Baby posture intelligent detection and early warning monitoring system

Abstract

The invention discloses an infant posture intelligent detection and early warning monitoring system which is characterized in that the system comprises an information terminal, a processing side end and a display terminal, and the information terminal comprises a camera and monitors the posture and state of an infant in real time; images and posture information acquired by the camera are transmitted to the processing side end through wireless signals to be analyzed and processed; the processing side end comprises a small computing device with an AI function and a server with an AI processing capability; the small computing device performs preliminary data processing, and then sends the processed data or unprocessed original data to a server with stronger AI processing capability for further analysis; the server analyzes the posture data of the baby through a deep learning algorithm, detects abnormal postures and makes early warning judgment, and an analysis result is transmitted to the display terminal through a wireless signal; the display terminal comprises a plurality of display devices. According to the invention, intelligent detection and real-time early warning of baby postures can be realized, and the stability is good.

Description

Technical Field

[0001] The present invention relates to the technical field of monitoring equipment, and more specifically to a child posture intelligent detection and early warning monitoring system. Background Art

[0002] Sudden Unexplained Death in Infancy / An Infant (SUDI) is defined as the sudden death of an infant, which can be divided into categories such as Sudden Infant Death Syndrome (SIDS), death due to uncertain or unknown causes, and Accidental Infant Suffocation and Strangulation in Bed (ASSB).

[0003] SIDS refers to a syndrome that causes sudden death in infants and young children. This disease is completely unpredictable based on the health status and medical history of the child, and conventional pathological anatomy cannot find an obvious cause of death. Sudden infant death syndrome is the most common cause of death in infants aged 2 weeks to 1 year, accounting for 30% of the mortality rate in this age group. There are many theories about its pathogenesis, but infection, allergy, pneumonia and vaccination have been proven to be unrelated to sudden infant death syndrome. A widely accepted theory is that the center responsible for sleep and wakefulness in the brain of some infants is delayed, making them prone to stop breathing in certain situations. Many studies have linked prone sleep to a high risk of sudden infant death syndrome. The treatment and prevention of SIDS often emphasizes cardiopulmonary monitoring, and cardiopulmonary monitors are an effective means of monitoring. Monitors are often connected to alarm devices, which automatically alarm when the baby's heart rate drops below 80 beats / minute or the respiratory arrest time exceeds 20 seconds. However, due to the rapid development of SIDS and the low level of popularization of family infant first aid skills, the mortality rate of SIDS children with respiratory arrest is still high, and SIDI is difficult to diagnose before the baby is alive. It is generally believed that maintaining a good supine sleeping position for infants and young children can prevent SIDS. According to reports, in 1994, the American Academy of Pediatrics, the American Sudden Infant Death Syndrome Association and others jointly launched the "Supine Sleeping" campaign, and the incidence of sudden infant death syndrome in the United States dropped by 40%.

[0004] ASSB is a state of blood hypoxia caused by internal or external obstruction of the infant's respiratory tract. It is named because aspiration asphyxia and accidental asphyxia in bed are the two most common types of asphyxia in infants. Infancy is a dangerous age for ASSB. According to reports, the leading cause of death for children aged 0 to 14 in China is accidental injury, and the incidence of accidental asphyxia among infant deaths is 22.86%. The incidence of accidental asphyxia among the causes of death of children under 1 year old reported by Sun Li'an et al. is even higher (70.19%). If it is not properly handled after the occurrence, it will cause cerebral ischemia and hypoxia or even death in 3 to 5 minutes. Infant asphyxia seriously threatens the life safety of infants. There are many specific causes of infant ASSB. In addition to prone sleep asphyxia, there are also reports of foreign objects, vomitus, or even the body of the parents in the same bed blocking the mouth and nose. Zang Chuannong et al. reported a case of a 29-day-old baby suffocated by a plastic bag. When the baby was unattended, the baby's upper limbs moved and the plastic bag next to the pillow covered his mouth and nose. He was found half an hour later with cyanotic lips and poor reaction. However, the baby did not develop SUDI because the discovery and treatment were timely. It can be seen that timely discovery of infant ASSB can prevent the occurrence of SUDI to a certain extent.

[0005] SUDI in infants can bring great pain and harm to society and families. Therefore, ensuring that infants are in the recommended sleeping position to prevent SIDS and timely detecting ASSB risk events are of positive significance in preventing SUDI.

