Drowning identification monitoring method based on multi-dimensional feature integrated network
Through multi-dimensional feature integration network and weighted cosine similarity calculation, the accuracy and response speed problems of traditional drowning monitoring methods are solved, and more efficient and accurate drowning identification and monitoring are achieved, ensuring the safety of swimmers.
Patent Information
- Application Number
- CN202411158056.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-22
- Publication Date
- 2024-12-03
AI Technical Summary
Traditional drowning monitoring methods cannot accurately identify and warn of drowning risks, resulting in incomplete monitoring and untimely response.
The multi-dimensional feature integration network is used to combine the BS neural network model and weighted cosine similarity calculation. By obtaining the body characteristic attributes of swimmers, building feature spaces, training neural networks, and correlation with drowning behavior is carried out, and local and global weighted cosine similarity is used. Calculate to realize real-time monitoring and drowning risk identification.
It improves the accuracy and response speed of drowning recognition, enhances the effect of water safety monitoring, reduces the occurrence of drowning accidents, and provides a safer environment for swimmers.
Smart Images

Figure CN119068414A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of drowning recognition intelligent monitoring, and in particular to a drowning recognition monitoring method of a multi-dimensional feature integrated network. Background Art
[0002] In traditional drowning monitoring methods, there are often problems of incomplete monitoring and untimely response. In view of these limitations, the present invention introduces advanced technical means and methods. The present invention aims to provide a drowning identification and monitoring method based on a multidimensional feature integration network to solve the problems existing in traditional methods. Traditional methods are often unable to accurately identify and warn of drowning risks. The present invention overcomes the limitations of traditional methods through a multidimensional feature integration network and weighted cosine similarity calculation, and achieves more efficient and accurate drowning identification and monitoring. By introducing a real-time monitoring system, the present invention further ensures the safety of drowning persons. Summary of the invention
[0003] The technical problem to be solved by the present invention is to provide a drowning recognition and monitoring method based on a multi-dimensional feature integrated network. The present invention introduces a BS neural network model and a multi-dimensional feature set, and is equipped with a weighted similarity calculation formula, aiming to timely discover and handle drowning incidents and improve the effect of water safety monitoring.
[0004] To achieve the above object, the present invention provides a drowning recognition and monitoring method based on a multi-dimensional feature integrated network, comprising the following steps:
[0005] S1. Obtain the physical characteristics of the swimmer to construct and train the feature space, wherein the anomaly detection formula is used to monitor the skin color during drowning, the electromyographic signal generation formula is used to draw the spectrum, and the bubble area calculation formula is used to calculate the bubble area;
[0006] S2, construct the dataset space of each swimmer and map it into the global coordinate system;
[0007] S3, training the neural network structure, building the association between normal swimmers and their behavioral characteristic attributes and drowning people and their behavioral characteristic attributes, wherein the Boosting model outputs the result formula, uses the attention mechanism calculation formula to calculate the model result weight, and uses the loss function calculation formula to obtain the model training loss;
[0008] S4. After the neural network is constructed and trained, similarity is calculated for the new swimmer data and the normal person feature vector and the drowning person feature vector, respectively. The local weighted cosine similarity formula is used to calculate the local similarity, the global weighted cosine similarity formula is used to calculate the global similarity, and the fusion formula is used to calculate the final similarity value.
[0009] S5. Apply the trained model to real-time monitoring of swimming scenes.
[0010] Furthermore, the anomaly detection formula in S1 is:
[0011]
[0012] Where d is the distance, which indicates the similarity between the two distributions. The larger the value, the greater the difference between the distributions. 1 H 2 Represent two histograms respectively. represents the mean of the two histograms, and i is the index of the histogram, indicating the position of each entry in the histogram. The closer this value is to 0, the more similar the two distributions are; the larger the value, the greater the difference between the two distributions. In skin color change detection, by comparing the color histogram of the current skin area with the baseline histogram under normal conditions, abnormal changes in skin color can be detected, thereby identifying potential precursors to drowning.
