Special child dangerous behavior early warning system based on artificial intelligence
By introducing a joint analysis of child behavior recognition and scene understanding, along with a multi-level feature fusion strategy, into the special children's dangerous behavior early warning system, and combining it with an improved swarm intelligence optimization algorithm, the problems of insufficient action behavior recognition and unreasonable hyperparameter settings in traditional systems are solved, thus achieving accurate and intelligent recognition of children's dangerous behaviors.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-10
AI Technical Summary
Traditional early warning systems for dangerous behaviors of children with special needs rely on action and behavior recognition, but lack comprehensive analysis of the child's individual physical state, spatial location, and environmental context. This results in one-sided judgments of dangerous behaviors. Existing models have unreasonable hyperparameter settings, making it difficult to balance global exploration and local development, resulting in poor model generalization performance and insufficient accuracy of output results.
This paper introduces joint analysis of children's behavior recognition and behavior scene understanding, performs cross-modal temporal fusion through bidirectional long short-term memory network, combines multi-level residual dense blocks and bi-branch attention mechanism, and adopts an improved swarm intelligence optimization algorithm for hyperparameter optimization to achieve joint judgment of children's behavior patterns and environmental risk factors and dynamic balance optimization of model parameters.
It improves the comprehensiveness, accuracy, and robustness of dangerous behavior identification, reduces the risk of misjudgment, achieves precise identification and intelligent early warning of dangerous behaviors of children with special needs, and improves the model's recognition accuracy and generalization ability.
Smart Images
Figure CN121838264A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of behavioral data processing, specifically to an artificial intelligence-based early warning system for dangerous behaviors of children. Background Technology
[0002] The AI-based early warning system for dangerous behaviors of children with special needs is an intelligent safety management system that uses artificial intelligence technology to monitor, analyze, identify behavioral patterns, and understand scene semantics of children with special needs in real time in specific environments. When potential dangerous behaviors are detected, it automatically generates early warning signals. This system can achieve immediate protection and management of children with special needs, reduce accident risks, and improve the efficiency of supervision and management by teachers, parents, or administrators.
[0003] However, traditional early warning systems for dangerous behaviors of children with special needs suffer from technical problems. They rely solely on action and behavior recognition, lacking a comprehensive analysis of the child's individual physical state, spatial location, and environmental context. This leads to one-sided results in dangerous behavior judgment. Existing methods for action and behavior recognition rely only on shallow convolutional features, lacking joint analysis of the evolution of actions over time and key spatial parts. This results in insufficient extraction of behavioral features, insufficient temporal dependence, and limited spatial and temporal discrimination capabilities, leading to low accuracy in dangerous behavior recognition. Furthermore, existing hyperparameter settings for dangerous behavior early warning models are unreasonable, and it is difficult to balance global exploration and local development during hyperparameter optimization, easily getting trapped in local optima. This results in poor model generalization performance and insufficient accuracy in model output results. Summary of the Invention
[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an artificial intelligence-based early warning system for dangerous behaviors in children with special needs. Addressing the technical problem of traditional systems relying solely on action recognition and lacking comprehensive analysis of the child's individual physical state, spatial location, and environmental context, leading to biased dangerous behavior assessments, this solution innovatively combines child behavior recognition with behavioral scene understanding and introduces an individualized analysis mechanism. By integrating optimized data on the child's individual body with action sequence features extracted from the child's behavioral image data, and performing cross-modal temporal fusion within a bidirectional long short-term memory network, this system achieves... The joint assessment of personalized children's behavioral patterns and environmental risk factors effectively improves the comprehensiveness, accuracy, and robustness of dangerous behavior identification, significantly reduces the risk of misjudgment caused by single action recognition, and achieves precise identification of dangerous behaviors of special children. It possesses real-time, accurate, and intelligent safety early warning effects for dangerous behaviors of special children. However, existing methods for action behavior recognition rely solely on shallow convolutional features and lack joint analysis of the temporal evolution of actions and key spatial components. This results in insufficient extraction of behavioral features, inadequate temporal dependence, and limited spatial and temporal discrimination capabilities, leading to low accuracy in dangerous behavior identification results. To address the technical challenges, this solution innovatively proposes a children's behavior temporal attention recognition network that combines multi-level residual dense blocks, a bi-branch attention mechanism, and a feature-guided fusion strategy. This enhances the sufficiency of feature extraction, improves spatial discrimination and temporal dependency modeling capabilities, and enhances the robustness and hierarchy of feature representation. It effectively improves the accuracy and stability of action recognition, thereby further enhancing the accuracy and comprehensiveness of dangerous behavior recognition, reducing the risk of false positives and false negatives, and achieving high-precision and intelligent recognition of children's dangerous behaviors. Furthermore, it addresses the issue of unreasonable hyperparameter settings in existing dangerous behavior early warning models. In the process of hyperparameter optimization, it is difficult to balance global exploration and local development, which easily leads to local optima, resulting in poor generalization performance and insufficient accuracy of model output. This solution innovatively introduces an improved swarm intelligence optimization algorithm with a multi-branch exploration and update strategy to dynamically balance global search and local development of model hyperparameters. This effectively avoids the local optimum trap in the hyperparameter optimization process, improves the search space coverage and optimization convergence speed, enhances the rationality of model hyperparameter selection, improves the recognition accuracy and generalization ability of the dangerous behavior early warning model, and realizes intelligent and reliable early warning of dangerous behaviors of children with special needs.