[0006] Currently, there are reports of strengthening the training of parents and other relevant personnel to improve their awareness of prevention, thereby preventing SIDS and ASSB. However, since it is difficult for manpower to be on duty 24 hours a day, when there is no one on duty, such as when the caregiver temporarily leaves the baby or the caregiver sleeps at night, the baby's sleeping posture changes and ASSB are often not discovered in time, resulting in baby SUDI. Similarly, even 24-hour human video surveillance is not necessarily reliable. In the past ten years, there have been news reports of medical staff failing to check the surveillance in time and infant suffocation not being treated in time, leading to SUDI. Therefore, how to ensure that when the baby is unattended, the baby's sleeping posture can still be monitored and identified, and the risk of ASSB events can be warned in real time, thereby preventing SUDI is a key scientific issue.

[0007] Currently, artificial intelligence (AI) has penetrated into every aspect of daily life. Among them, the application of AI tools has brought unprecedented changes to the way humans work. These tools use advanced algorithms and machine learning technologies to help humans complete tasks more efficiently and improve productivity. There are many types of AI tools, covering everything from basic data analysis, voice, and image recognition to advanced automated decision-making systems. Given the development of AI technology, in theory, AI should be able to identify ASSB risk events such as changes in infant sleeping position, changes in movement, and mouth and nose occlusion when there is no one on duty.

[0008] Therefore, the present invention aims to propose a system that can realize long-term automatic monitoring and early warning of sudden infant death risk events by combining video surveillance and artificial intelligence vision technology. Summary of the invention

[0009] The purpose of the present invention is to solve the deficiencies of the prior art and to provide a stable and reliable baby posture intelligent detection and early warning monitoring system that can accurately identify and monitor the baby's posture.

[0010] The present invention adopts the following technical solutions to achieve the above-mentioned purpose: an intelligent detection and early warning monitoring system for infant posture, characterized in that it includes an information terminal, a processing edge, and a display terminal.

[0011] The information terminal includes a camera to monitor the baby's posture and status in real time; the image and posture information obtained by the camera is transmitted to the processing edge via wireless signals for analysis and processing.

[0012] The processing edge includes a small computing device with AI function and a server with AI processing capabilities; the small computing device performs preliminary data processing, and then sends the processed data or unprocessed raw data to a server with stronger AI processing capabilities for further analysis; the server analyzes the baby's posture data through a deep learning algorithm, detects abnormal postures, and makes early warning judgments, and the analysis results are transmitted to the display terminal via wireless signals.

[0013] The display terminal includes several display devices. When an abnormal posture or potential danger is detected, the system will send an alarm to the caregivers through these display devices to remind them to check the baby's condition in time.

[0014] The workflow of the entire system is to monitor the baby's posture in real time through a camera, use edge computing devices and AI algorithms for data processing and posture recognition, and finally issue an early warning to the caregiver through a display terminal, thereby realizing intelligent detection and real-time early warning of the baby's posture.

[0015] As a further explanation of the above scheme, the processing edge includes a Linux operating system, a data management and storage unit, a system control interface, and a network communication module. The Linux operating system configures related dependency libraries to support the operation of the deep learning model; the Linux operating system deploys the YOLOv8 posture recognition model and configures the model inference script to process the image data transmitted by the camera in real time and recognize the baby's posture; the database or local storage is connected to the Linux operating system to save historical detection data, image data and warning logs to facilitate subsequent statistical analysis and backtracking; the Linux operating system is connected to the system control interface to support system settings, monitoring screen viewing, historical data query and warning information management through web pages and mobile applications; the network communication module is connected to the Linux operating system so that the system has networking capabilities and supports remote monitoring and data transmission; for data privacy and security, encrypted communication protocols (such as SSL / TLS) should be configured to protect the security of data transmission.

[0016] Furthermore, the data collected by the information terminal covers a variety of infant sleeping positions. High-definition cameras or infrared cameras are used to capture images of infant postures during sleep from multiple angles and distances. The system uses data annotation tools to annotate the images to meet the needs of the YOLOv8 posture recognition model. For each image, the infant's various postures need to be accurately annotated, including the position and posture of the head, torso, and limbs. A rectangular frame is used to frame various parts of the infant's body, while ensuring that the bounding box fits the infant's posture closely to avoid including too much background area.

[0017] Furthermore, the process of using the YOLOv8 posture recognition model to identify the baby's sleeping posture includes the following steps: 1) Image input and feature extraction: The input baby sleeping posture image first passes through the Backbone network to extract multi-level and multi-scale feature maps; these feature maps can capture the key details in the image; 2) Feature fusion and enhancement: After feature fusion of the Neck network, feature maps from different scales are fused, which enhances the model's ability to detect objects of different sizes and locations; 3) Bounding box and category prediction: In the Head part, YOLOv8 directly predicts the coordinates and category of the target bounding box on each grid cell; 4) Loss calculation and optimization: The model adopts a comprehensive loss function, including positioning loss, classification loss and confidence loss. Through the back-propagation algorithm, the model continuously adjusts the weights to reduce detection errors.