[0013] Furthermore, the S1 electromyographic signal generation formula is:
[0014]
[0015] Where x is the input signal, N is the signal length, and k is the frequency index.
[0016] Furthermore, the S1 bubble area calculation formula is:
[0017]
[0018] Where N is the total number of vertices, x i y i is the coordinate of the i-th vertex, and Area is the area of the bubble.
[0019] Furthermore, the output formula of the S3 Boosting model is:
[0020]
[0021] Among them, F m (x) represents the output of the mth round of Boosting model, a i is the weight of the i-th weak classifier, h i (x) represents the i-th weak classifier.
[0022] Furthermore, the calculation formula of the S3 attention mechanism is:
[0023] A=softmax(WZ+b)
[0024] H=A⊙Z
[0025] Where A is the attention weight, W and b are learnable weights and biases, and H is the weighted feature.
[0026] Furthermore, the S3 loss function calculation formula is:
[0027]
[0028] Where L is the loss function, y i is the actual sample label, is the predicted sample label, and N is the number of samples.
[0029] Furthermore, the local weighted cosine similarity formula in S4 is:
[0030]
[0031] Where L represents the local similarity value, X represents each new swimmer data, N represents the normal person feature vector, D represents the drowning person feature vector, and w = {w1, w2, w3, w4...} represents the local feature weight.
[0032] Furthermore, the global weighted cosine similarity formula in S4 is:
[0033]
[0034] Where Q represents the global similarity value, X represents each new swimmer data, N represents the normal person feature vector, D represents the drowning person feature vector, and W = {W1, W2, W3, ...} represents the global feature weight.
[0035] Furthermore, the fusion formula in S4 is:
[0036] S=α*L+(1-α)*Q
[0037] Here, α is a hyperparameter used to balance the contribution of local and global similarity.
[0038] Compared with the prior art, the present invention has the following advantages:
[0039] High-precision identification: By introducing the BS neural network model and multi-dimensional feature set, the present invention can more accurately identify normal swimming behavior and drowning behavior. Combined with the weighted cosine similarity calculation method, drowning risk identification is more accurate.
[0040] Real-time monitoring: The present invention applies the trained model to the real-time monitoring system, which can continuously monitor the behavioral characteristics of swimmers, calculate similarities in real time, and immediately issue an alarm when drowning risks are detected, greatly improving the response speed.
[0041] Multi-dimensional feature fusion: Utilizing a multi-dimensional feature integration network, multiple features such as facial expressions, skin color, electromyographic signals, drowning bubbles, sound amplitude, water pressure changes, thermal imaging, etc. are fully integrated to improve the comprehensiveness and accuracy of monitoring.
[0042] Strong adaptability: The importance of features is calculated through information gain or other feature selection methods, and the features are weighted, so that the model can adapt to the behavioral characteristics of different swimmers and in different environments, and has strong adaptability and robustness.
[0043] Improve safety: The present invention can detect and handle drowning incidents in a timely manner, significantly improve the effect of water area safety monitoring, reduce the occurrence of drowning accidents, and provide a safer environment for swimmers. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:
[0045] Figure 1 It is a flowchart of the drowning recognition and monitoring method based on multi-dimensional feature integration network;
[0046] Figure 2 This is a facial expression diagram of a drowning recognition and monitoring method based on a multi-dimensional feature integration network proposed by the present invention.
[0047] Figure 3 This is a skin color diagram of a drowning recognition and monitoring method based on a multi-dimensional feature integration network proposed by the present invention.
[0048] Figure 4 This is an electromyographic signal diagram of a drowning recognition and monitoring method based on a multi-dimensional feature integration network proposed by the present invention.
[0049] Figure 5 A drowning bubble diagram of a drowning recognition and monitoring method based on a multi-dimensional feature integrated network proposed by the present invention.
[0050] Figure 6 A sound amplitude diagram of a drowning recognition and monitoring method based on a multi-dimensional feature integration network proposed by the present invention.
[0051] Figure 7 This is a thermal imaging diagram of a drowning recognition and monitoring method based on a multi-dimensional feature integration network proposed by the present invention.