[0005] The technical solution adopted by the present invention is as follows: The special children's dangerous behavior early warning system based on artificial intelligence provided by the present invention includes a data acquisition module, a data optimization module, a dangerous behavior early warning model establishment module, an early warning model parameter optimization module, and a dangerous behavior intelligent early warning module;
[0006] The data acquisition module specifically obtains raw data for dangerous behavior warnings by collecting data;
[0007] The data optimization module specifically obtains optimized data for dangerous behavior warnings through image data optimization, non-image data optimization, and data synchronization.
[0008] The module for establishing a dangerous behavior early warning model specifically involves generating action sequence features through a child behavior temporal attention recognition network, and then performing cross-modal temporal fusion with the child's individual body optimization data in a bidirectional long short-term memory network to obtain child behavior recognition features. Subsequently, combined with the child behavior scene understanding features obtained from behavior scene understanding, judgment features are generated through an attention mechanism. Finally, dangerous behavior is judged based on the judgment features to obtain the child's dangerous behavior recognition result.
[0009] The early warning model parameter optimization module first completes the initial training of the early warning model under a fixed hyperparameter configuration, and then uses an improved swarm intelligence optimization algorithm to dynamically search and optimize the model hyperparameters to obtain the optimal hyperparameter combination. Based on the optimal hyperparameter combination, the model is retrained to obtain the high-performance dangerous behavior early warning model.
[0010] The intelligent early warning module for dangerous behavior specifically inputs the target child's dangerous warning data into the optimal dangerous behavior early warning model, outputs the child's real-time dangerous behavior identification results, and automatically generates corresponding early warning signals based on these results to notify the guardians.
[0011] Furthermore, the data acquisition module specifically collects data through video surveillance equipment and sensor devices worn by children to obtain raw data for dangerous behavior warnings; the raw data for dangerous behavior warnings includes reference child dangerous warning data and target child dangerous warning data; both the reference child dangerous warning data and the target child dangerous warning data include child behavior image data, child individual body data, child environmental image data, and child location data; the reference child dangerous warning data also includes reference child dangerous behavior identification results.
[0012] Furthermore, the data optimization module specifically includes the following steps:
[0013] Image data optimization specifically involves optimizing children's behavioral image data and children's environmental image data, including image cleaning, image format standardization, and image quality enhancement, to obtain optimized image data.
[0014] Non-image data optimization specifically involves optimizing individual child body data and child location data, including data cleaning, data standardization, and data encoding, to obtain non-image optimized data.
[0015] Data synchronization specifically involves using a dynamic time warping method to perform time-series matching and alignment between image-optimized data and non-image-optimized data, thereby obtaining time-aligned optimized data for dangerous behavior warnings.
[0016] Furthermore, the module for establishing a dangerous behavior early warning model specifically includes the following steps:
[0017] The process of child behavior recognition involves first inputting optimized data of children's behavior images into a children's behavior temporal attention recognition network to generate action sequence features. Then, the action sequence features and optimized data of the child's individual body are input into a bidirectional long short-term memory network for cross-modal temporal feature fusion to obtain children's behavior recognition features.
[0018] The child behavior temporal attention recognition network specifically includes the following steps:
[0019] Multi-level behavioral feature extraction specifically involves inputting optimized data of children's behavioral images into five concatenated residual dense blocks. The first residual dense block receives the optimized data of children's behavioral images, and the remaining residual dense blocks sequentially use the output of the previous residual dense block as input to complete the extraction of different levels of action features step by step, resulting in a multi-level behavioral feature set composed of the output features of the five residual dense blocks.
[0020] The dual-branch attention enhancement method involves mapping the output features of each residual dense block to query vectors, key vectors, and value vectors of spatial and temporal branches through convolution operations. Based on this, spatial attention weights are calculated to extract spatial enhancement features, and temporal attention weights are calculated to extract temporal enhancement features. Subsequently, the spatial enhancement features and temporal enhancement features are weighted and fused according to adaptive weights, and then residuals are added to the corresponding residual dense block output features to obtain the spatiotemporal features of dangerous behavior.
[0021] Multi-level behavioral feature fusion specifically involves first multiplying the output feature of the highest-level residual dense block element-wise with the corresponding spatiotemporal feature of dangerous behavior to obtain the fused feature for that level. For the remaining levels, a feature-guided fusion strategy is used to generate the corresponding fused features. Subsequently, global average pooling is performed on the fused features of each level, and they are concatenated along the channel dimension to obtain the action sequence features. The formula used is as follows:
[0022] ;
[0023] In the formula, This represents the fusion feature at the i-th level. This indicates a channel splicing operation. Indicates the first Spatiotemporal characteristics of dangerous behaviors This indicates an element-wise multiplication operation. This represents the spatiotemporal characteristics of the i-th dangerous behavior. This indicates the spatiotemporal characteristics of the fifth dangerous behavior. This represents a 1×1 convolution operation;
[0024] The behavioral scene understanding specifically involves processing optimized data of children's environmental images through a deep convolutional neural network to extract scene features, and simultaneously extracting position features from optimized data of children's positions through a multilayer perceptron neural network. Then, the scene features and position features are concatenated into vectors and jointly mapped through a fully connected layer and a normalization layer to obtain children's behavioral scene understanding features.
[0025] The dangerous behavior determination process involves weighting and combining the child's behavior recognition features and the child's behavior scene understanding features through an attention mechanism to generate determination features. These determination features are then input into a fully connected layer and a Softmax classifier to determine the risk level of the dangerous behavior and obtain the child's dangerous behavior recognition result.
[0026] Furthermore, the early warning model parameter optimization module specifically includes the following steps:
[0027] The initial training of the early warning model involves using the optimized reference child danger warning data as training data, iteratively updating the model parameters through forward and backward propagation, and training the early warning model to obtain an initial trained danger behavior early warning model.