[0018] Furthermore, in order to achieve the best performance in the task of infant sleeping position recognition, the model needs to go through a fine-tuning strategy, including the following steps: Freeze some layers for fine-tuning: In the early stage of model fine-tuning, freeze the first few layers of Backbone and only train the Neck and Head parts; Gradually unfreeze training: As fine-tuning progresses, each layer of Backbone is gradually unfrozen to perform joint training of the entire model; Category weight adjustment: By adjusting the category weights, the weight ratio in the loss function is balanced to improve the model's ability to identify categories with few samples; Adaptive threshold adjustment: In the inference phase, the confidence threshold and non-maximum suppression (NMS) threshold are adjusted experimentally to improve the accuracy of the model while minimizing false positives. Mixed Precision Training: Mixed Precision Training can not only speed up model training, but also reduce GPU memory usage, thus improving training efficiency without losing accuracy. Cross-validation: During the model fine-tuning process, 5-fold cross-validation is used to evaluate the stability and generalization ability of the model to ensure the consistency of the model on different data sets.

[0019] Furthermore, after the YOLOv8 gesture recognition model is trained and fine-tuned, the model needs to be deployed and optimized: Model quantization: During the deployment phase, model quantization technology is used to reduce the model size and accelerate the inference speed, so that it can run on embedded devices or mobile devices. Reasoning speed optimization: In actual applications, TensorRT or ONNX tools are used to optimize the model to accelerate the reasoning speed of the model and ensure that the system can respond quickly in real-time monitoring; Through the above description, the application of YOLOv8 in infant sleeping posture recognition achieves accurate recognition and monitoring of infant posture through advanced model structure, reasonable training parameter setting and effective fine-tuning strategy, effectively improving the safety and reliability of the system.

[0020] Furthermore, the small computing device with AI function adopts the development board model: NVIDIA Jetson XavierNX, processor: 6-core NVIDIA Carmel ARM v8.2 64-bit CPU, 6MB L2 + 4MB L3, GPU: 384-core NVIDIA Volta GPU with 48 Tensor Cores, memory: 8GB 128-bit LPDDR4x 51.2GB / s, storage: 500GB eMMC 5.1, supports external microSD card.

[0021] Furthermore, the server with AI processing capabilities uses processor: Intel Core i9-13900K 24-Core (8P+16E), GPU: NVIDIA GeForce RTX 4090 24GB GDDR6X, memory: 64GB DDR5 4800MHz, storage: 2TB NVMe SSD, operating system: Ubuntu 20.04 LTS.

[0022] The beneficial effects that can be achieved by the present invention using the above technical solution are:

[0023] The present invention uses the YOLOv8 model to achieve real-time detection and recognition of infant sleeping positions. The high precision and real-time performance of YOLOv8 enable the system to accurately identify infant postures in various environments. On the basis of traditional posture monitoring technology, the recognition ability of the system is improved through deep learning methods, especially when detecting various complex postures of infants, it shows higher accuracy and stability; a real-time early warning mechanism is adopted, through real-time monitoring of the infant's sleeping position, when the system detects a posture that may endanger the safety of the infant, such as the risk of suffocation when lying prone, an early warning notification is immediately issued. Compared with the existing technology, the present invention combines real-time computing and artificial intelligence algorithms, and can respond within milliseconds. , which greatly improves the timeliness and accuracy of early warning; multi-terminal monitoring and control, supports a variety of terminal devices, such as smart phones, tablets, and smart watches, to view the baby's monitoring screen in real time and receive early warning information, providing a flexible monitoring solution; on the basis of traditional video monitoring, through the integration of mobile devices and cloud services, remote real-time monitoring and data synchronization are realized, providing caregivers with a more convenient user experience; environmental adaptability and automatic adjustment, the system can automatically adjust the camera's exposure and contrast according to changes in ambient light, ensuring clear image data under various lighting conditions. In poor ambient light or at night, the image parameters can still be adjusted to maintain high accuracy in posture recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a schematic diagram of the system structure of the present invention.

[0025] Figure 2 It is an operation flow chart of the present invention.

[0026] Figure 3 Schematic diagram of labeling recognition sites for model training samples.

[0027] Figure 4 Schematic diagram of the identification sites for the sleeping posture annotation of the model training samples.

[0028] Figure 5This is the YOLOv8 network flow chart.