[0052] Figure 8 This is a global coordinate mapping diagram of a drowning recognition and monitoring method based on a multi-dimensional feature integration network proposed by the present invention.
[0053] Fig. 9This is a BS neural network structure diagram of a drowning recognition and monitoring method based on a multi-dimensional feature integration network proposed in the present invention.
[0054] Fig.10 This is a similarity calculation structure diagram of a drowning recognition and monitoring method based on a multi-dimensional feature integration network proposed in the present invention. DETAILED DESCRIPTION
[0055] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. It should be noted that the words "front", "rear", "left", "right", "upper" and "lower" used in the following description refer to directions in the accompanying drawings, and the words "inner" and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.
[0056] refer to Figure 1 , a drowning recognition and monitoring method based on multi-dimensional feature integration network
[0057] The method comprises the following steps:
[0058] S1. Obtain the physical characteristics of the swimmer to construct and train the feature space.
[0059] In this implementation, S1 specifically includes:
[0060] S11. By installing a high-resolution camera to capture the swimmer's facial expressions, a micro-expression recognition algorithm (Ekman micro-expression analysis) is used to detect whether the swimmer shows signs of drowning such as panic and tension.
[0061] like Figure 2 As shown in the figure, based on the Ekman algorithm, face detection is first performed to locate the positions of the two faces in the image; then the key feature points of each face are extracted. Subsequently, action unit detection is performed. For the "calm" expression on the left, few action units are detected, and the facial muscle movement is not significant; for the "frightened" expression on the right, multiple action units are detected, such as eyebrow raising (AU1, AU2), eyes widening (AU5), mouth stretching (AU20, AU26), etc. Finally, the detected action units are mapped to emotional states. The left side shows the facial expression of a person in the "calm" state. This expression usually indicates that the swimmer is in a normal and safe state. The right side shows the facial expression of the same person in the "frightened" state. This expression usually indicates that the swimmer may be in a dangerous state of drowning or about to drown, and needs timely attention and rescue.
[0062] S12. Use video image processing technology to detect changes in the swimmer's skin color. When drowning, the skin may change color (such as whitening or bluish) due to lack of oxygen. These changes can be detected through color space conversion and color histogram analysis.
[0063] First, perform color space conversion to convert the original image from RGB color space to HSV color space; secondly, use the threshold segmentation method in HSV color space to extract the skin area. By setting the threshold range, the skin area can be effectively segmented and a skin mask can be generated. In this article, the threshold range is set to be the Hue range: (0, 50), Saturation range: (20, 255), Value range: (70, 255); then, apply the skin mask to the original image and retain only the skin area; then, calculate the color histogram of the skin area and compare it with the normal skin color histogram; finally, based on the changes in the color histogram, determine whether the skin color has turned white or blue to detect potential drowning situations. The anomaly detection formula provided by this patent is:
[0064]
[0065] Where d is the distance, which indicates the similarity between the two distributions. The larger the value, the greater the difference between the distributions. 1 H 2 Represent two histograms respectively. represents the mean of the two histograms, and i is the index of the histogram, indicating the position of each entry in the histogram. The closer this value is to 0, the more similar the two distributions are; the larger the value, the greater the difference between the two distributions. In skin color change detection, by comparing the color histogram of the current skin area with the baseline histogram under normal conditions, abnormal changes in skin color can be detected, thereby identifying potential precursors to drowning.
[0066] like Figure 3 As shown, the left image shows the extracted skin regions, and the two faces show calm and frightened expressions respectively. The right image shows the color histograms of the skin regions of the two faces. The red curve represents the color distribution of the skin region of the frightened expression, and the green curve represents the color distribution of the skin region of the calm expression. The main peak of the red histogram is located in the higher hue area, and the frequency is significantly concentrated between hues of 75 and 125. This shows that the skin color of the frightened expression tends to the higher hue area. The main peak of the green histogram is located in the lower hue area, and the frequency is concentrated between hues of 0 and 50. This shows that the skin color of the calm expression is mainly distributed in the lower hue area. The skin region of the frightened expression shows a significant change in hue, which may be due to the whitening or bluish skin caused by lack of oxygen. The hue distribution of the skin region of the calm expression is relatively normal. Therefore, by analyzing the color histogram, it can be determined that the right face (frightened expression) is at risk of drowning due to skin color changes.
