Intelligent childcare video monitoring processing method and system

Abstract

The application provides an intelligent childcare video monitoring processing method and system. The method acquires first video monitoring data of a to-be-monitored childcare area at a first resolution. If it is determined that there is a target abnormal behavior in the first video monitoring data, the method switches to a second resolution with higher resolution to acquire second video monitoring data of the to-be-monitored childcare area. Thus, the video resolution is improved only when an abnormal behavior is detected, and unnecessary waste of data storage resources is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an intelligent childcare video monitoring processing method and system. BACKGROUND

[0002] With the increasing demand for childcare services in society, the safety and management of childcare institutions have attracted increasing attention. In order to ensure the safety and health of children in the childcare environment, real-time monitoring has become an indispensable management means. Traditional video monitoring systems usually use fixed resolution for continuous recording. Although this method can provide uninterrupted video recording, it also has obvious limitations.

[0003] On the one hand, high-resolution video data can provide clearer and more detailed pictures, but continuous recording consumes a large amount of storage space and requires a high demand for computing resources. This not only increases the operating costs of childcare institutions, but also may cause the system to be unable to run stably for a long time due to the limitation of storage space and computing capacity.

[0004] On the other hand, low-resolution video data can reduce the occupation of storage space and the burden of computing, but in the monitoring process, when abnormal behavior or emergency occurs, low-resolution video often cannot provide clear and accurate picture information, which brings difficulties to subsequent behavior analysis, event recording and evidence preservation.

[0005] Therefore, how to ensure the real-time monitoring and data storage economy while quickly obtaining high-quality video data when abnormal behavior occurs has become a technical problem to be solved in the field of childcare monitoring. SUMMARY

[0006] The present application provides an intelligent childcare video monitoring processing method and system, which avoids unnecessary waste of data storage resources by only increasing the video resolution when abnormal behavior is detected.

[0007] In a first aspect, the present application provides an intelligent childcare video monitoring processing method, comprising:

[0008] acquiring first video monitoring data of a to-be-monitored childcare area at a first resolution;

[0009] if it is determined that there is a target abnormal behavior in the first video monitoring data, switching to a second resolution to acquire second video monitoring data of the to-be-monitored childcare area, wherein the second resolution is higher than the first resolution.

[0010] In the above scheme, by intelligently regulating the resolution of video monitoring data, the dual goals of efficient monitoring and accurate identification of target abnormal behavior are achieved. First, the first video monitoring data of the to-be-monitored childcare area is obtained at the first resolution. This step fully considers the demand for storage space and computing resources of continuous video monitoring. The selection of the first resolution aims to balance the real-time monitoring and the economy of data storage. By using a relatively low resolution for preliminary monitoring, the amount of video data generated can be effectively reduced, thereby reducing the occupation of storage space and alleviating the computational burden of subsequent processing steps. When the model detects the existence of abnormal behavior, it immediately switches to the second resolution to obtain the second video monitoring data of the to-be-monitored childcare area. The second resolution is higher than the first resolution and can provide clearer and more detailed video pictures. The design of this switching mechanism ensures that when abnormal behavior occurs, the system can quickly obtain high-quality video data, providing strong support for subsequent behavior analysis, event recording, and evidence preservation. It can be seen that the above scheme, on the one hand, avoids unnecessary resource waste by only increasing the video resolution when abnormal behavior is detected, improving the overall efficiency of the monitoring system. On the other hand, the accurate capture of high-resolution video data provides a more abundant and accurate information source for subsequent abnormal behavior analysis, which helps to improve the accuracy and reliability of the analysis.

[0011] Optionally, after switching to the second resolution to obtain the second video monitoring data of the to-be-monitored childcare area, it further includes:

[0012] If it is determined that the duration of the absence of the target abnormal behavior in the second video monitoring data exceeds the preset duration threshold, the first resolution is switched to obtain the third video monitoring data of the to-be-monitored childcare area.

[0013] In the above scheme, in the case of long time without detecting target abnormal behavior, continuing to monitor video at high resolution (second resolution) will cause unnecessary waste of storage space and computing resources. By switching back to the first resolution to obtain the third video monitoring data, the amount of data generated can be significantly reduced, thereby reducing storage costs and alleviating the computational burden of the system, improving overall operating efficiency. Although the first resolution is switched back, the system still maintains the ability to continuously monitor the childcare area. This ensures that even in the case of lower resource consumption, the system can timely discover and respond to possible abnormal events.

[0014] Among them, the risk of abnormal behavior occurring in the childcare area is different at different time periods. For example, during the lunch break or quiet activity of infants and young children, the probability of abnormal behavior is lower. By dynamically adjusting the resolution, the system can flexibly adjust resource allocation according to the actual monitoring situation to adapt to the needs of different time periods.

[0015] Although the preset time threshold is used to determine when to switch back to the first resolution, the system still maintains the ability to monitor abnormal behavior in real time. Once abnormal behavior is detected again within the preset time threshold, the system can immediately switch back to the second resolution for high-precision monitoring, ensuring a quick response to sudden situations.

[0016] And due to factors such as environmental interference or sensor errors, the system may misjudge that there is abnormal behavior in a short period of time. By introducing a preset time threshold, the system can continuously observe the monitoring data for a period of time to reduce false positives caused by temporary interference.

[0017] Optionally, the first abnormal behavior monitoring model is used to monitor the target abnormal behavior from the first video monitoring data, and the second abnormal behavior monitoring model is used to monitor the target abnormal behavior from the second video monitoring data, wherein the identification accuracy of the first abnormal behavior monitoring model is higher than that of the second abnormal behavior monitoring model.

[0018] In the above scheme, in the first video monitoring stage, the first abnormal behavior monitoring model with higher identification accuracy is used for screening. This model can more accurately identify potential target abnormal behavior and reduce misjudgment and missed cases. This not only provides a more accurate target area for subsequent high-resolution monitoring, but also improves the accuracy of the entire monitoring system. It is worth noting that when switching to the second video monitoring stage, although the second abnormal behavior monitoring model with faster identification speed is used, since the second abnormal behavior monitoring model has a preset time threshold as a buffer time during the identification process, and the present scheme aims to record and save the monitoring video segment with the target abnormal behavior, and the second abnormal behavior monitoring model is used to determine the end time of high-resolution recording, and the identification accuracy of the second abnormal behavior monitoring model can be compensated through the buffer time, therefore, the identification accuracy of the second abnormal behavior monitoring model can be set to be lower than that of the first abnormal behavior monitoring model. Thus, both the monitoring efficiency and the monitoring accuracy of the abnormal behavior starting point and the monitoring reliability of the abnormal behavior ending point can be guaranteed.