[0029] Explanation of the accompanying drawings: 1-information terminal; 2-processing edge; 3-display terminal. DETAILED DESCRIPTION

[0030] In the description of the present invention, it should be noted that directional words, such as the terms "center", "lateral", "longitudinal", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", etc., indicating directions and positional relationships are based on the directions or positional relationships shown in the accompanying drawings, which are only for the convenience of narrating the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and cannot be understood as limiting the specific scope of protection of the present invention.

[0031] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features. Therefore, the definition of "first" and "second" features can explicitly or implicitly include one or more of the features. In the description of the present invention, "at least" means one or more, unless otherwise clearly and specifically defined.

[0032] In the present invention, unless otherwise specified and limited, the terms "assemble", "connect", and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can also be a mechanical connection; it can be directly connected, or it can be connected through an intermediate medium, or the two elements can be internally connected. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0033] In the invention, unless otherwise specified and limited, a first feature being “above” or “below” a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being “above”, “below”, and “above” a second feature includes that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being “above”, “below”, and “below” a second feature includes that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0034] The following is a further description of the specific implementation of the present invention in conjunction with the drawings of the specification, so that the technical solution and its beneficial effects of the present invention are clearer and more explicit. The following description of the embodiments with reference to the drawings is exemplary and intended to explain the present invention, but cannot be understood as limiting the present invention.

[0035] like Figure 1-Figure 5 As shown, the present invention is an intelligent detection and early warning monitoring system for infant posture, which includes an information terminal 1, a processing edge 2, and a display terminal 3. The information terminal 1 includes a camera to monitor the posture and state of the infant in real time; the image and posture information obtained by the camera is transmitted to the processing edge through a wireless signal for analysis and processing; the processing edge includes a small computing device with AI function and a server with AI processing capability; the small computing device performs preliminary data processing, and then sends the processed data or unprocessed raw data to a server with stronger AI processing capability for further analysis; the server analyzes the infant's posture data through a deep learning algorithm, detects abnormal postures, and makes early warning judgments, and the analysis results are transmitted to the display terminal through a wireless signal; the display terminal includes several display devices, and when an abnormal posture or potential danger is detected, the system will send an alarm to the caregivers through these display devices to remind them to check the infant's status in time; the workflow of the entire system is to monitor the infant's posture in real time through a camera, use edge computing devices and AI algorithms for data processing and posture recognition, and finally send an early warning to the caregiver through the display terminal, thereby realizing intelligent detection and real-time early warning of the infant's posture.

[0036] Among them, the small computing device with AI function adopts the development board model: NVIDIA Jetson Xavier NX, processor: 6-core NVIDIA Carmel ARM v8.2 64-bit CPU, 6MB L2 + 4MB L3, GPU: 384-coreNVIDIA Volta GPU with 48 Tensor Cores, memory: 8GB 128-bit LPDDR4x 51.2GB / s, storage: 500GB eMMC 5.1, support external microSD card. Deployment advantages: 1) Low power consumption: The power consumption of NVIDIA Jetson Xavier NX can be adjusted between 10-15W, which is suitable for deployment in energy-sensitive scenarios such as embedded devices, drones, and intelligent monitoring systems. 2) High performance: Although it is a small development board, Xavier NX has powerful computing power, especially its Tensor Cores can greatly accelerate the reasoning speed of the YOLOv8 model. 3) Flexibility: Supports a variety of external camera modules, and high-definition image acquisition can be achieved through the MIPI CSI interface. It can be connected to a Wi-Fi module or Ethernet interface to achieve data transmission and remote monitoring.

[0037] The server with AI processing capabilities uses the processor: Intel Core i9-13900K 24-Core (8P +16E), GPU: NVIDIA GeForce RTX 4090 24GB GDDR6X, memory: 64GB DDR5 4800MHz, storage: 2TBNVMe SSD, operating system: Ubuntu 20.04 LTS. Deployment advantages: 1) Ultra-high performance: Equipped with top-level Intel i9 CPU and NVIDIA RTX 4090 GPU, it can process high-resolution images and large-scale data sets, and support real-time complex tasks such as high-frame rate video processing and real-time multi-target detection. 2) Strong scalability: Storage and memory can be expanded according to needs, and a variety of external devices such as high-definition cameras and external storage devices are supported, which is suitable for scenarios that require long-term and large-scale data processing. 3) Flexible development environment: Supports a variety of development tools and libraries, such as PyTorch, TensorRT, and CUDA, which can efficiently develop, debug, optimize, and deploy models.