[0067] S13. Install electromyographic sensors underwater at the edge of a swimming pool or reservoir to detect the swimmer's muscle activity. When drowning, muscles may contract violently or become weak, and these characteristics can be detected through electromyographic signals.
[0068] like Figure 4 The figure shows the spectrum of myoelectric signal. The horizontal axis frequency (Hz) represents the different frequency components of the signal. The main focus is on the frequency range of 0 to 500 Hz, which is the common frequency range of myoelectric signal. The vertical axis amplitude represents the intensity of each frequency component, and its generation formula is as follows:
[0069]
[0070] Where x is the input signal, N is the signal length, and k is the frequency index.
[0071] It can be seen that there is an obvious peak at 50Hz, which is the main frequency component of the simulated signal. This indicates that the EMG signal has strong activity at 50Hz. In practical applications, by analyzing the EMG signal spectrum of a swimmer, abnormal changes in muscle activity can be detected. If abnormal high-frequency or low-frequency components appear in the spectrum, it may indicate severe muscle contraction or weakness, thereby identifying the risk of drowning.
[0072] S14. Use a highly sensitive underwater camera to capture the bubble pattern produced by a swimmer’s breathing. Bubbles may increase or decrease when drowning, and use image analysis technology to detect changes in the frequency and size of bubbles.
[0073] like Figure 5 As shown, the area of each bubble in the figure can be expressed by the following formula:
[0074]
[0075] Where N is the total number of vertices, x i y i are the coordinates of the i-th vertex.
[0076] In addition, the frequency of bubbles in the image can be calculated based on the image information, and these features can be used to detect changes in the swimmer's breathing pattern. For example, a sharp increase or decrease in the number of bubbles may occur during drowning, and by analyzing these features, potential drowning risks can be identified.
[0077] S15. Install underwater microphones to monitor the sound characteristics of swimmers, including cries for help, sounds of struggle, etc. Analyze the sound characteristics through sound recognition technology (such as convolutional neural network CNN) to detect abnormal sound signals.
[0078] First, extract features from the returned sound signal. Here, the sound signal is converted to the frequency domain through Fourier transform, and then the Mel frequency cepstral coefficients are calculated. Figure 6 , there is a significant peak between about 1.5Hz and 3.5Hz in the figure, and the maximum amplitude exceeds 1.0. This shows that the sound signal has strong frequency components in this frequency range. In the range of 0Hz to 1.5Hz and 3.5Hz to 10Hz, the amplitude is relatively low and stable, which indicates that these frequency components are weak. Through the above analysis, it can be seen that this spectrum difference map is used to detect specific frequency components in the sound signal and analyze its amplitude changes to identify potential abnormal sound signals (such as drowning calls or struggling sounds).
[0079] S16. Install a water pressure sensor underwater to detect changes in water pressure around the swimmer. When drowning, the swimmer may sink, causing significant changes in the surrounding water pressure. The pressure sensor is used to capture these changes. The pressure values are aggregated into a table:
[0080] P={P1,P2…Pn}
[0081] S17. Install thermal imaging cameras to detect changes in the swimmer’s body temperature. When drowning, the body temperature may drop rapidly, and thermal imaging technology can be used to capture these temperature changes.
[0082] White circles are used to represent high-temperature areas, and they are labeled "human body temperature areas". Histogram equalization is used to enhance the image contrast, making the temperature change more obvious, and they are labeled "enhanced contrast". The temperature range in the heat map is extracted, and the pixel values related to human body temperature are filtered out, and they are labeled "pixels within the temperature range". The heat map is segmented, and the human body temperature areas are identified and marked, and they are labeled "temperature areas after segmentation".