[0019] Optionally, the first abnormal behavior monitoring model is an abnormal motion detection model determined based on a convolutional neural network model, and the second abnormal behavior monitoring model is an abnormal motion detection model determined based on a lightweight convolutional neural network model.

[0020] In the above scheme, in the intelligent childcare monitoring, the first abnormal behavior monitoring model adopts a CNN model, which can accurately identify the preliminary signs of the target abnormal behavior, and the second abnormal behavior monitoring model adopts a Lightweight CNN model, which has lower computational complexity and faster processing speed while ensuring a certain identification accuracy. After switching to the second video monitoring stage (higher resolution), the Lightweight CNN model can ensure a certain accuracy while reducing the consumption of computing resources. The process aims to save the video segments with the target abnormal behavior at high resolution. By using convolutional neural network models of different complexities, the system can flexibly adjust the allocation of computing resources according to actual needs, thereby balancing the amount of stored data and the amount of identification computation.

[0021] Optionally, the first abnormal behavior monitoring model is an abnormal action detection model determined based on multi-modal behavior data fusion, and the second abnormal behavior monitoring model is an abnormal action detection model determined based on single-modal behavior data.

[0022] In the above scheme, the first abnormal behavior monitoring model can more comprehensively capture the behavior information of the childcare object by fusing behavior data of multiple modalities. The complementarity of multi-modal data enables the model to more accurately identify complex and subtle abnormal actions, reducing information loss and misjudgment that may be caused by single-modal data. Although the second abnormal behavior monitoring model is based on single-modal behavior data, its main purpose is to determine the time endpoint of high-resolution video recording. Since there is a preset time threshold as the identification buffer period, the time endpoint can be determined through single-modal behavior data.

[0023] Optionally, the first abnormal behavior monitoring model is an abnormal action detection model determined based on limb action features and audio sound features, and the second abnormal behavior monitoring model is an abnormal action detection model based only on the limb action features.

[0024] In the above scheme, the first abnormal behavior monitoring model can more comprehensively capture the behavior information of the childcare object by fusing limb action features and audio sound features. The limb action features provide spatial dimension information of the behavior, while the audio sound features provide temporal dimension information of the behavior (such as the rhythm and intensity of the sound), and the combination of the two can more accurately identify complex and subtle abnormal actions. For example, when an infant falls or has a sudden illness, there will not only be abnormal limb movements, but also abnormal crying or moaning sounds. This multi-feature fusion method can significantly improve the accuracy of identification. In the monitoring stage, after the first abnormal behavior monitoring model identifies potential abnormal behavior through multi-feature fusion, the second abnormal behavior monitoring model only uses limb action features to quickly confirm the end time node.

[0025] Optionally, the preset duration threshold is determined by a preset behavior duration calibration list, which is used to establish a mapping relationship between the preset duration threshold and the target abnormal behavior.

[0026] In the above scheme, the preset behavior duration calibration list provides accurate duration threshold references for the system by recording and analyzing the time characteristics of various target abnormal behaviors in detail. This mechanism ensures that the system not only considers the type and characteristics of the behavior when determining abnormal behavior, but also fully considers the time factor of the behavior duration, thereby improving the accuracy of the determination. For example, for certain specific abnormal behaviors of infants (such as continuous convulsions, long crying, etc.), the system can accurately determine according to the preset duration threshold, avoiding misjudgment or omission. Different abnormal behaviors often have different duration characteristics. Through the preset behavior duration calibration list, the system can set different duration thresholds for different types of abnormal behaviors, thereby more accurately adapting to the duration differences of different behaviors. This flexibility enables the system to make more accurate and reasonable judgments when facing a variety of abnormal behaviors.

[0027] In a second aspect, the present application provides an intelligent childcare video monitoring processing system, comprising:

[0028] An acquisition module is configured to acquire first video monitoring data of a to-be-monitored childcare area at a first resolution.

[0029] A processing module is configured to switch to a second resolution to acquire second video monitoring data of the to-be-monitored childcare area when it is determined that there is a target abnormal behavior in the first video monitoring data, wherein the second resolution is higher than the first resolution.

[0030] Optionally, the processing module is further configured to switch to the first resolution to acquire third video monitoring data of the to-be-monitored childcare area when it is determined that the duration of the target abnormal behavior not existing in the second video monitoring data exceeds a preset duration threshold.

[0031] Optionally, a first abnormal behavior monitoring model is used to monitor the target abnormal behavior from the first video monitoring data, and a second abnormal behavior monitoring model is used to monitor the target abnormal behavior from the second video monitoring data, wherein the identification accuracy of the first abnormal behavior monitoring model is higher than that of the second abnormal behavior monitoring model.

[0032] Optionally, the first abnormal behavior monitoring model is an abnormal motion detection model determined based on a convolutional neural network model, and the second abnormal behavior monitoring model is an abnormal motion detection model determined based on a lightweight convolutional neural network model.

[0033] Optionally, the first abnormal behavior monitoring model is an abnormal action detection model determined based on multi-modal behavior data fusion, and the second abnormal behavior monitoring model is an abnormal action detection model determined based on single-modal behavior data.

[0034] Optionally, the first abnormal behavior monitoring model is an abnormal action detection model determined based on limb action features and audio sound features, and the second abnormal behavior monitoring model is an abnormal action detection model based only on the limb action features.

[0035] Optionally, the preset time threshold is determined by a preset behavior time calibration list, and the preset behavior time calibration list is used to establish a mapping relationship between the preset time threshold and the target abnormal behavior.

[0036] In a third aspect, the present application provides an electronic device, comprising:

[0037] a processor; and

[0038] a memory for storing executable instructions of the processor;

[0039] The processor is configured to execute any one of the possible methods of the first aspect by executing the executable instructions.

[0040] In a fourth aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores computer execution instructions, and the computer execution instructions are used to implement any one of the possible methods of the first aspect when executed by a processor.