[0038] During the data collection process for model training on the information terminal, use a high-definition camera or infrared camera to capture images of the baby's posture during sleep from multiple angles and different distances. In order to ensure data diversity, images can be collected under different environments and lighting conditions. The collected data should cover various common baby sleeping positions, such as supine, side, and prone, as well as some potentially risky positions, such as face-down prone, head covered with quilt, etc. Try to ensure that the number of images for each posture is balanced. Use professional data annotation tools (such as Labelme, VGG Image Annotator, RectLabel, etc.) to annotate the images. The annotation tool should support functions such as rectangular box annotation and key point annotation to meet the needs of the YOLO model. For each image, it is necessary to accurately annotate each posture of the baby, including the position and posture of the head, trunk, and limbs. If it is multi-category detection, it is also necessary to assign corresponding category labels to each annotation box, such as "supine", "side", "prone", etc. In some cases, in addition to annotating the overall posture, key points of the baby (such as ears, eyes, nose, corners of the mouth, neck, shoulders, hips, knees, feet, wrists, palms, etc.) are also annotated to more accurately identify and analyze the baby's posture.

[0039] Bounding box annotation: Use a rectangular box to enclose each part of the baby's body (such as the entire body, face, mouth, etc.), while ensuring that the bounding box fits the baby's posture closely and avoids including too much background area.

[0040] Data augmentation and training division: In order to solve the problem of insufficient data or unbalanced posture categories, data augmentation operations can be performed on the images, such as rotation, flipping, scaling, color adjustment, brightness adjustment, adding noise, etc. These operations can effectively increase the diversity of the data set and enhance the generalization ability of the model. For some postures, different posture deformations can be simulated by fine-tuning the position of the annotation box or adjusting the coordinates of the key points to generate new training samples. Division of training set, validation set and test set: The labeled data set is divided into training set, validation set and test set, usually in a ratio of 8:1:1, to ensure that the types of postures in each data set are as balanced as possible.

[0041] The following is the model training method.

[0042] The YOLOv8 method is used to identify the baby's sleeping position. Figure 5As shown in the figure, it is the YOLOv8 network flow chart. Input module: The input data format is RGB image or video frame, usually with a fixed resolution, such as 640*640 or 1280*1280. The preprocessing operations are normalization, image scaling and data augmentation. Normalization is to normalize the pixel value to [0, 1] or [-1, 1] to adapt to the input requirements of the model. Image scaling is to scale the image to a fixed input size and maintain the aspect ratio (generally using padding edges or cropping). Data augmentation is an enhancement technique such as random rotation, flipping, cropping, color jittering, etc., which improves the robustness of the model to different postures and perspectives. Backbone network: (1) Convolution layer, the basic structure consists of multiple convolution layers, pooling layers, and activation functions (such as ReLU, Leaky ReLU) for feature extraction. The layers are divided into: shallow convolution includes extracting low-level features, such as edges, textures, etc.; deep convolution includes extracting high-level semantic features and identifying complex target structures and key points. (2) Feature pyramid network: This module aims to enhance the fusion of features at different scales, ensuring that both small and large objects can be effectively detected, especially for gesture recognition, ensuring that key points can be accurately located. The structure is to upsample and merge feature maps of different depths layer by layer to generate multi-scale feature maps from small to large. The design of the present invention includes: 1) Bottom-up path: extract high-level features layer by layer. 2) Top-down path: pass high-resolution information to low-resolution feature maps layer by layer. (3) Path aggregation network: The function of this module is to further enhance the global information transmission in the feature map, especially the flow of detail information. The path aggregation network allows high-level semantic information and low-level feature information to be more deeply integrated, which is very important for fine-grained gesture recognition. Head network: The head network of YOLOv8 is divided into two important branches, which are used for target detection and gesture recognition respectively. (1) Target detection branch: 1) Bounding box prediction: For each candidate region, YOLOv8 predicts the coordinates (center point, width and height) and confidence of the bounding box. 2) Category prediction: The category probability distribution of objects in each bounding box. The YOLOv8 model can predict a variety of different types of targets, such as humans, animals, etc. (2) Posture recognition branch, 1) Key point regression: predict the 2D coordinates of the joints for each target area, usually the key points of the human body (for example: head, shoulder, elbow, wrist, hip, knee, ankle, etc.). 2) Loss function, generally using a key point regression loss function, such as Mean Squared Error (MSE) or L1 Loss, to regress the coordinate position of each key point. 3) Human skeleton construction, connect the detected key points to construct a human skeleton diagram. In this step, the model outputs a visualization of the posture skeleton by connecting specific key points. Post-processing module: (1) Non-maximum suppression: filter overlapping bounding boxes, retain the bounding box with the highest confidence, ensure that each human body has only one detection box, and avoid multiple repeated bounding boxes.(2) Pose decoding: Decode the predicted joint information into 2D coordinates and connect them into a pose skeleton through certain rules or structures. (3) Skeleton connection rules: Predefined human skeleton topology, such as the head-shoulder-elbow-hand connection, is used to connect the predicted key points into a visual pose. Loss function: (1) Total loss function: The training goal of YOLOv8 is to optimize detection and pose recognition at the same time. The total loss function usually contains three parts; (2) Bounding box loss: Calculate the difference between the predicted bounding box and the true box, using IoU or CIoU loss; (3) Classification loss: Used to classify the object type in the bounding box, usually using cross entropy loss; (4) Key point loss: Used to regress the key point coordinates of the human pose, usually using MSE loss. Output module: (1) Target bounding box: Output the bounding box information of each detected target, including position, width, height, category and confidence. (2) Pose estimation results: 2D key point coordinates: Output the key point coordinates of each human pose and generate a pose skeleton graph. (3) Visualization results: Overlay the bounding box and posture skeleton on the original image or video frame. Visualization and output: (1) Visualization skeleton graph: Generate a visualization image with relevant nodes and skeleton connections, and overlay the key points and skeleton on the original image. (2) Output format: The results of posture estimation can be output, such as joint point coordinates in JSON format, or pictures or videos with posture tags.

