Wearable hot card prompting bracelet

Abstract

The invention is applicable to the technical field of wearable equipment, and provides a wearable hot card prompt bracelet, which comprises a bracelet main body, a front watch and a rear watch which are arranged on the front side and the rear side of the bracelet main body respectively, and each of the front watch and the rear watch comprises a camera, a motion sensor, a microphone, a processor, a vibration module, a display screen, a target hot card setting module and a diet scheme generation module. Multi-view images of food are collected through the front meter and the rear meter, the AI recognition and binocular distance measurement technology is combined, the heat (error is within + / -8%) of the food clamped at a time is calculated, the bracelet displays a numerical value in real time, and high-precision data are provided for card control people. Heat is fed back through vibration and screen numerical values within 0.3 s at the moment of clamping food, accumulated heat early warning is performed during dining, and the nutrition proportion of a color ring on a dial plate is displayed, so that a user is helped to master diet. A daily calorie target value can be self-defined according to the physical condition of the user, a personalized scheme is given through AI analysis of the whole meal, the system intelligently reminds to adjust the target value, precise management is achieved, and the health level is improved.

Description

Technical Field

[0001] This invention belongs to the field of wearable device technology, and particularly relates to a wearable heat card reminder wristband. Background Technology

[0002] As people's living standards improve, dietary health is receiving increasing attention. For patients with chronic diseases such as diabetes and obesity, as well as fitness enthusiasts, accurately controlling daily calorie intake is a crucial aspect of health management. Traditional methods of calculating food calories mainly rely on manually querying food calorie databases and making estimations. This process is cumbersome, prone to errors, and difficult to operate in real time, failing to meet people's need for quick and accurate access to food calorie information during meals. Especially when dining with multiple people, it is difficult for users to accurately determine the portion size and corresponding calories of the food they take, further increasing the difficulty of calorie control.

[0003] Although wearable devices such as health monitoring bracelets have appeared on the market, they have significant defects in food calorie management: (1) Inefficient calculation method and large error: The calorie calculation of existing devices still relies on manual query and estimation, which is cumbersome and time-consuming. It is difficult to apply in the fast-paced dining scene, and the error caused by human operation is difficult to avoid; (2) Lack of real-time recognition function: It cannot accurately identify the type of food when picking up food, nor can it calculate the accurate calories of the food picked up and provide real-time feedback, which cannot meet the user's need for real-time monitoring of calorie intake during the dining process; (3) Insufficient personalized control capability: It cannot set calorie intake targets according to the user's personalized needs, nor can it provide effective reminders when the calorie intake exceeds the standard, which is difficult to adapt to the differentiated control needs of different groups of people.

[0004] Therefore, this invention proposes a wearable heat card reminder wristband. Summary of the Invention

[0005] The purpose of this invention is to provide a wearable heat card reminder wristband, which aims to solve the problems mentioned in the background art.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] A wearable heat card reminder wristband includes a wristband body, with a front watch and a rear watch respectively disposed on the front and back sides of the wristband body. The front watch and the rear watch have the same structure and both include:

[0008] The camera, used to capture images of food, consists of dual front-facing cameras and dual rear-facing cameras;

[0009] Motion sensors are used to collect data on the movement trajectory of tableware;

[0010] Microphones are used to collect collision soundprint data;

[0011] The processor is used to run AI recognition and algorithms;

[0012] Vibration module for heat indication and early warning vibration feedback;

[0013] The display screen shows the dial color rings and calorie information;

[0014] The target calorie setting module is used to customize the daily calorie intake target value;

[0015] The diet plan generation module is used to generate reasonable diet plans;

[0016] The bracelet works in the following four stages:

[0017] I. Target Calorie Setting and Personalized Management Phase: Users set their daily calorie intake target value through the target calorie setting module, and the wristband intelligently adjusts the target value based on historical data and provides reminders;

[0018] II. Data Acquisition Stage: Motion sensors collect data on the movement trajectory of the tableware; cameras on the front and rear tables capture multi-view images the moment food is picked up; and microphones collect collision soundprint data.

[0019] III. Data Processing and Analysis Stage: After the processor preprocesses the image, it identifies the food type through an AI model, calculates the food volume using binocular ranging technology, and then calculates the calorie value based on the food calorie database. At the same time, the processor extracts features from the tableware movement trajectory data, visual data, and voiceprint data, and optimizes the calorie calculation, food type identification, and calorie calculation accuracy through a three-modal data fusion algorithm. Finally, the diet plan generation module generates personalized diet suggestions.

[0020] IV. Feedback and Execution Phase: The vibration module triggers vibration feedback based on the calorie value, and the display screen shows the calorie value and nutrient ratio, as well as the dietary plan.

[0021] Furthermore, when the processor preprocesses the image, it first uses the Laplacian variance algorithm to evaluate the image sharpness. When the variance of the Laplacian operator is less than 100, it filters out blurry data. Then, it performs image enhancement and normalization operations. Image enhancement includes random rotation and random scaling.