[0041] The intelligent childcare video monitoring processing method and system provided by the present application can obtain first video monitoring data of a to-be-monitored childcare area at a first resolution, and if it is determined that there is a target abnormal behavior in the first video monitoring data, switch to a second resolution with higher clarity to obtain second video monitoring data of the to-be-monitored childcare area, so that the video resolution is only improved when an abnormal behavior is detected, thereby avoiding unnecessary waste of data storage resources. BRIEF DESCRIPTION OF DRAWINGS

[0042] The accompanying drawings, which are incorporated into and form part of the specification, illustrate embodiments consistent with the present application and, together with the specification, serve to explain the principles of the application.

[0043] Figure 1 FIG. 1 is a flowchart of an intelligent childcare video monitoring processing method according to an example embodiment of the present application;

[0044] Figure 2 FIG. 2 is a flowchart of an intelligent childcare video monitoring processing method according to another example embodiment of the present application;

[0045] Figure 3 is a structural schematic diagram of an intelligent childcare video monitoring processing system according to an example embodiment of the present application;

[0046] Figure 4 is a structural schematic diagram of an electronic device according to an example embodiment of the present application.

[0047] The specific embodiments of the present application have been shown and described in the above drawings, and will be described in more detail hereinafter. These drawings and the written description are not intended to restrict the scope of the present application in any way, but to illustrate the concept of the present application by reference to specific embodiments. DETAILED DESCRIPTION

[0048] The example embodiments will be described in detail herein with reference to the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following example embodiments are not meant to represent all implementations consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with some aspects of the present application as detailed in the appended claims.

[0049] Figure 1 is a structural schematic diagram of an intelligent childcare video monitoring processing method according to an example embodiment of the present application. As shown in Figure 1 the method provided by the present embodiment includes:

[0050] S101, acquiring first video monitoring data of a to-be-monitored childcare area at a first resolution.

[0051] In this step, the system first acquires the first video monitoring data of the to-be-monitored childcare area at a relatively low first resolution. The selection of the first resolution aims to balance the real-time performance of monitoring and the economy of data storage. Since normal monitoring data is the main data volume, by using a relatively low resolution for preliminary monitoring, the generation amount of video data can be effectively reduced, thereby reducing the occupation of storage space. It is worth mentioning that the collected monitoring data can be stored by using the encryption storage method in the prior art (such as the existing methods of password encryption, face encryption, and intranet encryption) to prevent data leakage and protect privacy.

[0052] Specifically, the system can configure a camera device in the childcare area to collect video data in real time at a preset first resolution (such as 720P). The collected video data will be transmitted to the video processing module of the system for processing and storage.

[0053] S102, if it is determined that the target abnormal behavior exists in the first video monitoring data, switching to a second resolution to obtain second video monitoring data of the to-be-monitored childcare area.

[0054] If it is determined that the target abnormal behavior exists in the first video monitoring data, switching to a second resolution to obtain second video monitoring data of the to-be-monitored childcare area, wherein the second resolution is higher than the first resolution.

[0055] In this step, the system performs real-time analysis on the first video monitoring data to detect whether there is a target abnormal behavior. The target abnormal behavior can be one or more of a plurality of pre-set abnormal behaviors, such as falling, crying, and violent movement of infants and young children.

[0056] To achieve the detection of abnormal behavior, the system can use an abnormal behavior monitoring model based on machine learning. Specifically, the system can first train the abnormal behavior monitoring model using training data to enable it to accurately identify the target abnormal behavior. During the training process, the system can extract key features (such as limb movement features, audio sound features, etc.) from the video data and input these features into the abnormal behavior monitoring model for learning and optimization.

[0057] During real-time monitoring, the system analyzes each frame of the first video monitoring data, extracts key features, and inputs these features into the trained abnormal behavior monitoring model. The model outputs a target behavior probability of the target abnormal behavior based on the input features. If the target behavior probability is greater than a pre-set abnormal behavior probability threshold, the system determines that the target abnormal behavior exists in the first video monitoring data.

[0058] Once the target abnormal behavior is detected, the system immediately switches to a second resolution to obtain second video monitoring data of the to-be-monitored childcare area. The second resolution is higher than the first resolution (such as 1080P or higher), which can provide clearer and more detailed video images. This switching mechanism ensures that when an abnormal behavior occurs, the system can quickly obtain high-quality video data, providing strong support for subsequent behavior analysis, event recording, and evidence preservation.

[0059] Specifically, the system can adjust the resolution setting of the camera device or use the video processing module to perform resolution enhancement processing on the collected video data to switch from the first resolution to the second resolution. The second video monitoring data after switching will be transmitted to the video storage module of the system for saving for subsequent analysis and use.

[0060] In this embodiment, by acquiring first video monitoring data of the to-be-monitored childcare area at a first resolution, if it is determined that there is a target abnormal behavior in the first video monitoring data, second video monitoring data of the to-be-monitored childcare area is acquired at a second resolution with higher definition, so that the video resolution is only improved when an abnormal behavior is detected, thereby avoiding unnecessary waste of data storage resources.

[0061] Figure 2 is a flowchart of an intelligent childcare video monitoring processing method according to another example embodiment of the present application. As shown in Figure 2 the intelligent childcare video monitoring processing method provided by the present embodiment includes:

[0062] S201, acquiring first video monitoring data of a to-be-monitored childcare area at a first resolution.

[0063] In this step, the system first acquires first video monitoring data of the to-be-monitored childcare area at a relatively low first resolution. The selection of the first resolution aims to balance the real-time performance of monitoring and the economy of data storage. Since normal monitoring data is the main data volume, by adopting a relatively low resolution for preliminary monitoring, the generation amount of video data can be effectively reduced, thereby reducing the occupation of storage space.

[0064] S202, if it is determined that there is a target abnormal behavior in the first video monitoring data, acquiring second video monitoring data of the to-be-monitored childcare area at a second resolution.

[0065] Optionally, the first abnormal behavior monitoring model is used to monitor the target abnormal behavior in the first video monitoring data, the second abnormal behavior monitoring model is used to monitor the target abnormal behavior in the second video monitoring data, and the identification accuracy of the first abnormal behavior monitoring model is higher than that of the second abnormal behavior monitoring model.

[0066] Specifically, in the first video monitoring stage, the system adopts the first abnormal behavior monitoring model with higher identification accuracy. The system analyzes the first video monitoring data frame by frame, extracts key features (such as limb action features, audio sound features, etc.), and inputs these features into the first abnormal behavior monitoring model for identification.