[0043] Model training includes model structure, model principle, training parameter setting and model fine-tuning strategy.

[0044] (1) Model structure The main features of YOLOv8 include: Backbone: YOLOv8 uses the CSPNet (Cross Stage Partial Network) structure as the backbone network. CSPNet reduces the amount of calculation and retains more feature expression capabilities by sharing some gradient flows. Neck: In YOLOv8, the Neck part still uses the PANet (Path Aggregation Network) structure to further enhance the feature fusion capability and improve the detection capability of targets of different scales. PANet improves the effect of detecting small targets by fusing feature maps of different scales. Head: YOLOv8 uses an Anchor-free design in the Head part to directly predict the bounding box coordinates and category of the target, simplifying model training and reducing the complexity of Anchor settings. This can better adapt to various posture changes in infant posture detection. Prediction layer: The prediction layer of YOLOv8 outputs the category probability, bounding box position and size, and confidence of each grid unit. Due to the use of a more efficient network structure, YOLOv8 can further reduce the inference time while maintaining detection accuracy.

[0045] (2) Model principles The model principle of YOLOv8, the main detection process includes the following steps: 1) Image input and feature extraction: The input baby sleeping posture image first passes through the Backbone network to extract multi-level and multi-scale feature maps; these feature maps can capture the key details in the image. 2) Feature fusion and enhancement: After the feature fusion of the Neck network, the feature maps from different scales are fused to enhance the model's detection ability for targets of different sizes and positions. This step is particularly important because babies' sleeping postures are diverse and change rapidly, and the fused feature maps can help the model better capture these postures. 3) Bounding box and category prediction: In the Head part, YOLOv8 directly predicts the coordinates and category of the target bounding box on each grid unit; this anchor-free method avoids the complex setting of anchors and improves the detection speed and flexibility of the model. 4) Loss calculation and optimization: The model uses a comprehensive loss function, including positioning loss, classification loss and confidence loss. Through the back propagation algorithm, the model continuously adjusts the weights to reduce detection errors.

[0046] (3) Training parameter setting When using YOLOv8 to train the baby sleeping position recognition model, the following training parameters are crucial: 1) Learning Rate: The initial learning rate is set to 0.001, and the cosine annealing learning rate scheduler is used. The learning rate is gradually reduced during the training process to avoid overfitting of the model and improve the convergence speed. 2) Batch Size: According to the hardware conditions and the size of the dataset, the batch size is set to 32. The batch size should be selected as large as possible without overflowing the GPU memory to improve the training efficiency. 3) Optimizer: The AdamW optimizer is used, which combines the fast convergence of Adam and the robustness of L2 regularization, and is suitable for large-scale model training such as YOLOv8. 4) Number of training epochs: The number of training epochs is set to 500, and the specific number of epochs is adjusted according to the performance of the model on the validation set; in the early stage, a higher learning rate can be used for fast training, and then the learning rate can be gradually reduced to fine-tune the model. 5) Loss Function: The default comprehensive loss function in YOLOv8 is used, including CIoU loss (Complete IoU), Focal Loss (used to deal with category imbalance) and confidence loss. 6) Data Augmentation: During the training process, data augmentation techniques such as random cropping, horizontal flipping, color adjustment, and noise addition are used to improve the generalization ability of the model, especially when dealing with different postures of infants and environmental changes.