[0022] The formula for the Laplace variance algorithm is:

[0023] ;

[0024] in: The variance of the Laplace operator; The Laplacian operator for image I; via Calculate the variance of the Laplacian operator, where x and y are pixel coordinate values;

[0025] The formula for random rotation is:

[0026] ;

[0027] in: This represents the pixel value of the rotated image at coordinates (x, y). Pixel value; Let θ be the rotation matrix, and θ be a random rotation angle. It is the inverse of the rotation matrix; It is in vector form with coordinates (x, y);

[0028] The formula for random scaling is:

[0029] Isc(x,y)=I(s / x,s / y);

[0030] Where: Isc(x,y) are the scaled pixel values; s is the scaling factor, when s>1 the image is shrunk, and when s<1 the image is enlarged;

[0031] The normalization formula is:

[0032] ;

[0033] Where: μ is the pixel mean of the training set images; σ is the pixel standard deviation of the training set images; This represents the pixel value of the original image at coordinates (x, y). This represents the pixel value of the normalized image at coordinates (x, y).

[0034] Furthermore, the AI ​​model is based on a deep learning algorithm and is trained using image data of more than 3,000 kinds of food ingredients. During training, a Mixup data augmentation and category balancing sampling strategy is adopted.

[0035] The Mixup data augmentation formula is:

[0036] ;

[0037] ;

[0038] in: The enhanced pixel values; Enhanced food category labels; These are two randomly selected food images; For the corresponding food category label; It is the mixing coefficient that follows a Beta distribution;

[0039] The category-balanced sampling formula is:

[0040] ;

[0041] in: is the weight of the i-th food sample; N is the total number of samples; Let C be the number of samples in the i-th category; C is the total number of food categories.

[0042] Furthermore, the three-modal data fusion algorithm is implemented in the following way:

[0043] Feature-level fusion employs a weighted attention mechanism, the formula of which is:

[0044] F fusion =α rgb ⊙F rgb +α depth ⊙F depth +α thermal ⊙F thermal ;

[0045] Wherein: F fusion The fused feature vector; α rgb , α depth , α thermal These represent the attention weights for RGB, depth, and infrared modes, respectively; ⊙ represents element-wise multiplication; F rgb , F depth , F thermal These are the extracted RGB feature vector, depth feature vector, and infrared feature vector, respectively.

[0046] The decision-level fusion adopts Bayesian rules, and its formula is as follows:

[0047] P(Y∣X rgb X depth X thermal )∝P(X rgb |Y), P(X) depth |Y), P(X) thermal |Y)P(Y);

[0048] Where: P(Y|X) rgb ,X depth ,X thermal X represents the posterior probability of food category Y given three modal observations; Y is the food category; X represents the food category. rgb X depth X thermal These are the observations for RGB, depth, and infrared modes, respectively; P(X) rgb |Y), P(X) depth |Y), P(X) thermal|Y) is the likelihood function; P(Y) is the prior probability;

[0049] For dynamic food recognition, a temporal attention mechanism is used to process dynamic features, and its formula is as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] Where: h t For hidden state vectors; GRU is a gated recurrent unit; F fusion t Let be the fused feature vector at time t; The hidden state vector before the update; Let be the attention weight at time t; This is the weight parameter matrix that needs to be learned during model training; The bias term parameters are used for model training; i is the number of elements; T is the length of the time sequence. The fusion features are the final output.

[0054] Uncertainty-aware fusion is achieved through adaptive weight adjustment, and its formula is as follows:

[0055] (i∈{rgb,depth,thermal});

[0056] in: The weight values ​​are adaptively adjusted for perception fusion; , i, j are the variances of the i-th and j-th elements respectively; j is the index of the summation; i is the number of elements; rgb, depth, thermal are the data categories of RGB, depth and infrared modes respectively.

[0057] Furthermore, the vibration module converts real-time heat values ​​into vibration intensity, achieving mapping through a piecewise linear function. The basic piecewise linear mapping formula is:

[0058] ;

[0059] Where: VL is the vibration intensity level; Q is the calorie value of a single food pick-up; The minimum perceptible heat threshold; The step size for heat fractionation; This represents the maximum linearly mapped heat value. This represents the highest level of vibration intensity.

[0060] The formulas for how vibration amplitude and frequency change with vibration intensity level are:

[0061] ;

[0062] ;

[0063] Where: A is the vibration amplitude; f is the vibration frequency; A0 is the fundamental amplitude; f0 is the fundamental frequency.

[0064] Furthermore, the display screen dynamically displays a color ring, which is divided into different color blocks according to the proportion of carbohydrate, protein, and fat nutrients, to show the nutritional distribution of the food being picked up in real time.

[0065] Compared with the prior art, the beneficial effects of the present invention are:

[0066] 1. Precise Calorie Management: The system captures multi-view food images using dual front and dual rear cameras. Combined with AI recognition and binocular ranging technology, it accurately identifies food types and measures volume, simultaneously querying a food calorie database to calculate the calories per scoop, with an error margin within ±8%. The wristband displays real-time calorie values, providing high-precision data for individuals with diabetes, obesity, and other conditions requiring strict calorie control. This effectively assists in optimizing their diet and significantly improves the accuracy and reliability of calorie calculation compared to traditional estimation methods.

[0067] 2. Real-time and Convenient Feedback: A three-modal data fusion algorithm is triggered the instant food is picked up, providing calorie feedback within 0.3 seconds through vibration intensity and screen readings. This feedback continues throughout the meal, utilizing a cumulative calorie alert function and a color-coded nutrient ratio display on the dial to help users fully monitor their diet. This rapid, multi-dimensional feedback allows users to perceive calories instantly through touch and sight, enabling timely adjustments to their eating habits. Highly aligned with real-world dining scenarios, this significantly enhances the user experience and adherence, contributing to the development of healthy eating habits.