[0067] After switching to the second video monitoring stage, the system adopts a second abnormal behavior monitoring model with faster recognition speed but slightly lower recognition accuracy. This model has lower computational complexity and faster processing speed while ensuring a certain recognition accuracy. The system continues to analyze the second video monitoring data frame by frame, extracts key features, and inputs these features into the second abnormal behavior monitoring model for rapid identification. Although the recognition accuracy of this model is lower than that of the first model, since the determination has been made by the first model before switching to the second resolution, and there is a preset time threshold as a buffer in the second monitoring stage, the monitoring demand can still be met.

[0068] To train the first and second abnormal behavior monitoring models, video data of the childcare area can be collected and labeled with abnormal behaviors. The labeled data should include information such as the type, time of occurrence, and location of abnormal behaviors. The two models are trained separately using the labeled data. The first model focuses on improving recognition accuracy and can use more complex network structures and more training rounds; while the second model focuses on improving recognition speed and can use lightweight network structures and efficient training strategies.

[0069] During the training process, the system needs to optimize the parameters of the model to ensure that the model can perform optimally in different monitoring stages.

[0070] In the first video monitoring stage, the system uses the first abnormal behavior monitoring model to conduct preliminary monitoring and screening of video data. If potential target abnormal behavior is detected, the resolution switching mechanism is triggered.

[0071] After switching to the second video monitoring stage, the system uses the second abnormal behavior monitoring model to conduct rapid identification of high-resolution video data. If no target abnormal behavior is detected within a certain time (i.e., within the preset time threshold), it automatically switches back to the first resolution for subsequent monitoring.

[0072] In one possible implementation, the first abnormal behavior monitoring model is an abnormal action detection model based on a convolutional neural network model, and the second abnormal behavior monitoring model is an abnormal action detection model based on a lightweight convolutional neural network model.

[0073] This embodiment will provide specific implementation methods of the first abnormal behavior monitoring model and the second abnormal behavior monitoring model based on the previous ones, i.e., using an abnormal action detection model based on a convolutional neural network (CNN) and an abnormal action detection model based on a lightweight convolutional neural network (such as MobileNet, EfficientNet, etc.).

[0074] The first abnormal behavior monitoring model adopts a deep convolutional neural network structure, including multiple convolutional layers, pooling layers, fully connected layers, and activation function layers. The convolutional layers are used to extract spatial features in video frames, such as edges, textures, etc. The pooling layers are used to reduce the dimensionality of the feature maps and reduce the computational load. The fully connected layers are used to map the extracted features to the probability space of abnormal behavior. The activation function (such as ReLU) is used to increase the nonlinearity of the model and improve the expression ability of the model. Large-scale labeled childcare area video data can be used to train the model, and the data set contains multiple types of abnormal behavior. The cross-entropy loss function is used as the optimization objective, and the model parameters are updated through the backpropagation algorithm. Data augmentation techniques such as rotation, scaling, and cropping are used to increase the diversity of training data and improve the generalization ability of the model. The learning rate, batch size, and other hyperparameters are adjusted to optimize the training process of the model.

[0075] In the first video monitoring stage, the system processes the video data frame by frame, and inputs each frame into the first abnormal behavior monitoring model. The model outputs the probability of the presence of target abnormal behavior in each frame, and the system determines whether to trigger the resolution switching mechanism according to the preset probability threshold.

[0076] The second abnormal behavior monitoring model can be an abnormal action detection model based on a lightweight convolutional neural network (such as MobileNet). The second abnormal behavior monitoring model adopts a lightweight convolutional neural network structure, such as MobileNet, EfficientNet, etc. These network structures use techniques such as depthwise separable convolution and bottleneck structure to reduce the computational complexity and parameter quantity of the model while ensuring a certain recognition accuracy. The model also includes convolutional layers, pooling layers, fully connected layers, and activation function layers, but the specific structure and parameter settings are different from those of the first abnormal behavior monitoring model. The same training data and optimization objectives as the first abnormal behavior monitoring model are used to train the second model. Due to the characteristics of the lightweight network structure, the training process of the second model is more efficient and converges faster. Data augmentation techniques and hyperparameter adjustment strategies are also used to optimize the training process of the model.

[0077] In the second video monitoring stage, the system processes the high-resolution video data frame by frame, and inputs each frame into the second abnormal behavior monitoring model. The model quickly outputs the probability of the presence of target abnormal behavior in each frame, and the system determines whether to switch back to the first resolution or trigger the abnormal alarm mechanism according to the preset probability threshold and the preset time threshold.

[0078] In the first video monitoring phase, the system processes video data in real time and determines whether there are potential abnormal behaviors of the target based on the output of the first abnormal behavior monitoring model. If potential abnormal behavior is detected, the system immediately switches to the second video monitoring phase, using the second abnormal behavior monitoring model for high-resolution video monitoring. In the second video monitoring phase, the system continuously monitors video data and determines whether to switch back to the first resolution or trigger an abnormal alarm based on the output of the second abnormal behavior monitoring model and a preset duration threshold.

[0079] Furthermore, the first abnormal behavior monitoring model is an abnormal action detection model determined based on multimodal behavior data fusion, and the second abnormal behavior monitoring model is an abnormal action detection model determined based on single-modal behavior data.

[0080] In one possible implementation, to determine that there is abnormal target behavior in the first video monitoring data, one can obtain the skeletal key point sequence of the first video monitoring data, perform spatiotemporal graph convolution on the skeletal key point sequence to determine the skeletal feature vector, and simultaneously obtain the audio signal of the first video monitoring data to extract the acoustic feature vector from the audio signal. Then, generate a fused feature vector based on the skeletal feature vector and the acoustic feature vector, and input the fused feature vector into a time series model to output the probability of the target behavior. If the probability of the target behavior is greater than a preset abnormal behavior probability threshold, then it is determined that there is abnormal target behavior in the first video monitoring data.

[0081] Specifically, it could be the first video monitoring data acquired. Frame-by-frame skeletal keypoint sequence of childcare object ,in, For the first The three-dimensional coordinates of a preset joint. The total number of preset joints for the preset joint template.

