters include: Inputting the obtained model prediction performance evaluation value deviation into the offline reinforcement learning algorithm of the preset posture prediction model to output the actual increase in the sliding window length, the model prediction performance evaluation value deviation is used to quantify the degree of difference between the obtained model prediction performance evaluation value and the preset model prediction performance evaluation value; After the sliding window length is increased once, the obtained first model prediction performance evaluation value deviation is re-inputted into the offline reinforcement learning algorithm of the preset posture prediction model to output the actual reduction amplitude of the sliding window step size; If the reduction in the deviation of the second model prediction performance evaluation value obtained after the first windowing operation parameter optimization is not greater than the reference reduction in the corresponding database, an alarm instruction is sent; otherwise, the windowing operation parameter optimization is continued; If the deviation of the second model prediction performance evaluation value obtained by the windowing operation parameter optimization within the preset optimization times is not greater than 0, the windowing operation parameter optimization is completed, otherwise the preset personnel are prompted to intervene.

4. The method for analyzing child safety based on monitoring data according to claim 1, wherein: The recognition validity analysis also includes determining whether to perform recognition validity analysis based on the total input and output time obtained. The specific process is as follows: When the total input and output time obtained is within the allowable range of the total input and output time in the database, the pose detection model is considered to have converged successfully and recognition validity analysis is performed; When the total input and output time obtained is not within the allowed range of the total input and output time in the database, it is determined that the pose detection model has failed to converge and the regularization term is adjusted; The total input and output time is used to quantify the actual time of the pose detection model within the preset recognition period. The length of time of the work, the dangerous posture recognition process data represents the recognition validity data obtained by the pose posture detection model at the end of the preset recognition period; The recognition validity data includes input response time, output response time and total input and output time. The input response time is used to quantify the time length of the pose posture detection model receiving training data and successfully inputting it. The output response time is used to quantify the time length of the pose posture detection model starting to output the posture estimation result to the end of successful output.

5. The method for analyzing child safety based on monitoring data according to claim 4, wherein: The specific process of adjusting the regularization term is as follows: When the acquired total input and output duration is greater than the maximum allowable total input and output duration in the database, it is determined that the pose posture detection model is overfitting and the regularization term of the pose posture detection model is increased based on the acquired regularization term increase amplitude, where the regularization term increase amplitude represents the result obtained by mapping the difference between the acquired total input and output duration and the maximum allowable total input and output duration in the database; When the obtained total input and output time is less than the minimum allowed total input and output time in the database, it is determined that the pose posture detection model is underfitting and the regularization term of the pose posture detection model is reduced based on the obtained regularization term reduction amplitude. The reduction amplitude of the regularization term represents the result obtained by mapping the difference between the minimum allowed total input and output time and the obtained total input and output time in the database.

6. The method for analyzing child safety based on monitoring data according to claim 1, wherein: The specific process of performing recognition effectiveness analysis based on the dangerous posture recognition process data at the end of the preset recognition period includes: When the obtained total input and output duration is within the allowable range of the total input and output duration in the database, the difference between the obtained input response duration and the reference input response duration in the database is compensated by the input response duration compensation value to obtain an input response duration score; Compensating the difference between the obtained output response time and the reference output response time in the database by using the output response time compensation value to obtain an output response time score; Compensating the difference between the obtained input-output response time and the reference input-output response time in the database using the input-output response time compensation value to obtain an input-output response time score; The obtained input response duration score, output response duration score, and input-output response duration score are coupled to obtain the recognition effectiveness interference value, which represents the quantitative data of the degree of influence of the input response duration, output response duration, and input-output response duration on the recognition effectiveness of the pose posture detection model.

7. The method for analyzing child safety based on monitoring data according to claim 6, wherein: The specific process of model learning rate optimization is as follows: The obtained recognition validity interference value deviation is mapped in the database to obtain the initial learning rate reduction amplitude of the pose posture detection model, wherein the recognition validity interference value deviation is used to quantify the degree of difference between the obtained recognition validity interference value and the recognition validity interference value preset in the database; The obtained initial learning rate reduction amplitude and the preset initial learning rate reduction amplitude are inputted into the linear regression algorithm of the pose posture detection model to output the actual initial learning rate reduction amplitude, wherein the actual initial learning rate reduction amplitude is used to ensure that the pose posture detection model improves the response rate of the pose posture detection model to the feature data set when the pose posture detection model converges satisfactorily; If the deviation of the recognition validity interference value reacquired after the initial learning rate is reduced for a preset number of times is not greater than 0, the model learning rate optimization is completed, otherwise the preset personnel are prompted to intervene.

8. A system using the method for analyzing child safety based on monitoring data according to any one of claims 1 to 7, characterized in that: include: Data partitioning module, recognition validity analysis module and detection image generation module; The data segmentation module is used to segment the feature data set formed by the skeleton monitoring data to obtain a data segmentation result. The skeleton monitoring data is used to visualize the skeleton information of the children in the target monitoring area. The data segmentation module is used to generate a multidimensional data list based on the segmentation result of the skeleton monitoring data. The recognition effectiveness analysis module is used to determine whether to perform dangerous posture recognition based on the data division result at the end of the preset division period. If so, a recognition effectiveness analysis is performed based on the dangerous posture recognition process data at the end of the preset recognition period to obtain a recognition effectiveness analysis result. Otherwise, a windowing operation parameter optimization is performed. The recognition effectiveness analysis is used to quantify the effectiveness of the constructed pose posture detection model in recognizing dangerous postures on the skeleton monitoring data. The windowing operation parameter optimization means improving the prediction performance of the pose posture detection model by adjusting the windowing operation parameters. The detection image generation module is used to determine whether to generate a behavior detection image based on the recognition validity analysis results. If so, the generated behavior detection image is visualized; otherwise, the model learning rate optimization is performed. The model learning rate optimization means adjusting the initial learning rate to improve the response rate of the pose posture detection model.

Images