[0068] 3. Personalized Dietary Assistance: Optimized based on clinically validated data, the system allows users to customize their daily calorie intake target based on their physical condition, health goals, and activity level. The wristband provides dynamic alerts and generates targeted suggestions. By taking photos of complete meals, AI analyzes the nutritional components and provides personalized solutions. The system can also intelligently remind users to adjust target values ​​based on historical data. Combined with the color-coded nutrition visualization function on the watch face, it achieves precise management across the entire chain from target setting to dietary intervention, meeting the personalized dietary management needs of different users and possessing clinical application value. Attached Figure Description

[0069] Figure 1 A 3D model of a wearable thermal card reminder wristband after removing the display screen.

[0070] Figure 2 A 3D diagram of the vibration module in a wearable thermal card alert wristband.

[0071] Figure 3 A flowchart illustrating the working method of a wearable thermal card reminder wristband.

[0072] In the diagram: 1-Bracelet body; 2-Dual front cameras; 3-Dual rear cameras; 4-Motion sensor; 5-Microphone; 6-Processor; 7-Vibration module; 8-Target calorie setting module; 9-Diet plan generation module; 10-Front-side display; 11-Rear-side display. Detailed Implementation

[0073] In order to provide a clearer understanding of the technical features, objectives and beneficial effects of the present invention, the technical solution of the present invention will now be described in detail below, but it should not be construed as limiting the scope of implementation of the present invention.

[0074] like Figure 1 and Figure 2 As shown, this invention provides a wearable calorie reminder wristband, including a wristband body 1. A front watch 11 and a rear watch 12 are respectively set on the front and back sides of the wristband body 1. The front watch 11 and the rear watch 12 have the same structure, both including a front dual camera 2 and a rear dual camera 3 (collectively referred to as cameras, used to capture multi-view food images). It is also equipped with a motion sensor 4 (for collecting tableware movement trajectory data), a microphone 5 (for collecting collision soundprint data), a processor 6 (for running AI recognition and algorithms), a vibration module 7 (for calorie reminder and warning vibration feedback), a display screen (for displaying information such as the dial color ring and calorie value), a target calorie setting module 8 (for customizing the daily calorie intake target value), and a diet plan generation module 9 (for generating a reasonable diet plan based on a food photo and related data of a meal). All components work together through circuitry to realize the functions of calorie reminder, control, and diet advice.

[0075] During the meal, motion sensor 4 collects data on the movement trajectory of the tableware and triggers the cameras on the front and rear cameras 11 and 12 to capture multi-view images of the food at the moment it is picked up, acquiring visual data. Simultaneously, microphone 5 collects collision sound data. Processor 6 preprocesses the images, identifies the food type using a pre-trained AI model, calculates the food volume using binocular ranging technology combined with data from motion sensor 4, and then calculates the calorie value for a single pick-up based on a food calorie database. Simultaneously, processor 6 extracts features from the tableware movement trajectory data, visual data, and sound data, optimizing the accuracy of food-picking behavior judgment and calorie calculation through a three-modal data fusion algorithm. Furthermore, users can manually take photos of their entire meal. The diet plan generation module 9 identifies the food type, analyzes nutritional components, and generates personalized dietary suggestions based on the user's set health goals, dietary preferences, and physical condition, displayed on the screen in graphic and textual form. Processor 6 converts real-time calorie values ​​into vibration intensity and screen display values, and dynamically issues warnings based on the difference between the cumulative calorie intake and the target value, continuously adjusting suggestions during the meal to help users achieve scientific dietary management.

[0076] This invention provides a method for operating the above-mentioned wearable heat card reminder wristband, the process of which is as follows: Figure 3 As shown, it includes the following four stages:

[0077] I. Target Calorie Setting and Personalized Management Phase;

[0078] Personalized goal setting: Users can set their daily calorie intake target value through the target calorie setting module 8 of the bracelet, based on their own physical condition, health goals (such as weight loss, blood sugar control), daily activity level, etc. For example, users who want to lose weight can set it to 1500kcal, and diabetic patients can set it to 1800kcal according to their doctor's advice.

[0079] Intelligent Adjustment and Reminders: The band can provide intelligent suggestions based on user habits and historical data. When a user's actual calorie intake deviates significantly from the target value for several consecutive days, it will remind the user to reassess and adjust the target value to ensure that the target is reasonable and achievable. At the same time, it will regularly remind the user to pay attention to calorie intake, increasing the user's awareness and motivation for dietary management.

[0080] II. Data Collection Phase;

[0081] Multimodal data acquisition: During dining, motion sensor 4 collects data on the movement trajectory of the tableware, including information such as acceleration, angular velocity changes, movement speed, and trajectory shape. Simultaneously, the cameras on the front and rear displays 11 are triggered the instant food is picked up, capturing multi-view images of the food and acquiring visual data such as its shape, color, and texture. Microphone 5 collects collision soundprint data (soundprint data generated by collisions between tableware and food, and between tableware themselves, during the food-picking process), serving as a data basis to assist in judging the food-picking behavior and determining the food's state (such as the softness or hardness of the food, which may affect volume judgment).

[0082] Manual food collection for the whole meal: Before or during a meal, users can manually trigger the wristband's shooting function to take photos of all the food in a meal and record the food information for overall analysis of the diet plan.