[0082] Specifically, first video monitoring data of the childcare area to be monitored is acquired at a first resolution. Then, using computer vision technology, particularly a deep learning-based human pose estimation algorithm, a sequence of skeletal key points of the childcare object is extracted from each frame of the first video monitoring data. These key points typically correspond to major joints of the human body, such as the head, shoulders, elbows, wrists, hips, knees, and ankles. In this embodiment, we assume that the total number of preset joints in the preset joint template is n. Therefore, for each frame of video data, we can obtain a sequence of skeletal key points containing n three-dimensional coordinates, where each coordinate represents the position of a preset joint in three-dimensional space.

[0083] Using formula (1), and for the sequence of skeletal key points Perform spatiotemporal graph convolution to determine skeletal feature vectors. wherein formula (1) is:

[0084]

[0085] wherein, is a spatio-temporal graph convolution weight matrix, is the th spatio-temporal adjacency matrix, is a ReLU activation function, is a spatio-temporal graph convolution order;

[0086] Specifically, the extracted skeleton key point sequence can be processed by using spatio-temporal graph convolution to determine the skeleton feature vector. The spatio-temporal graph convolution is a convolution operation specially used for processing graph structure data, which can capture the dependence relationship in space and time at the same time. In this embodiment, formula (1) is used to calculate the skeleton feature vector, which contains parameters such as spatio-temporal graph convolution weight matrix, spatio-temporal adjacency matrix, ReLU activation function and spatio-temporal graph convolution order. Through this step, the system can convert the original skeleton key point sequence into a more expressive feature vector, providing a basis for subsequent behavior analysis.

[0087] For formula (1), the skeleton key point sequence is processed by spatio-temporal graph convolution, which realizes the effective extraction of the action features of the child care object's limbs. This process not only considers the spatial relationship between the skeleton points, but also integrates the information of the time dimension, making the extracted feature vector more comprehensive and accurate. The application of spatio-temporal graph convolution can capture subtle changes and trends in limb movements, thereby improving the recognition accuracy of abnormal behaviors. This is particularly important for monitoring the behavior of infants and young children, as their movements are often more complex and variable.

[0088] For the determination of the above-mentioned spatio-temporal graph convolution weight matrix, the following steps can be followed:

[0089] Initialize the weight matrix: before the start of training, the weight matrix is randomly initialized or uses some initialization strategy (such as Xavier initialization, He initialization, etc.).

[0090] Forward propagation: multiply the input feature matrix with the weight matrix, and get the output feature matrix through the activation function.

[0091] Calculate the loss: according to the task target (such as classification, regression, etc.), calculate the difference between the predicted value and the true value, and get the loss function value.

[0092] Backpropagation: calculate the gradient of the loss function with respect to the weight matrix using the chain rule, and update the weight matrix using an optimization algorithm (such as SGD, Adam, etc.).

[0093] Repeat training: Repeat the process of forward propagation, loss calculation, and backpropagation until the loss function converges or reaches a predetermined number of training rounds.

[0094] For the determination of the above-mentioned spatio-temporal adjacency matrix, the following steps can be followed:

[0095] Spatial adjacency matrix: For skeletal joints, an adjacency matrix can be constructed based on their connection relationships. If node i and node j are directly connected, the corresponding element in the adjacency matrix is 1, otherwise it is 0.

[0096] Temporal adjacency matrix: In spatio-temporal graph convolution, the temporal adjacency matrix is used to describe the connection relationship between nodes at different time steps. Adjacent time step nodes are considered to be directly connected, i.e. the temporal adjacency matrix is a matrix with elements 1 on the diagonal and 0 elsewhere (or plus a certain smoothing coefficient).

[0097] Spatio-temporal adjacency matrix: The spatio-temporal adjacency matrix combines the information of the spatial adjacency matrix and the temporal adjacency matrix, which can be directly added or multiplied to obtain the final spatio-temporal adjacency matrix.

[0098] For the determination of the above-mentioned spatio-temporal graph convolution order, the following steps can be followed:

[0099] Select the appropriate order according to the requirements of the specific task. For example, in the action recognition task, it may be necessary to capture the dependence relationship between distant skeletal joints, so a higher order can be selected. The higher the order, the higher the computational cost. Therefore, a trade-off between model performance and computational cost needs to be made.

[0100] The order can be selected by comparing the performance of the model under different orders through experiments. It can also be determined according to the structure of the graph and the characteristics of the task using a heuristic method. For example, in the skeletal joint graph, the appropriate order can be determined according to the physiological structure of the human body. A dynamic order method can also be used, i.e. different orders are used at different time steps or different nodes.

[0101] Next, the audio signal of the childcare object in the first video monitoring data is obtained frame, to extract the acoustic feature vector from the audio signal.

[0102] Specifically, audio signals of the childcare object are also extracted from each frame of the first video monitoring data. These audio signals contain various sounds emitted by the childcare object, such as speech, laughter, crying, and the like. In order to extract useful acoustic features from these audio signals, the system generates a mel-spectrogram using short-time Fourier transform and determines acoustic feature vectors. These acoustic feature vectors can reflect the changes of the audio signals in frequency and time, providing additional information for subsequent behavior analysis.

[0103] Using formula (2), and according to the skeletal feature vector and the acoustic feature vector determine the attention weight where formula (2) is:

[0104]

[0105] wherein, is the query matrix, is the key matrix, is the feature dimension.

[0106] Specifically, after obtaining the skeletal feature vector and the acoustic feature vector, the system uses formula (2) to calculate the attention weight. The attention mechanism can determine the importance of each feature according to the correlation between the skeletal feature vector and the acoustic feature vector.

[0107] For formula (2), the attention mechanism can be used to calculate the attention weight between the skeletal feature vector and the acoustic feature vector, achieving effective evaluation of feature importance. This step can highlight the role of key features in behavior recognition and reduce the interference of noise and irrelevant information. Through the introduction of attention weight, the system can focus more on features that are important for behavior recognition, thereby improving the robustness and stability of recognition.

[0108] wherein, for determination of the above query matrix and key matrix, the following steps can be taken:

[0109] Initialization: The query matrix and the key matrix are usually obtained by random initialization. Before the start of model training, initialization methods such as Xavier or He can be used to initialize these matrices according to the dimensions of the input features.

[0110] Feature extraction: The features of the query matrix and the key matrix are extracted from the skeletal feature vector and the acoustic feature vector. Therefore, a deep learning model (such as a convolutional neural network) is first used to process the skeletal data and the audio data respectively to extract high-level feature representations.

[0111] Dimension matching: The dimensions of the query matrix and the key matrix need to match in order to perform subsequent dot product operations. If the extracted feature dimensions do not match, the dimensions can be adjusted through linear transformations (such as fully connected layers).