[0028] Intelligent hyperparameter optimization specifically involves obtaining the optimal combination of hyperparameters for the model through improved optimization algorithms; it includes the following steps:
[0029] Initialize the search individuals by encoding the hyperparameters of the dangerous behavior warning model into search individual position vectors, and generate N search individual position vectors through a random initialization method. Each individual is encoded to represent a candidate combination of dangerous behavior warning model hyperparameters, thus obtaining the initial search population.
[0030] The fitness value of a search individual is calculated by calculating the fitness value of the search individual in the population; the performance of the danger behavior early warning model established based on the location of the search individual is used as the fitness value of the search individual.
[0031] The search process involves dividing individuals into groups, specifically selecting the individual with the best fitness as the leader and the rest as followers.
[0032] The leader's individual position is updated using a multi-branch exploration strategy; the formula used is as follows:
[0033] ;
[0034] In the formula, Indicates the first In the next iteration, the leader's position in the j-th dimension. Let T represent the current optimal individual position in the j-th dimension, and let T represent the current iteration number. Indicates the maximum number of iterations. and They represent Random numbers within a range and Let represent the upper and lower bounds of the search in the j-th dimension for the optimal individual, respectively. and These represent the upper and lower bounds of the global search space, respectively. Indicates the probability threshold. This represents the position of the worst individual in the j-th dimension. express Random numbers within a certain range;
[0035] The follower's position is updated by setting the position of each follower to the average of its own position and the position of the previous searched individual; the formula used is as follows:
[0036] ;
[0037] In the formula, Indicates the first In the next iteration, the position of the i-th follower individual in the j-th dimension. Indicates the first In the next iteration, the position of the i-th follower individual in the j-th dimension. Indicates the first During the nth iteration, the 1st The position of a follower individual in the j-th dimension, where i represents the index of the follower individual;
[0038] The optimal position update for a search individual involves evaluating the fitness value of the search individual after the update in the current iteration, and comparing the fitness value of the current search individual with the global optimal position of the current search individual. If the fitness value of the current search individual is better, the global optimal position of the search individual is updated.
[0039] The iterative search terminates when the fitness value of the search individual is higher than the fitness threshold or when the maximum number of iterations is reached, and the global optimal position of the search individual is obtained; the global optimal position of the search individual specifically refers to the optimal combination of hyperparameters of the model.
[0040] The final optimization of the early warning model involves adjusting the hyperparameters of the initially trained dangerous behavior early warning model based on the optimal hyperparameter combination, and using the reference children's dangerous warning data after data optimization as training data to retrain the model and update the parameters, thus obtaining the best-performing dangerous behavior early warning model.
[0041] Furthermore, the intelligent early warning module for dangerous behavior specifically inputs the target child's dangerous warning data after data optimization into the best-performing dangerous behavior early warning model to obtain the child's real-time dangerous behavior identification result. If the child's real-time dangerous behavior identification result is not safe, then the module automatically generates a corresponding early warning signal based on the child's real-time dangerous behavior identification result and immediately notifies teachers, parents, and administrators.
[0042] The beneficial effects achieved by the present invention using the above solution are as follows:
[0043] (1) In view of the technical problem that traditional special children’s dangerous behavior early warning system relies solely on action behavior recognition and lacks comprehensive analysis of the child’s individual physical state, spatial location and environmental context, resulting in one-sided dangerous behavior judgment results, this solution innovatively combines child behavior recognition with behavior scene understanding and introduces an individualized analysis mechanism. By combining the child’s individual physical optimization data with the action sequence features extracted from the child’s behavior image data, cross-modal temporal fusion is carried out in a bidirectional long short-term memory network, which realizes the joint judgment of personalized child behavior patterns and environmental risk factors. This effectively improves the comprehensiveness, accuracy and robustness of dangerous behavior recognition, significantly reduces the risk of misjudgment caused by single action recognition, realizes the accurate recognition of dangerous behavior of special children, and has the real-time, accurate and intelligent special children’s dangerous behavior safety early warning effect.
[0044] (2) In view of the technical problem that existing methods for action and behavior recognition rely only on shallow convolutional features and lack joint analysis of the evolution of actions over time and key spatial parts, resulting in insufficient extraction of behavioral features, insufficient temporal dependence, and limited spatial and temporal discrimination capabilities, thus causing low accuracy of dangerous behavior recognition results, this solution innovatively proposes a children's behavior temporal attention recognition network that combines multi-level residual dense blocks, a dual-branch attention mechanism, and a feature-guided fusion strategy. This enhances the sufficiency of feature extraction, improves spatial discrimination and temporal dependence modeling capabilities, and enhances the robustness and hierarchy of feature expression. It effectively improves the accuracy and stability of action recognition, thereby further improving the accuracy and comprehensiveness of dangerous behavior recognition, reducing the risk of false alarms and false misses, and realizing high-precision and intelligent recognition of children's dangerous behaviors.
[0045] (3) In view of the technical problems that existing hyperparameter settings for dangerous behavior early warning models are unreasonable and that it is difficult to balance global exploration and local development during the hyperparameter optimization process, which easily leads to local optima, resulting in poor model generalization performance and insufficient accuracy of model output results, this solution innovatively introduces an improved swarm intelligence optimization algorithm with a multi-branch exploration and update strategy to perform dynamic balance optimization of model hyperparameters between global search and local development. This can effectively avoid the local optimum trap in the hyperparameter optimization process, improve the search space coverage and optimization convergence speed, enhance the rationality of model hyperparameter selection, improve the recognition accuracy and generalization ability of dangerous behavior early warning models, and realize intelligent and reliable early warning of dangerous behaviors of special children. Attached Figure Description
[0046] Figure 1 A schematic diagram of the modules of the artificial intelligence-based early warning system for dangerous behaviors of children provided by the present invention;
[0047] Figure 2 A flowchart illustrating the process of establishing a dangerous behavior early warning model module;
[0048] Figure 3 A flowchart illustrating the parameter optimization module for the early warning model;
[0049] Figure 4 This is a flowchart illustrating the intelligent parameter optimization process in the parameter optimization module of the early warning model.