[0083] III. Data Processing and Analysis Stage;

[0084] 1. AI recognition and calorie calculation:

[0085] 1.1 Image Preprocessing: For the images captured by the cameras in the front table 11 and the rear table 12, firstly, to ensure the quality of visual data, the Laplacian variance algorithm is used to evaluate the image sharpness, thereby filtering out blurry data; then, image enhancement is performed to improve sharpness and contrast, making the details of the food more obvious; then, normalization is performed to unify the image data to a specific range, which is convenient for subsequent model processing and eliminates the influence of different images caused by differences in shooting equipment and lighting. For example, food images under different lighting conditions (such as indoor warm light and outdoor strong light) can be mapped to a uniform distribution to improve the robustness of the model to changes in lighting.

[0086] (1) The formula for the Laplace variance algorithm is:

[0087] ;

[0088] in: This represents the variance of the Laplacian operator; the smaller the value, the blurrier the image. Let be the Laplacian operator for image I, reflecting the second derivative of the image; through Calculate the variance of the Laplacian operator, where x and y are pixel coordinates. When a detection is made... When the value is less than 100, the system automatically filters out the blurry image, thus ensuring that the food images used for training are clearly identifiable.

[0089] (2) Image enhancement includes random rotation and random scaling to enhance the model's ability to generalize to food of different angles and sizes. For example, the same plate of stir-fried dish may have different angles and distances when taken by different users.

[0090] The formula for random rotation is:

[0091] ;

[0092] in: This represents the pixel value of the rotated image at coordinates (x, y). Pixel value; θ is a rotation matrix, where θ is a random rotation angle (e.g., from −15° to 15°). This is the inverse of the rotation matrix, used for coordinate transformations; It is a vector form of coordinates (x, y).

[0093] The formula for random scaling is:

[0094] Isc(x,y)=I(s / x,s / y);

[0095] Where: Isc(x,y) is the scaled pixel value; s is the scaling factor (such as a random value between 0.8 and 1.2), when s>1 the image is shrunk, and when s<1 the image is enlarged.

[0096] (3) The normalization formula is:

[0097] ;

[0098] Where: μ is the pixel mean of the training set images; σ is the pixel standard deviation of the training set images; This represents the pixel value of the original image at coordinates (x, y). This represents the pixel value of the normalized image at coordinates (x, y).

[0099] 1.2 Food Classification and Recognition: Preprocessed images are input into a pre-trained AI model. This model, based on deep learning algorithms, is trained using image data of over 3000 food ingredients (covering different angles, lighting, cooking methods, and cut shapes). It boasts high recognition accuracy and can be continuously optimized through online learning to adapt to more food types and complex food presentation formats (such as different cooking methods and cut shapes). The AI ​​model extracts features from the food in the images and accurately classifies food types (such as braised pork, broccoli, etc.). The recognition results serve as a key basis for subsequent calorie calculations.

[0100] When training the AI ​​model, we used image data of over 3000 kinds of food ingredients, combined with Mixup data augmentation and category-balanced sampling strategies:

[0101] The Mixup data augmentation formula is:

[0102] ;

[0103] ;

[0104] in: The enhanced pixel values; Enhanced food category labels; These are two randomly selected food images; For the corresponding food category label (one-hot encoding); It is the mixing coefficient that follows a Beta distribution. Mixup can generate virtual food samples (such as "half a plate of fried rice + half a plate of salad"), enhancing the model's ability to recognize food combinations, especially suitable for recognizing mixed dishes.

[0105] The category-balanced sampling formula is:

[0106] ;

[0107] in: is the weight of the i-th food sample; N is the total number of samples; Let C be the number of samples in the i-th class; C is the total number of food categories. This method can solve the problem of food category imbalance in the dataset (e.g., there are many samples of "apple" and few samples of "truffle"), and improve the recognition accuracy of rare food categories through weighted sampling.

[0108] 1.3 Volume Recognition: Utilizing binocular ranging technology, the spatial three-dimensional coordinates of the food are calculated based on the parallax of images captured by the front and rear cameras (11 and 12), thereby obtaining the food volume. Simultaneously, by combining the motion trajectory data of the tableware collected by the motion sensor (4), the food's range and volume calculation boundaries are precisely defined, reducing measurement errors and providing reliable volume data for calorie calculation.

[0109] 1.4 Calorie Calculation: Based on the food type identified by the AI ​​model and the food volume measured by binocular ranging technology, the built-in food calorie database (containing nutritional information such as calorie values ​​per unit mass or volume of various ingredients) is queried to calculate the calorie value of the food being picked up, and the calorie calculation error is controlled within ±8%.

[0110] 2. Trimodal data fusion optimization:

[0111] 2.1 Feature Extraction: Features such as speed change and direction change are extracted from the tableware motion trajectory data; features such as frequency distribution and amplitude are extracted from the collision soundprint data; features such as food outline, color ratio and texture details are extracted from the visual data to prepare for subsequent fusion processing.

[0112] 2.2 Feature Fusion: A trimodal data fusion algorithm is used to comprehensively analyze and correlate multimodal features (such as the correlation between motion trajectory features and changes in food position in visual data, and the correlation between voiceprint features and food texture (hardness / softness)) to make the data complementary and corroborative. The fused data not only effectively improves the accuracy of judging food-grabbing behavior and accurately distinguishes between normal food grabbing and misoperation, but also enhances the reliability of food recognition and calorie calculation, more accurately determining the type of food through multimodal data. This enables the wristband to accurately complete calorie assessment tasks in complex and changing dining scenarios, reducing misjudgments and errors caused by single data modalities. The fusion analysis results are output to realize functions such as calorie calculation and food grabbing prompts, improving the accuracy of calorie assessment by the wristband in complex dining scenarios.