[0112] Training Updates: During model training, the query matrix and key matrix are updated using the backpropagation algorithm. Based on the attention weights calculated by the attention mechanism and the model's loss function, the gradients of these matrices can be calculated, and optimization algorithms (such as Adam, SGD, etc.) can be used to update their values.

[0113] The determination of the aforementioned feature dimensions can be carried out according to the following steps:

[0114] In intelligent childcare monitoring, the goal is to accurately identify abnormal behaviors in children, requiring higher feature dimensionality to capture more detailed information. Higher feature dimensionality leads to greater model complexity and computational cost. Therefore, a trade-off must be struck between model performance and computational cost.

[0115] Experiments can be conducted to verify the model's performance under different feature dimensions. In practical applications, multiple candidate values ​​for feature dimensions can be set, models can be trained separately, and their performance on the validation set can be compared.

[0116] In addition to experimental verification, experience and research findings in related fields can be consulted to determine feature dimensions. For example, in research on attention mechanisms, there are some commonly used feature dimension settings that can be used as a reference.

[0117] Next, using formula (3) and based on the skeletal feature vector... and acoustic feature vectors Generate fused feature vectors Formula (3) is:

[0118]

[0119] For the above formula (3), the skeletal feature vector and the acoustic feature vector are fused to generate a fused feature vector containing richer information. This process achieves a comprehensive description of the childcare object's behavior, including both limb movements and vocal characteristics, providing a more accurate and comprehensive information basis for subsequent abnormal behavior judgment. The generation of the fused feature vector enables the system to comprehensively utilize information from multiple modalities for behavior recognition, thereby enhancing the recognition capability and accuracy.

[0120] fuse feature vectors Input to a time series model to output the probability of target behavior exhibiting anomalous behavior. The time series model includes formula (4), which is:

[0121]

[0122] wherein, is the weight matrix of the fully connected layer, is the bias term of the fully connected layer.

[0123] For the determination of the above weight matrix and bias term, the following steps can be referred to:

[0124] First, the weight matrix and bias term need to be initialized before the start of model training. Common initialization methods include random initialization, zero initialization, normal distribution initialization, Xavier initialization, etc. In the above scheme, considering the processing ability of the time series model for long sequence data and the need to avoid gradient vanishing or explosion, Xavier initialization or similar strategies can be used to adaptively adjust the scale of the initial weight according to the number of input and output neurons, which helps to keep the distribution of activation values and gradients in the entire network relatively stable, thereby improving the training efficiency and convergence of the model.

[0125] Once the weight matrix and bias term are initialized, they will be continuously adjusted and optimized during the model training process through the backpropagation algorithm. The training data includes a large amount of labeled video monitoring data, which contains various behavior scenarios of the childcare object, including normal behavior and abnormal behavior. These data are used to calculate the loss function, which measures the difference between the predicted target behavior probability of the model and the actual label.

[0126] In order to minimize the loss function, optimization algorithms such as stochastic gradient descent (SGD), Adam, RMSprop, etc. are usually used to update the weight matrix and bias term. These algorithms adjust the parameter values step by step according to the gradient information of the loss function with respect to the parameters, according to a certain learning rate, until the performance of the model on the validation set reaches the optimal or converges.

[0127] During the training process, in order to prevent the model from overfitting to the training data and reducing the generalization ability, regularization strategies need to be introduced. Common regularization methods include L1 regularization, L2 regularization, Dropout, etc. In this embodiment, L2 regularization term may be applied to the weight matrix, i.e. adding the sum of the squares of the weights multiplied by a regularization coefficient in the loss function, so as to encourage the model to learn smaller weight values, reduce the complexity of the model, and improve the generalization ability.

[0128] Finally, the final determination of the weight matrix and bias term also depends on the selection of a series of hyperparameters, such as learning rate, batch size, regularization coefficient, number of training rounds, etc. The selection of these hyperparameters has an important influence on the performance of the model, and usually needs to be optimized through strategies such as cross-validation, grid search or random search.

[0129] For the above formula (4), the fusion feature vector is processed by the time series model, and the target behavior probability of the target abnormal behavior is output. This process realizes the prediction and judgment of the behavior trend of the childcare object, and can timely discover and respond to abnormal behavior. The application of the time series model enables the system to quickly output the behavior probability, thereby improving the response speed and processing efficiency of abnormal behavior. This is particularly important for monitoring the infant care environment, as timely response and processing can maximize the safety and health of infants.

[0130] In the above scheme, by obtaining the skeleton key point sequence of the childcare object in each frame of the first video monitoring data, the limb movement information of the childcare object can be accurately captured. The extraction of the skeleton key point sequence fully considers the characteristics of infant behavior, such as large amplitude and rapid change, thereby ensuring the complete capture of key information.

[0131] Secondly, the skeleton key point sequence is processed using spatio-temporal graph convolution to determine the skeleton feature vector. Spatio-temporal graph convolution can consider information in both time and space dimensions, thereby more comprehensively reflecting the action features of the childcare object. The application of this step significantly improves the accuracy and robustness of behavior recognition, providing strong support for subsequent behavior judgment.

[0132] Furthermore, the acoustic feature vector is extracted from the audio signal and combined with the skeleton feature vector to determine the fusion feature vector using an attention mechanism. The introduction of the acoustic feature vector provides an additional information source for behavior recognition, such as the rhythm and intensity of sound, which complements the limb movement features and together forms a comprehensive behavior description. The application of the attention mechanism further highlights the role of key features in behavior recognition, improving the accuracy of recognition.

[0133] Finally, the fusion feature vector is input into the time series model to output the target behavior probability of the target abnormal behavior. The time series model can consider the time series characteristics of behavior, thereby predicting and judging the behavior trend. By comparing the target behavior probability with the preset abnormal behavior probability threshold, it can accurately determine whether there is a target abnormal behavior in the first video monitoring data.

[0134] If it is determined that the target behavior probability is greater than the preset abnormal behavior probability threshold , it is determined that there is a target abnormal behavior in the first video monitoring data.