[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation
[0051] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0052] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0053] Example 1, see Figure 1The present invention provides an artificial intelligence-based early warning system for dangerous behaviors of children, including a data acquisition module, a data optimization module, a dangerous behavior early warning model establishment module, an early warning model parameter optimization module, and a dangerous behavior intelligent early warning module;
[0054] The data acquisition module specifically collects data to obtain raw data for dangerous behavior warnings and sends the data to the data optimization module.
[0055] The data optimization module receives data sent by the data acquisition module and performs quality improvement and structured processing on the original data. Specifically, it obtains dangerous behavior warning optimization data through image data optimization, non-image data optimization and data synchronization, and sends the data to the warning model parameter optimization module and the dangerous behavior intelligent warning module.
[0056] The module for establishing a dangerous behavior early warning model is used to construct a dangerous behavior early warning model that can simultaneously perceive children's action patterns and environmental context. Specifically, it generates action sequence features through a children's behavior temporal attention recognition network, and performs cross-modal temporal fusion with children's individual body optimization data in a bidirectional long short-term memory network to obtain children's behavior recognition features. Then, it combines children's behavior scene understanding features obtained from behavior scene understanding, and generates judgment features through an attention mechanism. Finally, it completes the dangerous behavior judgment based on the judgment features, obtains the children's dangerous behavior recognition results, and sends the data to the early warning model parameter optimization module.
[0057] The early warning model parameter optimization module receives data from the data optimization module and the dangerous behavior early warning model establishment module, and is used to optimize the parameters of the dangerous behavior early warning model. Specifically, the early warning model is first trained under a fixed hyperparameter configuration. Then, the improved swarm intelligence optimization algorithm is used to dynamically search and optimize the model hyperparameters to obtain the optimal hyperparameter combination. Based on the optimal hyperparameter combination, the model is retrained to obtain the best-performing dangerous behavior early warning model, and the data is sent to the dangerous behavior intelligent early warning module.
[0058] The intelligent early warning module for dangerous behaviors receives data from the data optimization module and the early warning model parameter optimization module, and is used to provide real-time early warnings for dangerous behaviors of the target special children. Specifically, it inputs the target child's dangerous warning data into the optimal dangerous behavior early warning model, outputs the child's real-time dangerous behavior identification results, and automatically generates corresponding early warning signals based on these results to notify the guardians, thereby realizing intelligent safety early warning and protection for special children.
[0059] Example 2, see Figure 1This embodiment is based on the above embodiment. Specifically, the data acquisition module collects data through video surveillance equipment and sensor devices worn by children to obtain raw data for dangerous behavior warnings. The raw data for dangerous behavior warnings includes reference child danger warning data and target child danger warning data. Both the reference child danger warning data and the target child danger warning data include child behavior image data, child individual body data, child environmental image data, and child location data.
[0060] The reference child danger warning data also includes reference child dangerous behavior identification results;
[0061] The reference child risk behavior identification results include safe behavior, low-risk behavior, medium-risk behavior, and high-risk behavior.
[0062] The child behavior image data comes from the image acquisition of video surveillance equipment and belongs to the action layer data. By detecting and segmenting the child target in the video frame, it is used to perform child behavior pattern recognition.
[0063] The individual physical data of the children are used to reflect the children's exercise intensity and physical condition. They belong to the body layer data and include acceleration, angular velocity, cadence and range of motion, heart rate, body temperature and respiratory rate.
[0064] The children's behavioral image data comes from video surveillance equipment and belongs to scene layer data. By performing scene semantic segmentation on video frames, environmental objects are extracted to achieve understanding of the scene in which the child is located and analysis of dangerous contexts.
[0065] The child's location data includes real-time spatial coordinates, movement trajectory, and area label information, and belongs to the spatial layer data.
[0066] Example 3, see Figure 1 This embodiment is based on the above embodiment, and the data optimization module specifically includes the following steps:
[0067] Image data optimization specifically involves optimizing children's behavioral image data and children's environmental image data, including image cleaning, image format standardization, and image quality enhancement, to obtain optimized image data.
[0068] The image optimization data shown includes optimized image data of children's behavior and optimized image data of children's environment;
[0069] The image cleaning is used to remove interference information and abnormal frames in the image to ensure the accuracy and usability of the image data. Specifically, it uses a mean filtering algorithm to filter noise in the frame image, removes blurry, overly bright or dark abnormal frame images, and uses an inter-frame difference detection algorithm to delete invalid background frames.
[0070] The image standardization process is used to unify image data from different sources and resolutions into a standard input format to ensure the consistency of data processing. Specifically, it involves scaling the image to a preset resolution using a bilinear interpolation algorithm and unifying the color format using a histogram equalization algorithm.
[0071] The image enhancement is used to improve image clarity and feature distinguishability. Specifically, it improves image quality through a linear contrast stretching algorithm and expands image data samples through random geometric transformation and random brightness.
[0072] Non-image data optimization specifically involves optimizing individual child body data and child location data, including data cleaning, data standardization, and data encoding, to obtain non-image optimized data.