[0083] The final result is Figure 7 As shown in the figure, brighter areas in the image represent higher temperatures and darker areas represent lower temperatures. "Human temperature area" marks the human temperature area, indicating that this is the high temperature area detected. The second histogram equalization enhances the image contrast, making the temperature changes more obvious. "Enhanced contrast" marks the area where the contrast is enhanced after the image is processed by histogram equalization. The third image shows the filtered image, retaining only the pixel values related to human body temperature. Temperature range filtering helps to filter out background and other interference. The fourth image shows the results of image segmentation, identifying and marking the human temperature area. Image segmentation is used to extract the target area for further analysis of temperature changes.
[0084] Through the acquisition of original thermal images, histogram equalization, temperature range screening and image segmentation, the changes in the swimmer's body surface temperature can be effectively detected, abnormal situations such as drowning can be identified, and the accuracy and response speed of water safety monitoring can be improved.
[0085] S2. Construct the dataset space of each swimmer and map it into the global coordinate system.
[0086] Based on the data obtained in step S1, for the convenience of calculation, facial expressions, skin color, electromyographic signals, drowning bubbles, sound amplitude, water pressure changes, and thermal imaging are first mapped into the following formula:
[0087] {Facial expression: 1, Skin color: 2, EMG signal: 3, Drowning bubbles: 4, Sound amplitude: 5, Water pressure change: 6, Thermal imaging: 7}
[0088] That is, the facial expression feature is mapped to action 1, skin color is mapped to action 2, electromyography is mapped to action 3, drowning bubbles are mapped to action 4, sound amplitude is mapped to action 5, water pressure changes are mapped to action 6, and thermal imaging is mapped to action 7.
[0089] Then each action has two attributes, namely: {Facial expression: [calm, panic}, Skin color: {normal, pale or blue}, EMG signal: {normal, high frequency or low frequency}, Drowning bubbles: {normal, sharp increase or decrease}, Sound amplitude: {normal, sharp}, Water pressure change: {normal, increase}, Thermal imaging: {higher, lower}}, Map the attributes of each action to 0, 1. That is, the final representation is:
[0090] {1:{0,1},2:{0,1},3:{0,1},4:{0,1},5:{0,1},6:{0,1},7:{0,1}}
[0091] The global coordinate system of the reservoir is established, and the target position information of each mode is transformed to map the position information of each swimmer to the global coordinate system. Figure 8 As shown,
[0092] In the figure, the X-axis represents the action space, 1, 2, 3, 4, 5, 6, and 7 represent facial expressions, skin color, electromyographic signals, drowning bubbles, sound amplitude, water pressure changes, and thermal imaging, respectively. The Y-axis represents the attributes of each action, and each color in the figure represents the same swimmer.
[0093] By mapping the various characteristic attributes of each swimmer to the global coordinate system, the status of each person on different characteristics can be intuitively displayed. This method can not only facilitate the visualization analysis of data, but also help identify potential abnormal behaviors or characteristic patterns.
[0094] S3. Train the neural network structure to build the association between normal swimmers and their behavioral characteristic attributes and between drowning people and their behavioral characteristic attributes.
[0095] like Fig. 9 As shown, in this patent, the BS neural network is used to construct the association between normal swimmers and their behavioral characteristic attributes and drowning people and their behavioral characteristic attributes. The BS neural network combines the advantages of Boosting and Stacking, and introduces an attention mechanism in the Stacking stage, paying special attention to factors with great influence.
[0096] First, the facial expressions, skin color, electromyographic signals, drowning bubbles, sound amplitude, water pressure changes, and thermal imaging of normal and drowning people are input into the model. The input data is standardized and normalized, and data enhancement techniques such as adding noise and transformation are used to increase the diversity and robustness of the data.
[0097] Secondly, in the Boosting stage, the Boosting algorithm is used to perform multiple rounds of training on the data to generate multiple enhanced basic classifiers. The Boosting algorithm gradually improves the performance of the classifier by adjusting the model weights, and obtains a series of enhanced classifier output results. Specifically, the model output of the Boosting stage can be expressed as:
[0098]
[0099] Among them, F m (x) represents the output of the mth round of Boosting model, a i is the weight of the i-th weak classifier, h i (x) represents the i-th weak classifier.