[0113] a. Trimodal feature extraction and fusion;

[0114] (1) Feature extraction formula:

[0115] RGB modal feature extraction: F rgb = (I rgb ); where: I rgb It is an RGB image; For RGB-specific convolutional neural networks (such as ResNet-50); F rgb ∈R d rgb The extracted RGB feature vector has a dimension of d. rgb .

[0116] Deep modal feature extraction: F depth = (I depth ); where: I depth This is a depth image (reflecting the distance of objects from the camera); For deep-specific convolutional neural networks (such as Depthwise Separable CNN); F depth ∈R d depth The extracted deep feature vector has dimension d. depth .

[0117] Infrared thermal imaging feature extraction: F thermal = (I thermal ); where: Ithermal is an infrared thermal imaging image (reflecting the temperature distribution on the surface of an object); For infrared-specific convolutional neural networks (such as an improved version of the Inception architecture); F thermal ∈R d thermal The extracted infrared feature vector has a dimension of d.thermal .

[0118] RGB images provide information about the color and texture of food (e.g., red may indicate meat or tomatoes), depth images supplement the three-dimensional structure of food (e.g., the fluffiness of bread), and infrared thermal imaging reflects the temperature of food (e.g., the difference between hot soup and cold drinks).

[0119] (2) Feature-level fusion formula:

[0120] Weighted attention fusion mechanism:

[0121] α rgb =σ(W rgb F rgb +b rgb );

[0122] α depth =σ(W depth F depth +b depth );

[0123] α thermal =σ(W thermal F thermal +b thermal );

[0124] F fusion =α rgb ⊙F rgb +α depth ⊙F depth +α thermal ⊙F thermal ;

[0125] Where: α rgb , α depth , α thermal σ represents the attention weights for RGB, depth, and infrared modes, respectively; σ is the Sigmoid activation function, which normalizes the weights to the [0,1] interval; W rgb W depth W thermal These are the learnable weight matrices for RGB, depth, and infrared modes, respectively; b rgb , b depth , b thermal These are the RGB, depth, and infrared modal bias vectors, respectively; F fusion represents the fused feature vector; ⊙ represents element-wise multiplication.

[0126] (3) Decision-level fusion formula:

[0127] Bayesian fusion rules:

[0128] P(Y∣X rgb , Xdepth , X thermal )∝P(X rgb |Y), P(X) depth |Y), P(X) thermal |Y)P(Y);

[0129] Where: P(Y|X) rgb ,X depth ,X thermal X represents the posterior probability of food category Y given three modal observations; Y is the food category (e.g., "steak" or "salad"); X represents the food category (e.g., "steak" or "salad"). rgb , X depth , X thermal These are the observations for RGB, depth, and infrared modes, respectively; P(X) rgb |Y), P(X) depth |Y), P(X) thermal |Y) is the likelihood function, representing the probability of each modality observation given a category Y; P(Y) is the prior probability (based on the category distribution of the training data).

[0130] (4) Multi-task learning fusion formula:

[0131] L=λ1L class +λ2L calorie +λ3L temperature ;

[0132] Where: L is the total loss after fusion; L class For food classification loss (such as cross-entropy loss); L calorie For calorie prediction loss (e.g., mean squared error loss); L temperature λ1, λ2, and λ3 are the task weight coefficients for food classification loss, calorie prediction loss, and temperature estimation loss, respectively, which are dynamically adjusted using the Pareto optimality method.

[0133] b. Three-modal spatiotemporal fusion (for dynamic food recognition);

[0134] In real-world dining scenarios, food states may change over time (e.g., hot soup cooling down, or the color change during steak cooking). To accurately capture these changes, a spatiotemporal sequence data acquisition and fusion mechanism is introduced. Dynamic changes are detected using the inter-frame difference method.

[0135] ;

[0136] Among them: I t (x,y) is the food image at time t; I t-1 (x,y) represents the food image from the previous moment; D t(x,y) represents the pixel difference between two frames, used to detect changes in food state (such as stirring or heating).

[0137] Furthermore, a temporal attention mechanism is employed to process dynamic features:

[0138] ;

[0139] ;

[0140] ;

[0141] Where: h t For hidden state vectors; GRU is a gated recurrent unit that processes timing information; F fusion t Let be the fused feature vector at time t; The hidden state vector before the update; Let be the attention weight at time t; This is the weight parameter matrix that needs to be learned during model training; The bias term parameters are used for model training; i is the number of elements; T is the length of the time sequence. This refers to the fused features of the final output.

[0142] Taking "pan-fried steak" as an example, this mechanism can capture the color change of steak from raw to cooked (red to brown), combine it with temperature changes (from room temperature to high temperature), and combine multi-frame features with attention weights through temporal fusion to improve the accuracy of recognition during the cooking stage and ensure accurate recognition of dynamic food states in complex dining scenarios.