[0073] The non-image-optimized data includes children's individual body optimization data and children's location optimization data;
[0074] The data cleaning process aims to improve the integrity, rationality, and consistency of the original structured data. Specifically, it involves imputing missing values and removing outliers. The imputation of missing values is used to ensure data integrity by filling in missing data items in the original data using the mean imputation method. The removal of outliers is used to ensure the rationality of data distribution by detecting and removing extreme values and logical outliers in the original data using box plots.
[0075] The data standardization is used to unify the numerical range of numerical features, specifically by normalizing the numerical data in the original data using the min-max normalization method.
[0076] The data encoding is used to convert categorical variables in structured data into numerical vectors, specifically by using one-hot encoding to encode the categorical fields in the original data.
[0077] Data synchronization is used to synchronize multi-source data in time to ensure that different types of data have consistency at the same time. Specifically, it uses dynamic time warping to match and align the time series of image-optimized data and non-image-optimized data to obtain time-aligned optimized data for dangerous behavior warnings.
[0078] Example 4, see Figure 1 and Figure 2 This embodiment is based on the above embodiment, and the module for establishing a dangerous behavior early warning model specifically includes the following steps:
[0079] Child behavior recognition is used for temporal modeling and fusion of optimized image data of children's behavior and optimized data of children's individual bodies. Specifically, the optimized image data of children's behavior is first input into a temporal attention recognition network for children's behavior to generate action sequence features. Then, the action sequence features and the optimized data of children's individual bodies are jointly input into a bidirectional long short-term memory network for cross-modal temporal feature fusion to obtain the child behavior recognition features. The formula used is as follows:
[0080] ;
[0081] In the formula, This represents the child's behavioral characteristics at time t. This indicates a bidirectional long short-term memory network operation. This represents the action sequence characteristics at time t. This represents the individual physical optimization data of a child at time t. This indicates a vector concatenation operation;
[0082] The child behavior temporal attention recognition network specifically includes the following steps:
[0083] Multi-level behavioral feature extraction is used to extract hierarchical features from the input optimized data of children's behavioral images, capturing action features from low-level contours to high-level semantics. Specifically, the optimized data of children's behavioral images is input into five concatenated residual dense blocks, where the first residual dense block receives the optimized data of children's behavioral images, and the remaining residual dense blocks sequentially use the output of the previous residual dense block as input, thus completing the extraction of action features at different levels step by step, resulting in a multi-level behavioral feature set composed of the output features of the five residual dense blocks.
[0084] The residual dense block comprises L convolutional layers. After the L convolutional layers have completed their operations, the outputs of all convolutional layers are concatenated with the initial input features of the residual dense block along the channel dimension. Then, a 1×1 convolution is used to compress the number of channels of the concatenated high-dimensional features back to the initial number of channels. Finally, the compressed features are element-wise added to the initial input feature map of the residual dense block to obtain the output features of the residual dense block. The formula used is as follows:
[0085] ;
[0086] In the formula, This represents the output of the l-th convolutional layer. This represents a mapping function consisting of a 1×1 convolution, a ReLU activation function, a 3×3 convolution, and a ReLU activation function. The initial input features represent the residual dense blocks. This represents the output of the first convolutional layer. Indicates the first The output of the convolutional layer;
[0087] A dual-branch attention enhancement method is used to enhance the spatial and temporal discriminative features of the output features of each residual dense block. Specifically, each residual dense block output feature is mapped to the query vector, key vector, and value vector of the spatial branch and temporal branch respectively through a convolution operation. Based on this, spatial attention weights are calculated to extract spatial enhanced features, and temporal attention weights are calculated to extract temporal enhanced features. Then, the spatial enhanced features and temporal enhanced features are weighted and fused according to adaptive weights, and the residuals are added to the corresponding residual dense block output features to obtain the spatiotemporal features of dangerous behavior.
[0088] The formula used is as follows:
[0089] ;
[0090] ;
[0091] ;
[0092] In the formula, This represents the output feature of the i-th residual dense block, with dimension . Where C is the number of channels, H×W is the spatial dimension, and D is the number of time frames. , and This indicates that the process is achieved through a 1×1 convolution from... The mapped spatial branch query vector, key vector, and value vector are , and This indicates that the process is achieved through a 1×1 convolution from... The mapped time-branch query vector, key vector, and value vector are... and These are the dimensions of the key vectors for the spatial branch and the temporal branch, respectively. Let represent the i-th spatial enhancement feature, used to highlight key action-related areas of the child in a single frame image. This represents the i-th temporal enhancement feature, used to highlight important frames in the video sequence that reflect the evolution of actions. Indicates the adaptive weight parameters. This represents the spatiotemporal characteristics of the i-th dangerous behavior. This represents a 1×1 convolution operation;
[0093] Multi-level behavioral feature fusion specifically involves first multiplying the output feature of the highest-level residual dense block element-wise with the corresponding spatiotemporal feature to obtain the fused feature for that level. For the remaining levels, a feature-guided fusion strategy is used to generate the fused feature for that level. Subsequently, global average pooling is performed on the fused features of each level, and they are concatenated along the channel dimension to obtain the action sequence feature. The formula used is as follows:
[0094] ;
[0095] ;
[0096] ;
[0097] In the formula, This indicates the output characteristics of the 5th residual dense block. This indicates the spatiotemporal characteristics of the fifth dangerous behavior. This represents the fusion feature at the 5th level. This represents the fusion feature at the i-th level. This indicates a channel splicing operation. Indicates the first Spatiotemporal characteristics of dangerous behaviors This indicates an element-wise multiplication operation. This represents the fusion feature at the first level. This represents the fusion feature at the second level. This represents the fusion feature at the third level. This represents the fusion feature at the fourth level. This indicates a global average pooling operation. Represents action sequence features, dimension Where M represents the number of channels after splicing, and represents the trajectory of a child's movement evolution over a continuous time period. Its data representation is a sequence of feature vectors arranged in chronological order. Each feature vector is a visual movement feature extracted from optimized data based on children's behavioral images.