[0047] (4) Model fine-tuning strategy In order to achieve the best performance in the task of infant sleeping posture recognition, the model fine-tuning strategy is a key step: 1) Freeze some layers for fine-tuning: In the early stage of model fine-tuning, you can freeze the first few layers of Backbone and only train the Neck and Head parts. This method can focus on fine-tuning the detection and classification capabilities of the model while retaining the feature extraction capability. 2) Gradually unfreeze the training: As the fine-tuning progresses, you can gradually unfreeze the layers of Backbone and conduct joint training of the entire model. This step can further optimize the overall performance of the model, especially improve the detection capability of subtle postures. 3) Class weight adjustment: Since there may be class imbalance in infant postures (such as less data for some postures), you can adjust the class weights to balance the weight ratio in the loss function and improve the model's recognition capability for few-sample categories. 4) Adaptive threshold adjustment: In the inference stage, the confidence threshold and non-maximum suppression (NMS) threshold are adjusted experimentally to minimize false positives while improving the accuracy of the model. 5) Mixed precision training: Using mixed precision training can not only speed up the training of the model, but also reduce the GPU memory usage, and improve the training efficiency without losing accuracy. 6) Cross-validation: During the model fine-tuning process, 5-fold cross-validation is used to evaluate the stability and generalization ability of the model to ensure the consistency of the model on different data sets.

[0048] (5) Model deployment and optimization After training and fine-tuning, the model needs to be deployed and optimized: 1) Model quantization: In the deployment phase, model quantization technology (such as INT8 quantization) can be used to reduce the model size and accelerate the inference speed, so as to adapt to running on embedded devices or mobile devices. 2) Inference speed optimization: In practical applications, tools such as TensorRT and ONNX are used to optimize the model, accelerate the inference speed of the model, and ensure that the system can respond quickly in real-time monitoring. 3) Through the above description, the application of YOLOv8 in infant sleeping posture recognition has achieved accurate recognition and monitoring of infant posture through advanced model structure, reasonable training parameter settings and effective fine-tuning strategies, effectively improving the safety and reliability of the system.

[0049] Compared with the prior art, the present invention 1) adopts intelligent posture detection technology and uses the YOLOv8 model to achieve real-time detection and recognition of infant sleeping postures. The selection and application of the model are the core of the present invention. The high precision and real-time performance of YOLOv8 enable the system to accurately identify infant postures in various environments. On the basis of traditional posture monitoring technology, the recognition ability of the system is improved through deep learning methods, especially when detecting various complex postures of infants, showing higher accuracy and stability.

[0050] 2) Real-time warning mechanism: through real-time monitoring of the baby's sleeping position, when the system detects a posture that may endanger the baby's safety (such as the risk of suffocation when lying on the stomach), it will immediately issue a warning notification; compared with the existing technology, the present invention combines real-time computing and artificial intelligence algorithms, and can respond within milliseconds, greatly improving the timeliness and accuracy of the warning.

[0051] 3) Multi-terminal monitoring and control, supporting multiple terminal devices (such as smart phones, tablets, smart watches) to view the baby's monitoring screen in real time and receive warning information, providing a flexible monitoring solution; based on traditional video monitoring, through the integration of mobile devices and cloud services, remote real-time monitoring and data synchronization are realized, providing caregivers with a more convenient user experience.

[0052] 4) Environmental adaptability and automatic adjustment. The system can automatically adjust the exposure and contrast of the camera according to changes in ambient light to ensure that clear image data can be obtained under various lighting conditions. The present invention can still maintain high accuracy of gesture recognition by adjusting image parameters under poor ambient light or night conditions.

[0053] The above is only a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. An intelligent baby posture detection and early warning monitoring system, characterized in that: It includes information terminal, processing edge, and display terminal. The information terminal includes a camera to monitor the baby's posture and status in real time; the image and posture information obtained by the camera is transmitted to the processing edge via wireless signals for analysis and processing; The processing edge includes a small computing device with AI functions and a server with AI processing capabilities. The small computing device performs preliminary data processing and then sends the processed data or unprocessed raw data to a server with stronger AI processing capabilities for further analysis. The server analyzes the baby's posture data through a deep learning algorithm, detects abnormal postures, and makes early warning judgments. The analysis results are transmitted to the display terminal via wireless signals. The display terminal includes several display devices. When abnormal posture or potential danger is detected, the system will send an alarm to the caregivers through these display devices to remind them to check the baby's condition in time; The workflow of the entire system is to monitor the baby's posture in real time through a camera, use edge computing devices and AI algorithms for data processing and posture recognition, and finally issue an early warning to the caregiver through a display terminal, thereby realizing intelligent detection and real-time early warning of the baby's posture.