[0143] c. The fusion formula for uncertainty perception;

[0144] (1) Uncertainty in Monte Carlo Dropout estimation:

[0145] ;

[0146] ;

[0147] ;

[0148] in: , , The prediction uncertainty for each modality is represented by K; K represents the sample size. The softmax function transforms the model output into a probability distribution. , , The outputs of the models for the k-th Dropout sampling are RGB, depth, and infrared modes, respectively. , , These represent the average prediction probabilities from multiple samplings of the RGB, depth, and infrared modes, respectively.

[0149] Adaptive weight fusion:

[0150] (i∈{rgb,depth,thermal});

[0151] in: The weight values ​​are adaptively adjusted for perception fusion; , i, j are the variances of the i-th and j-th elements respectively; j is the index of the summation; i is the number of elements; rgb, depth, thermal are the data categories of RGB, depth and infrared modes respectively.

[0152] Modalities with lower uncertainty receive higher fusion weights, thus achieving robust fusion.

[0153] 3. Diet plan generation:

[0154] 3.1 Food Identification and Analysis: AI technology is used to identify the types of food, measure the volume, and calculate the content of nutrients such as calories, carbohydrates, protein, fat, vitamins, and minerals in photos of meals taken manually by users.

[0155] 3.2 Nutritional Assessment: Combining the calorie and nutrient data of the whole meal, the wristband assesses the total calorie and nutritional balance of the meal, calculates the proportions of nutrients such as carbohydrates, protein, and fat, as well as the content of various vitamins and minerals, and analyzes whether it meets the user's healthy eating requirements and nutritional balance standards.

[0156] 3.3 Personalized Recommendation Generation: Based on the user's pre-set health goals (weight loss, muscle gain, blood sugar control, etc.), dietary preferences (taste, food preferences), physical condition (allergies, specific diseases), and the overall meal calorie and nutritional assessment results, the band generates personalized dietary recommendations for the user through its built-in diet plan generation algorithm. For example, when the overall meal calorie intake exceeds the limit, it suggests that the user reduce the intake of certain high-calorie foods; when the nutrition is unbalanced, it suggests that the user increase the intake of foods rich in specific nutrients; when the food combination is unreasonable, it provides healthier combination schemes, etc.

[0157] IV. Feedback and Implementation Phase;

[0158] 1. Real-time heat feedback;

[0159] 1.1 Feedback Method: The calorie value of a single grab is converted into vibration intensity according to a preset mapping relationship (e.g., 25kcal corresponds to a light vibration, 50kcal corresponds to a medium vibration, and 100kcal and above corresponds to a strong vibration), and feedback is completed within 0.3 seconds. This allows users to intuitively perceive the approximate calorie range of the food they are grabbing through touch, making it convenient for them to adjust their food intake in real time during meals. At the same time, the calorie value will be clearly displayed on the wristband screen (e.g., "Calories of food grabbed this time: 28kcal"), reminding users of their current calorie intake through both tactile and visual means.

[0160] 1.2 Mapping Relationship Implementation: The mapping from heat value to vibration intensity is achieved through a piecewise linear function. At the same time, fuzzy logic is introduced to handle boundary cases, ensuring the accuracy and recognizability of the user's tactile feedback.

[0161] (1) Basic piecewise linear mapping formula:

[0162] ;

[0163] Where: VL is the vibration intensity level (an integer, the higher the level, the greater the vibration amplitude / frequency); Q is the calorie value of a single food pick-up (unit: kcal). Set the minimum perceptible heat threshold (e.g., 5 kcal, to avoid false triggering). The step size for heat classification (e.g., 25 kcal, corresponding to the basic vibration level interval). The maximum linearly mapped heat value (e.g., 200kcal, exceeding which triggers the highest level of vibration). This represents the maximum vibration intensity level (e.g., level 5, corresponding to a strong vibration mode).

[0164] (2) Operational logic:

[0165] No-action scenario: When Q=0 (no food is being picked up), the vibration level is 0 (no feedback).

[0166] Linear piecewise scenarios:

[0167] For example, if =5kcal, =25kcal, then:

[0168] When Q=15kcal ( <Q≤ ), (Slight vibration);

[0169] When Q=40kcal ( <Q≤ ), (Zhongzhen);

[0170] Exceeding limits: When Q> (For example, when picking up a high-calorie food of 250kcal), the highest level of vibration is directly triggered. This is accompanied by a warning icon on the display screen.

[0171] (3) Vibration characteristic parameter formula:

[0172] ;

[0173] ;

[0174] Where: A is the vibration amplitude (peak displacement, unit: mm); f is the vibration frequency (unit: Hz); A0 is the basic amplitude (e.g., 0.5 mm, corresponding to level 1); f0 is the basic frequency (e.g., 20 Hz, corresponding to level 1). For each increase of 1 in vibration intensity level, the vibration amplitude increases by 20% and the vibration frequency increases by 10%.

[0175] Example:

[0176] Level 1 (Minor Vibration): A=0.5mm, f=20Hz;

[0177] Level 3 (Strong Vibration): A = 0.5 × 1.22 = 0.72 mm, f = 20 × 1.12 = 24.2 Hz.

[0178] 1.3 Design Highlights: By adjusting both amplitude and frequency, the design ensures a clear tactile difference between different vibration levels, avoiding user confusion (e.g., light vibration is a single, short vibration, while strong vibration is a continuous high-frequency vibration); the mapping calculation time is controlled within 10ms, and combined with hardware optimization (e.g., a dedicated DSP chip), feedback is ensured to be completed within 0.3s; the formula... Parameters such as A0 and f0 can be customized by the user through the bracelet settings interface to meet personalized tactile sensitivity needs (such as increasing the amplitude threshold for people with low tactile sensitivity).