[0098] Behavioral scene understanding is used to obtain scene understanding features of the child's environment and its relationship with risk factors, supplementing the scene context information of action recognition. Specifically, it processes the optimized data of the child's environmental image through a deep convolutional neural network to extract scene features. At the same time, it extracts position features from the optimized data of the child's position through a multilayer perceptron neural network. Then, the scene features and position features are concatenated into vectors and jointly mapped through a fully connected layer and a normalization layer to obtain the child's behavioral scene understanding features.
[0099] The dangerous behavior determination process involves weighting and combining the child's behavior recognition features and the child's behavior scene understanding features through an attention mechanism to generate determination features. These determination features are then input into a fully connected layer and a Softmax classifier to determine the risk level of the dangerous behavior and obtain the child's dangerous behavior recognition result.
[0100] By performing the above operations, this solution addresses the technical problem in traditional early warning systems for dangerous behaviors of children with special needs: relying solely on action recognition and lacking comprehensive analysis of the child's individual physical state, spatial location, and environmental context, leading to biased judgments of dangerous behaviors. This solution innovatively combines child behavior recognition with behavioral scene understanding and introduces an individualized analysis mechanism. By integrating optimized data on the child's individual body with action sequence features extracted from child behavior image data, and performing cross-modal temporal fusion in a bidirectional long short-term memory network, it achieves joint judgment of personalized child behavior patterns and environmental risk factors. This effectively improves the comprehensiveness, accuracy, and robustness of dangerous behavior recognition, significantly reduces the risk of misjudgment caused by single action recognition, and achieves precise identification of dangerous behaviors of children with special needs. It provides real-time, accurate, and intelligent safety measures for children with special needs. Early warning effect: Addressing the technical problem that existing methods for action and behavior recognition rely solely on shallow convolutional features, lacking joint analysis of the temporal evolution of actions and key spatial components, resulting in insufficient feature extraction, inadequate temporal dependence, and limited spatial and temporal discrimination capabilities, thus leading to low accuracy in dangerous behavior recognition, this solution innovatively proposes a children's behavior temporal attention recognition network that combines multi-level residual dense blocks, a dual-branch attention mechanism, and a feature-guided fusion strategy. This enhances the sufficiency of feature extraction, improves spatial discrimination and temporal dependence modeling capabilities, and enhances the robustness and hierarchy of feature representation. It effectively improves the accuracy and stability of action recognition, thereby further enhancing the accuracy and comprehensiveness of dangerous behavior recognition, reducing the risk of false alarms and false negatives, and achieving high-precision and intelligent recognition of children's dangerous behaviors.
[0101] Example 5, see Figure 1 , Figure 3 and Figure 4 This embodiment is based on the above embodiment, and the early warning model parameter optimization module specifically includes the following steps:
[0102] The early warning model is initially trained to update the weights and bias parameters of the model with training data under a fixed hyperparameter configuration. Specifically, the reference child danger early warning data after data optimization is used as training data, and the model parameters are iteratively updated through forward propagation and back propagation to train the early warning model and obtain the initially trained dangerous behavior early warning model.
[0103] Intelligent hyperparameter optimization is used to search for and adjust hyperparameters during model training. Specifically, it obtains the optimal combination of hyperparameters for the model through improved optimization algorithms; it includes the following steps:
[0104] Initialize the search individuals by encoding the hyperparameters of the dangerous behavior warning model into search individual position vectors, and generate N search individual position vectors through a random initialization method. Each individual is encoded to represent a candidate combination of dangerous behavior warning model hyperparameters, thus obtaining the initial search population.
[0105] The hyperparameters of the dangerous behavior early warning model include the learning rate, the number of convolutional layers per block, the number of BiLSTM layers, and the hidden dimension of the fully connected layers.
[0106] The fitness value of a search individual is calculated by calculating the fitness value of the search individual in the population; the performance of the danger behavior early warning model established based on the location of the search individual is used as the fitness value of the search individual.
[0107] The search process involves dividing individuals into groups, specifically selecting the individual with the best fitness as the leader and the rest as followers.
[0108] Leader position update is used to achieve a dynamic balance between global exploration and local development during the hyperparameter search process of the dangerous behavior early warning model. Specifically, it involves updating the leader's position through a multi-branch exploration update strategy; the formula used is as follows:
[0109] ;
[0110] In the formula, Indicates the first In the next iteration, the leader's position in the j-th dimension. Let T represent the current optimal individual position in the j-th dimension, and let T represent the current iteration number. Indicates the maximum number of iterations. and They represent Random numbers within a range and Let represent the upper and lower bounds of the search in the j-th dimension for the optimal individual, respectively. and These represent the upper and lower bounds of the global search space, respectively. This represents the probability threshold, with a value range of [value range missing]. , This represents the position of the worst individual in the j-th dimension. Represents the natural constant. express Random numbers within a certain range;
[0111] Follower position updates are used to adjust the positions of non-leader followers during the hyperparameter search of the dangerous behavior warning model, ensuring that the population as a whole converges towards the optimal solution. Specifically, the position of each follower is updated to the average of its own position and the position of the previous searched individual; the formula used is as follows:
[0112] ;
[0113] In the formula, Indicates the first In the next iteration, the position of the i-th follower individual in the j-th dimension. Indicates the first In the next iteration, the position of the i-th follower individual in the j-th dimension. Indicates the first During the nth iteration, the 1st The position of a follower individual in the j-th dimension, where i represents the index of the follower individual;
[0114] The optimal position update for a search individual involves evaluating the fitness value of the search individual after the update in the current iteration, and comparing the fitness value of the current search individual with the global optimal position of the current search individual. If the fitness value of the current search individual is better, the global optimal position of the search individual is updated.