[0135] Specifically, the system inputs the fused feature vector into a time-series model (e.g., a recurrent neural network, a long short-term memory network, or a Transformer) to output a target behavior probability of the target abnormal behavior. The time-series model can capture the temporal dependency of the feature vector, thereby more accurately identifying the abnormal behavior. In this embodiment, formula (4) (which includes the weight matrix and bias term of the fully connected layer) is used to calculate the target behavior probability. If the target behavior probability is greater than a preset abnormal behavior probability threshold, the system determines that the target abnormal behavior exists in the first video monitoring data, and triggers the subsequent high-resolution monitoring process.

[0136] Further, before determining the target behavior probability greater than the preset abnormal behavior probability threshold , formula (5) can also be used to determine the abnormal behavior probability threshold based on the sequence of skeletal key points of the child in the first frame , the sequence of skeletal key points of the child in the second frame , and the audio signal in the third frame , thereby achieving dynamic adjustment of the probability threshold, where formula (5) is:

[0137]

[0138] wherein, is a preset basic determination threshold, is a first dynamic threshold adjustment coefficient, is a second dynamic threshold adjustment coefficient, the coordinate of the i-th key point in the sequence of skeletal key points of the child in the first frame , the coordinate of the i-th key point in the sequence of skeletal key points of the child in the second frame , the coordinate of the i-th key point in the sequence of skeletal key points of the child in the third frame , , is the energy ratio of the target sound to the background noise in the audio signal.

[0139] ​​The preset basic determination threshold can be the distribution of the statistical normal behavior feature vector on the training set (such as mean ± 3 times standard deviation), and the threshold is set to cover more than 90% of normal samples to reduce false positives. The first dynamic threshold adjustment coefficient and the second dynamic threshold adjustment coefficient can be determined by analyzing the infant behavior data set to find the feature difference between normal behavior and abnormal behavior. In a laboratory environment, various behavior scenarios of infants can be simulated, and the threshold and adjustment coefficient can be adjusted to observe the changes in recognition effect, so as to find the optimal parameter combination. In addition, the threshold and adjustment coefficient can be preliminarily set in combination with the knowledge and experience of experts in the field, and fine-tuned through actual monitoring data.

[0140] In the above scheme, a dynamic threshold adjustment mechanism is introduced, which dynamically adjusts the preset basic determination threshold by considering the sequence change of the skeletal key points of the childcare object and the energy ratio in the audio signal. The technical effect of this step is that it can flexibly adjust the threshold of abnormal behavior determination according to the actual behavior characteristics of the childcare object and the environmental noise, thereby improving the accuracy and adaptability of the determination.

[0141] Specifically, by calculating the change amount of the key point coordinates in the skeletal key point sequence, the system can capture the subtle changes in the limb movements of the childcare object. This change amount reflects the activity level of the infant's behavior, and when the behavior becomes abnormally intense or suddenly changes, the change amount will significantly increase. At the same time, the energy ratio of the target sound to the background noise in the audio signal provides information in the sound dimension, and when the infant makes abnormal sounds (such as crying, screaming), the energy ratio will change.

[0142] Combining these two aspects of information, the system can more accurately determine whether the infant is in an abnormal state. When both behavior changes and sound characteristics indicate the possibility of an abnormality, the dynamic threshold is lowered, making the system more sensitive to abnormal behavior. Conversely, when the behavior is stable and the sound is normal, the dynamic threshold is relatively increased, reducing the possibility of false positives.

[0143] In addition, this dynamic threshold adjustment mechanism also enhances the robustness of the system. In actual application, environmental noise, light changes and other factors may interfere with the monitoring data, leading to false positives. By introducing dynamic threshold adjustment, the system can adaptively adjust the determination threshold according to real-time monitoring data, thereby reducing the impact of these interference factors on the determination result.

[0144] Further, this dynamic adjustment mechanism also helps to improve the overall performance of the system. In the process of abnormal behavior monitoring, the system needs to balance the decision accuracy and the consumption of computing resources. By dynamically adjusting the threshold, the system can reduce unnecessary computation and data storage as much as possible under the premise of ensuring decision accuracy, thereby improving the operation efficiency and resource utilization of the system.

[0145] In addition, the above extracting the acoustic feature vector from the audio signal may be including:

[0146] performing short-time Fourier transform on the audio signal to generate a mel-spectrogram ;

[0147] using formula (6) and determining the acoustic feature vector according to the mel-spectrogram , wherein formula (6) is:

[0148]

[0149] wherein, is a max-pooling operation, is a one-dimensional convolution operation.

[0150] In the above scheme, by performing short-time Fourier transform on the audio signal to generate a mel-spectrogram, this step converts the audio signal from time domain to frequency domain, and performs nonlinear compression on the frequency scale based on the hearing characteristics of human ears, so as to more accurately reflect the key features in the audio signal. The mel-spectrogram not only retains the time information of the audio signal, but also highlights the frequency components more relevant to human auditory perception.

[0151] Then, by using the combination of max-pooling operation and one-dimensional convolution operation (as shown in formula (5)), the mel-spectrogram is further processed to determine the acoustic feature vector. The max-pooling operation can reduce the dimension of the data, reduce the amount of calculation, and at the same time retain the most important feature information. The one-dimensional convolution operation can extract local features and realize effective traversal and feature extraction of the entire mel-spectrogram through the sliding of the convolution kernel. This combination operation not only improves the efficiency of feature extraction, but also ensures that the extracted acoustic feature vector can accurately reflect the key information in the audio signal.

[0152] It can be seen that the acoustic feature vector extracted from the audio signal can be one of the important inputs of the subsequent abnormal behavior monitoring model, combined with limb movement features such as skeletal key point sequences, to form multi-modal behavior data for more comprehensively capturing and identifying abnormal behaviors of the childcare object. The use of this multi-modal data fusion can significantly improve the accuracy and reliability of abnormal behavior monitoring.

[0153] ​S203, if it is determined that the duration of the absence of the target abnormal behavior in the second video monitoring data exceeds the preset duration threshold, switching to the first resolution to obtain third video monitoring data of the to-be-monitored childcare area.

[0154] In this step, after the system switches to the second resolution for high-resolution video monitoring, it continues to analyze the second video monitoring data to detect whether there is a target abnormal behavior. If the system does not detect a target abnormal behavior within a certain time (i.e., within the preset duration threshold), it will automatically switch back to the first resolution to continue obtaining third video monitoring data.