[0179] 2. Cumulative Dynamic Alert: The wristband records the calorie value of each food item picked up in real time and compares it with the target value set by the user through the target calorie setting module 8. When the cumulative calories approach the target value, it reminds the user to control their diet through specific alert modes such as continuous short vibrations; if it exceeds the target value, it emits a long vibration accompanied by a flashing dial, which is a clear alert to help users avoid excessive calorie intake and achieve dynamic and precise control of dietary calories.

[0180] 3. Color ring display of nutritional proportions on the watch face: The wristband display screen is designed with a dynamic color ring display area. According to the proportion of nutrients such as carbohydrates, proteins, and fats, the color ring is divided into different color blocks (such as green representing the proportion of carbohydrates, blue representing the proportion of proteins, and red representing the proportion of fats). This provides a real-time and intuitive display of the nutritional proportions of the food being picked up, allowing users to understand the calorie value and clearly recognize the nutritional balance of the food, thus helping them make reasonable dietary choices.

[0181] 4. Visualized Diet Plans: The generated whole-meal dietary suggestions are displayed on the wristband screen in the form of pictures and text (such as recommended food pairing charts and recommended serving sizes for each ingredient), presenting a nutritionally balanced plan in an intuitive way.

[0182] 5. Dynamic behavioral intervention: The wristband can remind users to follow dietary recommendations based on real-time food picking and cumulative calorie intake during meals, guide users to adjust their eating behavior, ensure that daily calorie intake meets preset goals, and help users achieve scientific and reasonable dietary goals.

[0183] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0184] Example 1: A middle-aged user with type 2 diabetes used a wearable calorie-tracking wristband during daily family meals. In a typical family meal scenario, each time food was picked up, the wristband's front and rear cameras (11 and 12) quickly captured images of the food from different angles. The AI ​​model accurately identified the food type (e.g., rice, vegetables, braised pork, etc.) and calculated the calories. When a piece of braised pork was picked up, the wristband vibrated slightly (indicating approximately 25 kcal), and the display showed "Calories of food picked up: 28 kcal," with the dial's color ring showing the nutritional composition of the food, with a relatively high fat content. As the meal continued, the wristband accumulated the user's calorie intake. When the user's accumulated calorie intake approached their preset daily target of 1600 kcal, the wristband emitted continuous short vibrations as a warning and displayed "Accumulated calories are approaching the target value; please pay attention to controlling your diet," reminding the user to control their food intake. If the user continued to eat beyond the target value, the wristband vibrated continuously with the dial flashing, displaying "Accumulated calories have exceeded the limit; please control your diet." In addition, before each meal, users took photos of all the food they ate. The diet plan generated by the wristband suggested reducing the intake of braised pork and increasing the proportion of vegetables to better control calories and balance nutrition. After a period of use, users' average daily calorie intake decreased by about 18% compared to before using the band, their blood sugar levels became more stable, and their diet gradually became more reasonable, providing strong support for diabetes dietary management.

[0185] Example 2: An obese patient used a wearable calorie-tracking bracelet while dining with friends at a restaurant. Faced with a variety of dishes, the bracelet captures images of the food the user picks up and performs recognition and calculations. When the user unknowingly consumes too many high-calorie foods such as fried foods, and the cumulative calorie intake is about to exceed the set daily target of 1800kcal, the bracelet promptly issues a warning. The display clearly shows the current cumulative calorie value and an over-limit reminder, "Cumulative calories have exceeded the limit, please control your diet," allowing the user to adjust their food choices in time and avoid excessive calorie intake. Simultaneously, based on the nutritional ratios displayed on the watch face, the user consciously increases their intake of nutritionally balanced foods such as vegetables. Furthermore, before the meal, the user takes photos of all the food in the meal, and the bracelet's generated dietary plan suggests replacing some high-calorie fried foods with low-calorie vegetables to reduce the overall calorie intake and improve nutritional balance. After long-term consistent use, the user's weight was effectively controlled, and various health indicators improved, fully demonstrating the practicality and effectiveness of the bracelet in dietary management for obese individuals, helping them better manage their daily diet and move towards a healthier life.

[0186] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A wearable thermal card reminder wristband, characterized in that, The bracelet includes a main body, with a front watch and a rear watch respectively disposed on the front and rear sides of the main body. The front watch and the rear watch have the same structure and both include: The camera, used to capture images of food, consists of dual front-facing cameras and dual rear-facing cameras; Motion sensors are used to collect data on the movement trajectory of tableware; Microphones are used to collect collision soundprint data; The processor is used to run AI recognition and algorithms; Vibration module for heat indication and early warning vibration feedback; The display screen shows the dial color rings and calorie information; The target calorie setting module is used to customize the daily calorie intake target value; The diet plan generation module is used to generate reasonable diet plans; The bracelet works in the following four stages: I. Target Calorie Setting and Personalized Management Phase: Users set their daily calorie intake target value through the target calorie setting module, and the wristband intelligently adjusts the target value based on historical data and provides reminders; II. Data Acquisition Stage: Motion sensors collect data on the movement trajectory of the tableware; cameras on the front and rear tables capture multi-view images the moment food is picked up; and microphones collect collision soundprint data. III. Data Processing and Analysis Stage: After the processor preprocesses the image, it identifies the food type through an AI model, calculates the food volume using binocular ranging technology, and then calculates the calorie value based on the food calorie database. At the same time, the processor extracts features from the tableware movement trajectory data, visual data, and voiceprint data, and optimizes the calorie calculation, food type identification, and calorie calculation accuracy through a three-modal data fusion algorithm. Finally, the diet plan generation module generates personalized diet suggestions. IV. Feedback and Execution Phase: The vibration module triggers vibration feedback based on the calorie value, and the display screen shows the calorie value and nutrient ratio, as well as the dietary plan.