[0115] The iterative search terminates when the fitness value of the search individual is higher than the fitness threshold or when the maximum number of iterations is reached, and the global optimal position of the search individual is obtained; the global optimal position of the search individual specifically refers to the optimal combination of hyperparameters of the model.
[0116] The final optimization of the early warning model involves adjusting the hyperparameters of the initially trained dangerous behavior early warning model based on the optimal hyperparameter combination, and using the reference children's dangerous warning data after data optimization as training data to retrain the model and update the parameters, thus obtaining the best-performing dangerous behavior early warning model.
[0117] By performing the above operations, this solution addresses the technical problems of existing hyperparameter settings for dangerous behavior early warning models being unreasonable, and the difficulty in balancing global exploration and local development during hyperparameter optimization, easily falling into local optima, resulting in poor model generalization performance and insufficient accuracy of model output results. This solution innovatively introduces an improved swarm intelligence optimization algorithm with a multi-branch exploration and update strategy to dynamically balance and optimize model hyperparameters between global search and local development. This effectively avoids local optimum traps during hyperparameter optimization, improves search space coverage and optimization convergence speed, enhances the rationality of model hyperparameter selection, improves the recognition accuracy and generalization ability of dangerous behavior early warning models, and achieves intelligent and reliable early warning of dangerous behaviors in children with special needs.
[0118] Example 6, see Figure 1 This embodiment is based on the above embodiment. Specifically, the intelligent early warning module for dangerous behavior inputs the target child's dangerous warning data after data optimization into the best-performing dangerous behavior early warning model to obtain the child's real-time dangerous behavior identification result. If the child's real-time dangerous behavior identification result is not safe, a corresponding early warning signal is automatically generated based on the child's real-time dangerous behavior identification result, and the signal is immediately notified to teachers, parents, and administrators through sound prompts or mobile terminal information push, so as to realize intelligent safety early warning and protection for special children.
[0119] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0120] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
[0121] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A special children's dangerous behavior early warning system based on artificial intelligence, characterized in that: It includes a data acquisition module, a data optimization module, a dangerous behavior early warning model establishment module, an early warning model parameter optimization module, and a dangerous behavior intelligent early warning module; The data acquisition module specifically obtains raw data for dangerous behavior warnings by collecting data. The data optimization module specifically obtains optimized data for dangerous behavior warnings through image data optimization, non-image data optimization, and data synchronization. The module for establishing a dangerous behavior early warning model specifically involves generating action sequence features by combining a child behavior temporal attention recognition network with multi-level residual dense blocks, a bi-branch attention mechanism, and a feature-guided fusion strategy. These features are then fused with the child's individual body optimization data in a bidirectional long short-term memory network to obtain child behavior recognition features. Subsequently, these features are combined with child behavior scene understanding features obtained through behavior scene understanding and weighted by an attention mechanism to generate judgment features. Finally, dangerous behavior is judged based on the judgment features to obtain the child's dangerous behavior recognition result. The early warning model parameter optimization module first completes the initial training of the early warning model, then uses an improved swarm intelligence optimization algorithm with an improved multi-branch exploration update strategy to optimize the hyperparameters, obtain the optimal hyperparameter combination of the model, and retrains the model based on the optimal hyperparameter combination to obtain the high-performance dangerous behavior early warning model. The intelligent early warning module for dangerous behavior specifically inputs the target child's dangerous warning data into the optimal dangerous behavior early warning model, outputs the child's real-time dangerous behavior identification results, and automatically generates corresponding early warning signals based on these results to notify the guardians.
2. The artificial intelligence-based early warning system for dangerous behaviors of children according to claim 1, characterized in that: The module for establishing a dangerous behavior early warning model specifically includes the following steps: The process of child behavior recognition involves first inputting optimized data of children's behavior images into a children's behavior temporal attention recognition network to generate action sequence features. Then, the action sequence features and optimized data of the child's individual body are input into a bidirectional long short-term memory network for cross-modal temporal feature fusion to obtain children's behavior recognition features. The behavioral scene understanding specifically involves processing optimized data of children's environmental images through a deep convolutional neural network to extract scene features, and simultaneously extracting position features from optimized data of children's positions through a multilayer perceptron neural network. Then, the scene features and position features are concatenated into vectors and jointly mapped through a fully connected layer and a normalization layer to obtain children's behavioral scene understanding features. The dangerous behavior determination process involves weighting and combining the child's behavioral recognition features and the child's behavioral scene understanding features through an attention mechanism to generate determination features. These determination features are then input into a fully connected layer and a Softmax classifier to determine the risk level of the dangerous behavior and obtain the result of the child's dangerous behavior identification.