[0155] Figure 3 FIG. 1 is a structural schematic diagram of an intelligent childcare video monitoring processing system according to an example embodiment of the present application. As shown in FIG. 1, the intelligent childcare video monitoring processing system 300 provided in this embodiment includes: Figure 3

[0156] The acquisition module 310 is configured to acquire first video monitoring data of a to-be-monitored childcare area at a first resolution.

[0157] The processing module 320 is configured to, when it is determined that there is a target abnormal behavior in the first video monitoring data, switch to a second resolution to acquire second video monitoring data of the to-be-monitored childcare area, wherein the second resolution is higher than the first resolution.

[0158] Optionally, the processing module 320 is further configured to, if it is determined that the duration of the absence of the target abnormal behavior in the second video monitoring data exceeds a preset duration threshold, switch to the first resolution to acquire third video monitoring data of the to-be-monitored childcare area.

[0159] Optionally, the first abnormal behavior monitoring model is a first abnormal action detection model determined based on a convolutional neural network model, and the second abnormal behavior monitoring model is a second abnormal action detection model determined based on a lightweight convolutional neural network model.

[0160] Optionally, the first abnormal behavior monitoring model is an abnormal action detection model determined based on multi-modal behavior data fusion, and the second abnormal behavior monitoring model is an abnormal action detection model determined based on single-modal behavior data.

[0161] Optionally, the first abnormal behavior monitoring model is an abnormal action detection model determined based on multi-modal behavior data fusion, and the second abnormal behavior monitoring model is an abnormal action detection model determined based on single-modal behavior data.

[0162] ​Optionally, the first abnormal behavior monitoring model is an abnormal action detection model determined based on limb action features and audio sound features, and the second abnormal behavior monitoring model is an abnormal action detection model based on only the limb action features.

[0163] Optionally, the preset time threshold is determined by a preset behavior time threshold calibration list, and the preset behavior time threshold calibration list is used to establish a mapping relationship between the preset time threshold and the target abnormal behavior.

[0164] Figure 4 is a structural schematic diagram of an electronic device according to an example embodiment. As shown in Figure 4 The electronic device 400 provided in the embodiment includes a processor 401 and a memory 402.

[0165] The memory 402 is used to store a computer program, and the memory can also be a flash memory.

[0166] The processor 401 is used to execute the execution instructions stored in the memory to implement each step in the above method. For details, please refer to the related description in the method embodiment.

[0167] Optionally, the memory 402 can be independent or integrated with the processor 401.

[0168] When the memory 402 is independent of the processor 401, the electronic device 400 can further include:

[0169] The bus 403 is used to connect the memory 402 and the processor 401.

[0170] The embodiment further provides a readable storage medium, and the readable storage medium stores a computer program. When at least one processor of an electronic device executes the computer program, the electronic device executes the method provided in the various embodiments.

[0171] The embodiment further provides a program product, and the program product includes a computer program stored in a readable storage medium. At least one processor of an electronic device can read the computer program from the readable storage medium, and the at least one processor executes the computer program to make the electronic device implement the method provided in the various embodiments.

[0172] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the specification and examples be considered as exemplary only, with the true scope and spirit of the application being indicated by the following claims.

[0173] It is to be understood that the application is not limited to the precise construction herein disclosed and shown in the drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the application. The scope of the application is limited only by the claims that follow.

Claims

1. A method for intelligent childcare video monitoring and processing, characterized in that, include: The first video monitoring data of the childcare area to be monitored is acquired at a first resolution, and the first video monitoring data is transmitted to the video storage module for storage; If it is determined that there is abnormal behavior of the target in the first video monitoring data, the system switches to the second resolution to obtain the second video monitoring data of the childcare area to be monitored, and transmits the second video monitoring data to the video storage module for storage, wherein the second resolution is higher than the first resolution; If it is determined that the duration of the abnormal behavior of the target in the second video monitoring data does not exceed a preset duration threshold, then switch to the first resolution to obtain the third video monitoring data of the childcare area to be monitored, and transmit the third video monitoring data to the video storage module for storage; The first abnormal behavior monitoring model is used to monitor the abnormal behavior of the target from the first video monitoring data, and the second abnormal behavior monitoring model is used to monitor the abnormal behavior of the target from the second video monitoring data, wherein the recognition accuracy of the first abnormal behavior monitoring model is higher than that of the second abnormal behavior monitoring model. The first abnormal behavior monitoring model is an abnormal behavior detection model determined based on limb movement features and audio sound features, while the second abnormal behavior monitoring model is an abnormal behavior detection model based solely on the limb movement features.

2. The intelligent childcare video monitoring and processing method according to claim 1, characterized in that, The first abnormal behavior monitoring model is an abnormal action detection model determined based on multimodal behavior data fusion, and the second abnormal behavior monitoring model is an abnormal action detection model determined based on single-modal behavior data.

3. The intelligent childcare video monitoring and processing method according to claim 1 or 2, characterized in that, The preset duration threshold is determined by a preset behavior duration calibration list, which is used to establish a mapping relationship between the preset duration threshold and the target abnormal behavior.

4. An intelligent childcare video monitoring and processing system, characterized in that, include: The acquisition module is used to acquire first video monitoring data of the childcare area to be monitored at a first resolution, and transmit the first video monitoring data to the video storage module for storage. The processing module is used to switch to the second resolution to obtain the second video monitoring data of the childcare area to be monitored when it is determined that there is abnormal behavior of the target in the first video monitoring data, and to transmit the second video monitoring data to the video storage module for storage, wherein the second resolution is higher than the first resolution; The processing module is further configured to, when determining that the duration for which no abnormal behavior of the target exists in the second video monitoring data exceeds a preset duration threshold, switch to the first resolution to acquire the third video monitoring data of the childcare area to be monitored, and transmit the third video monitoring data to the video storage module for storage; The first abnormal behavior monitoring model is used to monitor the abnormal behavior of the target from the first video monitoring data, and the second abnormal behavior monitoring model is used to monitor the abnormal behavior of the target from the second video monitoring data, wherein the recognition accuracy of the first abnormal behavior monitoring model is higher than that of the second abnormal behavior monitoring model. The first abnormal behavior monitoring model is an abnormal behavior detection model determined based on limb movement features and audio sound features, while the second abnormal behavior monitoring model is an abnormal behavior detection model based solely on the limb movement features.

5. An electronic device, characterized in that, include: processor; as well as, Memory for storing the executable instructions of the processor; The processor is configured to execute the method of any one of claims 1 to 3 by executing the executable instructions.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 3.

Images