2. The intelligent detection and early warning monitoring system for infant posture according to claim 1, characterized in that: The processing edge includes a Linux operating system, a data management and storage unit, a system control interface, and a network communication module. The Linux operating system configures related dependency libraries to support the operation of the deep learning model; the Linux operating system deploys a YOLOv8 posture recognition model and configures a model reasoning script to process image data transmitted in real time by the camera and recognize the posture of the baby; The database or local storage is connected to the Linux operating system to save historical detection data, image data and warning logs; the Linux operating system is connected to the system control interface, which supports system settings, monitoring screen viewing, historical data query and warning information management through web pages and mobile applications; the network communication module is connected to the Linux operating system, so that the system has networking capabilities and supports remote monitoring and data transmission; for data privacy and security, an encrypted communication protocol should be configured to protect the security of data transmission.

3. The baby posture intelligent detection and early warning monitoring system according to claim 1, characterized in that: The data collected by the information terminal covers a variety of infant sleeping positions. High-definition cameras or infrared cameras are used to capture images of infants' postures during sleep from multiple angles and distances. The system uses data annotation tools to annotate the images. For each image, the infant's various postures need to be accurately annotated, including the position and posture of the head, torso, and limbs. A rectangular frame is used to frame various parts of the infant's body, while ensuring that the bounding box fits the infant's posture closely to avoid including too much background area.

4. The intelligent detection and early warning monitoring system for infant posture according to claim 2, characterized in that: The process of using the YOLOv8 posture recognition model to identify the baby's sleeping posture includes the following steps: 1) Image input and feature extraction: The input baby sleeping posture image first passes through the Backbone network to extract multi-level and multi-scale feature maps; 2) Feature fusion and enhancement: After feature fusion of the Neck network, feature maps from different scales are fused; 3) Bounding box and category prediction: In the Head part, YOLOv8 directly predicts the coordinates and category of the target bounding box on each grid cell; 4) Loss calculation and optimization: The model adopts a comprehensive loss function, including positioning loss, classification loss and confidence loss. Through the back-propagation algorithm, the model continuously adjusts the weights.

5. The intelligent detection and early warning monitoring system for infant posture according to claim 3, characterized in that: The model needs to go through a fine-tuning strategy, which includes the following steps: 1) Freeze some layers for fine-tuning: In the early stage of model fine-tuning, freeze the first few layers of Backbone and only train the Neck and Head parts; 2) Gradually unfreeze training: As fine-tuning progresses, each layer of Backbone is gradually unfrozen to perform joint training of the entire model; 3) Category weight adjustment: Balance the weight ratio in the loss function by adjusting the category weight; 4) Adaptive threshold adjustment: In the inference stage, the detection threshold and non-maximum suppression threshold are adjusted through experiments; 5) Mixed precision training: Use mixed precision training; 6) Cross-validation: During the model fine-tuning process, 5-fold cross-validation is used to evaluate the stability and generalization ability of the model to ensure the consistency of the model on different data sets.

6. The baby posture intelligent detection and early warning monitoring system according to claim 4, characterized in that: After the YOLOv8 gesture recognition model is trained and fine-tuned, the model needs to be deployed and optimized: 1) Model quantization: In the deployment phase, model quantization technology is used to reduce the model size and accelerate the inference speed, so that it can run on embedded devices or mobile devices; 2) Reasoning speed optimization: In actual applications, TensorRT or ONNX tools are used to optimize the model to accelerate the reasoning speed of the model and ensure that the system can respond quickly in real-time monitoring.

7. The baby posture intelligent detection and early warning monitoring system according to claim 1, characterized in that: The small computing device with AI function uses the development board model: NVIDIA Jetson Xavier NX, processor: 6-core NVIDIACarmel ARM v8.2 64-bit CPU, 6MB L2 + 4MB L3, GPU: 384-core NVIDIA Volta GPUwith 48 Tensor Cores, memory: 8GB 128-bit LPDDR4x 51.2GB / s, storage: 500GB eMMC 5.1, supports external microSD card.

8. The baby posture intelligent detection and early warning monitoring system according to claim 1, characterized in that: The server with AI processing capabilities uses processor: Intel Core i9-13900K 24-Core (8P + 16E), GPU: NVIDIAGeForce RTX 4090 24GB GDDR6X, memory: 64GB DDR5 4800MHz, storage: 2TB NVMe SSD, operating system: Ubuntu 20.04 LTS.

Images