3. The artificial intelligence-based early warning system for dangerous behaviors of children according to claim 1, characterized in that: The child behavior temporal attention recognition network specifically includes the following steps: Multi-level behavioral feature extraction specifically involves inputting optimized data of children's behavioral images into five concatenated residual dense blocks. The first residual dense block receives the optimized data of children's behavioral images, and the remaining residual dense blocks sequentially use the output of the previous residual dense block as input to complete the extraction of different levels of action features step by step, resulting in a multi-level behavioral feature set composed of the output features of the five residual dense blocks. The dual-branch attention enhancement method involves mapping the output features of each residual dense block to query vectors, key vectors, and value vectors of spatial and temporal branches through convolution operations. Based on this, spatial attention weights are calculated to extract spatial enhancement features, and temporal attention weights are calculated to extract temporal enhancement features. Subsequently, the spatial enhancement features and temporal enhancement features are weighted and fused according to adaptive weights, and then residuals are added to the corresponding residual dense block output features to obtain the spatiotemporal features of dangerous behavior. Multi-level behavioral feature fusion specifically involves first multiplying the output feature of the highest-level residual dense block element-wise with the corresponding spatiotemporal feature of dangerous behavior to obtain the fused feature for that level. For the remaining levels, a feature-guided fusion strategy is used to generate the corresponding fused features. Subsequently, global average pooling is performed on the fused features of each level, and they are concatenated along the channel dimension to obtain the action sequence features. The formula used is as follows: ； In the formula, This represents the fusion feature at the i-th level. This indicates a channel splicing operation. Indicates the first Spatiotemporal characteristics of dangerous behaviors This indicates an element-wise multiplication operation. This represents the spatiotemporal characteristics of the i-th dangerous behavior. This indicates the spatiotemporal characteristics of the fifth dangerous behavior. This represents a 1×1 convolution operation.
4. The artificial intelligence-based early warning system for dangerous behaviors of children according to claim 1, characterized in that: The early warning model parameter optimization module specifically includes the following steps: The initial training of the early warning model involves using the optimized reference child danger warning data as training data, iteratively updating the model parameters through forward and backward propagation, and training the early warning model to obtain an initial trained danger behavior early warning model. Intelligent hyperparameter optimization; The final optimization of the early warning model involves adjusting the hyperparameters of the initially trained dangerous behavior early warning model based on the optimal hyperparameter combination, and using the reference children's dangerous warning data after data optimization as training data to retrain the model and update the parameters, thus obtaining the best-performing dangerous behavior early warning model.
5. The artificial intelligence-based early warning system for dangerous behaviors of children according to claim 4, characterized in that: The intelligent optimization of hyperparameters specifically includes the following steps: Initialize the search individuals by encoding the hyperparameters of the dangerous behavior warning model into search individual position vectors, and generate N search individual position vectors through a random initialization method. Each individual is encoded to represent a candidate combination of dangerous behavior warning model hyperparameters, thus obtaining the initial search population. The fitness value of a search individual is calculated by calculating the fitness value of the search individual in the population; the performance of the danger behavior early warning model established based on the location of the search individual is used as the fitness value of the search individual. The search process involves dividing individuals into groups, specifically selecting the individual with the best fitness as the leader and the remaining individuals as followers. Leader's individual position update is specifically achieved through a multi-branch exploration update strategy. Follower individual position update, specifically, updating the position of each follower individual to the average position of itself and the previous search individual; The optimal position update for a search individual involves evaluating the fitness value of the search individual after the update in the current iteration, and comparing the fitness value of the current search individual with the global optimal position of the current search individual. If the fitness value of the current search individual is better, the global optimal position of the search individual is updated. The iterative search terminates when the fitness value of the search individual is higher than the fitness threshold or when the maximum number of iterations is reached, and the global optimal position of the search individual is obtained; the global optimal position of the search individual specifically refers to the optimal combination of hyperparameters of the model.
6. The artificial intelligence-based early warning system for dangerous behaviors of children according to claim 1, characterized in that: The formula used to update the leader's individual position is as follows: ； In the formula, Indicates the first In the next iteration, the leader's position in the j-th dimension. Let T represent the current optimal individual position in the j-th dimension, and let T represent the current iteration number. Indicates the maximum number of iterations. and They represent Random numbers within a range and Let represent the upper and lower bounds of the search in the j-th dimension for the optimal individual, respectively. and These represent the upper and lower bounds of the global search space, respectively. Indicates the probability threshold. This represents the position of the worst individual in the j-th dimension. express A random number within a given range.
7. The artificial intelligence-based early warning system for dangerous behaviors of children according to claim 1, characterized in that: The intelligent early warning module for dangerous behavior specifically inputs the target child's dangerous warning data, which has been optimized, into the best-performing dangerous behavior early warning model to obtain the child's real-time dangerous behavior identification result. If the child's real-time dangerous behavior identification result is not safe, the module automatically generates a corresponding early warning signal based on the child's real-time dangerous behavior identification result and immediately notifies teachers, parents, and administrators.
8. The artificial intelligence-based early warning system for dangerous behaviors of children according to claim 1, characterized in that: The raw data for dangerous behavior warnings includes reference child danger warning data and target child danger warning data; Both the reference child danger warning data and the target child danger warning data include child behavior image data, child individual body data, child environmental image data, and child location data; the reference child danger warning data also includes reference child dangerous behavior identification results.
9. The artificial intelligence-based early warning system for dangerous behaviors of children according to claim 1, characterized in that: The data optimization module specifically includes the following steps: Image data optimization specifically involves optimizing children's behavioral image data and children's environmental image data, including image cleaning, image format standardization, and image quality enhancement, to obtain optimized image data. Non-image data optimization specifically involves optimizing individual children's body data and children's location data, including data cleaning, data standardization, and data encoding, to obtain non-image optimized data; Data synchronization specifically involves using a dynamic time warping method to perform time-series matching and alignment between image-optimized data and non-image-optimized data, thereby obtaining time-aligned optimized data for dangerous behavior warnings.