[0100] Then, the output results of multiple enhanced classifiers generated in the Boosting stage are used as new features to form a new feature set. These feature sets are input into the Stacking model for final classification. At this point, the new feature set can be expressed as:
[0101] Z=[h 1 (x),h 2 (x),h 3 (x)…h n (x)]
[0102] Among them, Z represents the new feature set, h i (x) represents the output of the i-th enhanced classifier.
[0103] In the Stacking stage, the logistic regression model is used as a meta-learner for training. In order to pay special attention to the features with great influence, the attention mechanism is introduced in the Stacking model. By calculating the attention weight, special attention is paid to the important features and weighted processing is performed on them. The attention mechanism can be expressed as:
[0104] A=softmax(WZ+b)
[0105] H=A⊙Z
[0106] Where A is the attention weight, W and b are learnable weights and biases, and H is the weighted feature.
[0107] Finally, the Stacking model is used to perform the final classification on the weighted features, and the final prediction result can be expressed as:
[0108]
[0109] in is the final prediction result, and g(H) is the output of the Stacking model.
[0110] To train the model, we first collected data on normal behavior of swimmers swimming in a reservoir and abnormal data on drowning. With this data, the model was able to learn the difference between normal and abnormal behavior.
[0111] The loss function used during model training is:
[0112]
[0113] Where L is the loss function, y i is the actual sample label, is the predicted sample label, and N is the number of samples.
[0114] The learning rate is set to 0.01, and the model parameters are gradually updated through optimization algorithms (such as Adam optimizer) to minimize the loss function. The decrease in model loss is recorded in the loss decrease table to monitor the performance changes during model training.
[0115]
[0116]
[0117] Through this innovative method combining Boosting and Stacking, the BS neural network can effectively build the association between normal swimmers and their behavioral characteristics and between drowning people and their behavioral characteristics. The model is cross-validated and hyperparameters are tuned to further optimize performance, and is ultimately applied to real-time monitoring systems and multimodal data fusion to improve classification performance and detection accuracy.
[0118] S4. After the neural network is constructed and trained, similarity calculations are performed on the new swimmer data and the normal person feature vector and drowning person feature vector. Fig.10As shown. In order to further improve the accuracy and robustness of the model, this patent proposes a multiple weighted cosine similarity calculation method. This method not only takes into account the weighting of features, but also introduces the fusion of local and global similarities. Specifically, for each new swimming personnel data X, its local weighted cosine similarity and global weighted cosine similarity with the normal person feature vector N and the drowning person feature vector D are calculated. The local weighted cosine similarity formula is:
[0119]
[0120] The global weighted cosine similarity formula is:
[0121]
[0122] Among them, w = {w1, w2, w3, w4...} represents the local feature weight, and W = {W1, W2, W3,...} represents the global feature weight.
[0123] Then, the local and global similarities are combined to obtain a fused similarity score:
[0124] S=α*L+(1-α)*Q
[0125] Here, α is a hyperparameter used to balance the contribution of local and global similarity.
[0126] Finally, the weighted similarity calculation method is used to calculate the similarity between the new swimmer data and the normal person feature vector and the drowning person feature vector. By comparing the weighted similarity values, it is determined whether the new swimmer data is closer to a normal person or a drowning person, so as to perform classification warning. This method not only combines the advantages of Boosting and Stacking, but also improves the accuracy of similarity calculation through feature weighting and similarity fusion, and can more effectively build the association between normal swimmers and their behavioral characteristic attributes and between drowning people and their behavioral characteristic attributes.
[0127] S5. Finally, the trained model is applied to the real-time monitoring of swimming scenes through the real-time monitoring system. The system can continuously monitor the behavioral characteristics of swimmers, calculate similarities in real time, and immediately issue an alarm when drowning risks are detected, thereby improving the response speed and accuracy of drowning incidents. Through real-time monitoring, the system can effectively utilize the classification performance and detection accuracy of the model to ensure the safety of swimmers.
[0128] It should be noted that the above content only illustrates the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.