2. The wearable heat card reminder wristband according to claim 1, characterized in that, When the processor preprocesses the image, it first uses the Laplacian variance algorithm to evaluate the image sharpness. When the variance of the Laplacian operator is less than 100, it filters out blurry data. Then, it performs image enhancement and normalization operations. Image enhancement includes random rotation and random scaling. The formula for the Laplace variance algorithm is: ； in: The variance of the Laplace operator; The Laplacian operator for image I; via Calculate the variance of the Laplacian operator, where x and y are pixel coordinates; The formula for random rotation is: ； in: This represents the pixel value of the rotated image at coordinates (x, y). Pixel value; Let θ be the rotation matrix, and θ be a random rotation angle. It is the inverse of the rotation matrix; It is in vector form with coordinates (x, y); The formula for random scaling is: Isc(x,y)=I(s / x,s / y); Where: Isc(x,y) are the scaled pixel values; s is the scaling factor, when s>1 the image is shrunk, and when s<1 the image is enlarged; The normalization formula is: ； Where: μ is the pixel mean of the training set images; σ is the pixel standard deviation of the training set images; This represents the pixel value of the original image at coordinates (x, y). This represents the pixel value of the normalized image at coordinates (x, y).

3. The wearable heat card reminder wristband according to claim 1, characterized in that, The AI ​​model is based on deep learning algorithms and was trained using image data of more than 3,000 kinds of food ingredients. During training, Mixup data augmentation and category balancing sampling strategies were used. The Mixup data augmentation formula is: ； ； in: The enhanced pixel values; Enhanced food category labels; These are two randomly selected food images; For the corresponding food category label; It is the mixing coefficient that follows a Beta distribution; The category-balanced sampling formula is: ； in: is the weight of the i-th food sample; N is the total number of samples; Let C be the number of samples in the i-th category; C is the total number of food categories.

4. The wearable heat card reminder wristband according to claim 1, characterized in that, The three-modal data fusion algorithm is implemented in the following way: Feature-level fusion employs a weighted attention mechanism, the formula of which is: F fusion =a rgb ⊙F rgb +a depth ⊙F depth +a thermal ⊙F thermal ; Wherein: F fusion The fused feature vector; α rgb , α depth , α thermal These represent the attention weights for RGB, depth, and infrared modes, respectively; ⊙ represents element-wise multiplication; F rgb , F depth , F thermal These are the extracted RGB feature vector, depth feature vector, and infrared feature vector, respectively. The decision-level fusion adopts Bayesian rules, and its formula is as follows: P(Y∣X rgb , X depth , X thermal )∝P(X rgb ∣Y), P(X depth ∣Y), P(X thermal ∣Y)P(Y); Where: P(Y|X) rgb ,X depth ,X thermal X represents the posterior probability of food category Y given three modal observations; Y is the food category; X represents the food category. rgb X depth X thermal These are the observations for RGB, depth, and infrared modes, respectively; P(X) rgb |Y), P(X) depth |Y), P(X) thermal |Y) is the likelihood function; P(Y) is the prior probability; For dynamic food recognition, a temporal attention mechanism is used to process dynamic features, and its formula is as follows: ； ； ； Where: h t For hidden state vectors; GRU is a gated recurrent unit; F fusion t Let be the fused feature vector at time t; The hidden state vector before the update; Let be the attention weight at time t; This is the weight parameter matrix that needs to be learned during model training; The bias term parameters are used for model training; i is the number of elements; T is the length of the time sequence. The fusion features are the final output. Uncertainty-aware fusion is achieved through adaptive weight adjustment, and its formula is as follows: (i∈{rgb,depth,thermal}); in: The weight values ​​are adaptively adjusted for perception fusion; , i, j are the variances of the i-th and j-th elements respectively; j is the index of the summation; i is the number of elements; rgb, depth, thermal are the data categories of RGB, depth and infrared modes respectively.

5. The wearable heat card reminder wristband according to claim 1, characterized in that, The vibration module converts real-time heat values ​​into vibration intensity, and the mapping is achieved through a piecewise linear function. The basic piecewise linear mapping formula is: ； Where: VL is the vibration intensity level; Q is the calorie value of a single food pick-up; The minimum perceptible heat threshold; The step size for heat fractionation; This represents the maximum linearly mapped heat value. This represents the highest level of vibration intensity. The formulas for how vibration amplitude and frequency change with vibration intensity level are: ； ； Where: A is the vibration amplitude; f is the vibration frequency; A0 is the fundamental amplitude; f0 is the fundamental frequency.

6. The wearable heat card reminder wristband according to claim 1, characterized in that, The display screen dynamically displays a color ring, which is divided into different color blocks according to the proportion of carbohydrates, proteins, and fats, so as to show the nutritional distribution of the food being picked up in real time.

Images