Claims
1. A drowning recognition and monitoring method based on a multi-dimensional feature integration network, which is characterized by the following method steps: S1. Obtain the physical characteristics of the swimmer to construct and train the feature space, wherein the anomaly detection formula is used to monitor the skin color during drowning, the electromyographic signal generation formula is used to draw the spectrum, and the bubble area calculation formula is used to calculate the bubble area; S2, construct the dataset space of each swimmer and map it into the global coordinate system; S3, training the neural network structure, building the association between normal swimmers and their behavioral characteristic attributes and drowning people and their behavioral characteristic attributes, wherein the Boosting model outputs the result formula, uses the attention mechanism calculation formula to calculate the model result weight, and uses the loss function calculation formula to obtain the model training loss; S4. After the neural network is constructed and trained, similarity is calculated for the new swimmer data and the normal person feature vector and the drowning person feature vector, respectively. The local weighted cosine similarity formula is used to calculate the local similarity, the global weighted cosine similarity formula is used to calculate the global similarity, and the fusion formula is used to calculate the final similarity value. S5. Apply the trained model to real-time monitoring of swimming scenes.
2. According to claim 1, the drowning recognition and monitoring method based on multi-dimensional feature integrated network, Features: The anomaly detection formula in S1 is: Where d is the distance, which indicates the similarity between the two distributions. The larger the value, the greater the difference between the distributions. 1 H 2 Represent two histograms respectively. represents the mean of the two histograms, and i is the index of the histogram, indicating the position of each entry in the histogram. The closer this value is to 0, the more similar the two distributions are; the larger the value, the greater the difference between the two distributions. In skin color change detection, by comparing the color histogram of the current skin area with the baseline histogram under normal conditions, abnormal changes in skin color can be detected, thereby identifying potential precursors to drowning.
3. The drowning recognition and monitoring method based on a multidimensional feature integrated network according to claim 1, Features: The S1 electromyographic signal generation formula is: Where x is the input signal, N is the signal length, and k is the frequency index.
4. The drowning recognition and monitoring method based on a multidimensional feature integrated network according to claim 1, Features: The S1 bubble area calculation formula is: Where N is the total number of vertices, x i y i is the coordinate of the i-th vertex, and Area is the area of the bubble.
5. The drowning recognition and monitoring method based on a multi-dimensional feature integrated network according to claim 1, Features: The output formula of the S3 Boosting model is: Among them, F m (x) represents the output of the mth round of Boosting model, a i is the weight of the i-th weak classifier, h i (x) represents the i-th weak classifier.
6. The drowning recognition and monitoring method based on a multi-dimensional feature integrated network according to claim 1, Features: The calculation formula of the S3 attention mechanism is: A=softmax(WZ+b) H=A⊙Z Where A is the attention weight, W and b are learnable weights and biases, and H is the weighted feature.
7. The drowning recognition and monitoring method based on a multi-dimensional feature integrated network according to claim 1, Features: The S3 loss function calculation formula is: Where L is the loss function, y i is the actual sample label, is the predicted sample label, and N is the number of samples.
8. The drowning recognition and monitoring method based on a multi-dimensional feature integrated network according to claim 1, Features: The local weighted cosine similarity formula in S4 is: Where L represents the local similarity value, X represents each new swimmer data, N represents the normal person feature vector, D represents the drowning person feature vector, and w = {w1, w2, w3, w4...} represents the local feature weight.
9. The drowning recognition and monitoring method based on a multi-dimensional feature integrated network according to claim 1, Features: The global weighted cosine similarity formula in S4 is: Where Q represents the global similarity value, X represents each new swimmer data, N represents the normal person feature vector, D represents the drowning person feature vector, and W = {W1, W2, W3, ...} represents the global feature weight.
10. The drowning recognition and monitoring method based on multi-dimensional feature integrated network according to claim 1, Features: The fusion formula in S4 is: S=α*L+(1-α)*Q Here, α is a hyperparameter used to balance the contribution of local and global similarity.
Citation Information
Cited By
Intelligent wearable device for people on water and method for positioning people on water
CN119846682A
Intelligent wearable device for water personnel and water personnel positioning method
CN119846682B
