Fall detection method and system based on wristband type smart watch
Through parallel spatial and timing attention branches, the movement data characteristics of wristband smart watches are extracted, combined with the deep learning model, the problem of insufficient fall detection accuracy caused by wrist movement complexity is solved, and accurate recognition and real-time alarms are achieved for the elderly.
Patent Information
- Application Number
- CN202510653947.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-21
- Publication Date
- 2025-08-19
AI Technical Summary
The traditional fall detection method based on wrist movement data is insufficient in identifying fall behavior in the elderly, especially when the wrist movement status is complex and changeable, it is easy to cause misjudgment and misjudgment.
Parallel spatial attention branches and timing attention branches are used to extract the features of wristband-type smart watch motion data, combined with deep learning models for fall detection, and through the spatial attention feature extraction module and timing attention feature extraction module, the motion data features are integrated and pattern recognition is performed.
It improves the accuracy and robustness of fall detection, realizes accurate identification and real-time alarms of fall behavior in the elderly, and reduces misjudgment and misjudgment.
Smart Images
Figure CN120501419A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of artificial intelligence and pattern recognition, and relates to a fall detection method and system based on a wristband-type smart watch. Background Art
[0002] With the increasing aging of society and the increasing number of elderly people living alone, the health and safety of the elderly has become a major concern. Falls are a common and highly dangerous accident for the elderly, posing a serious threat to their safety and quality of life. According to relevant data, approximately one-third of people aged 65 and over experience at least one fall each year. Falls often lead to serious consequences such as fractures and head injuries, and in some cases, can be fatal. More importantly, if medical intervention is not received quickly after a fall, especially within the "golden hour" after the fall, the injury can be significantly aggravated, increasing disability and mortality rates. Furthermore, even if a fall does not cause significant physical damage, the psychological trauma of the elderly can still lead to a fear of falling, which can lead to restricted activities, reduced social interactions, and a decline in overall self-care ability, creating a vicious cycle.
[0003] With the rapid development of wearable technology, smartwatches, as convenient and intelligent mobile devices, have been widely used in the field of health monitoring. Smartwatches are easy to wear and can continuously and in real time acquire dynamic data such as acceleration and angular velocity during human motion. They have high coverage and are not restricted by usage scenarios. Furthermore, smartwatches offer excellent portability and wireless communication capabilities. Upon detecting a fall, they can quickly send an alarm message via wireless communication methods such as Wi-Fi or Bluetooth, enabling timely notification to caregivers or family members, significantly shortening medical response time. Thirdly, smartwatch solutions can provide all-weather, personalized fall detection and alarms without infringing on user privacy, offering greater practicality and promotional value. However, due to the greater flexibility of the wrist and its complex and variable motion states, fall detection can be prone to misjudgment and omission, increasing the difficulty of accurate identification.
[0004] Traditional methods based on rules or simple feature thresholds struggle to effectively address the diversity of wrist movements. With the development of deep learning technology, by building multi-level feature extraction and discrimination models, it is possible to automatically extract effective features from complex motion data, significantly improving the accuracy and robustness of fall detection. Therefore, the development of deep learning detection methods and systems for wristband watches can significantly improve the safety of special populations such as the elderly, and has important social value and broad application prospects. Summary of the Invention
[0005] In response to the problems of poor detection effect of traditional detection methods and large interpersonal differences in wrist motion data, the present invention proposes a fall detection method and system based on wristband watch motion data, which can achieve accurate identification of fall behavior of the elderly and real-time alarm.
[0006] The specific technical solutions of the present invention are:
[0007] A fall detection method based on a wristband smartwatch extracts spatial and temporal features of motion data through parallel spatial attention branches and temporal attention branches, respectively, and applies an attention mechanism to adjust the importance of each feature in the fusion process. The method includes the following steps:
[0008] Step (1) collecting sensor data from a wristband smartwatch, the sensor including a three-axis accelerometer, a gyroscope, and a magnetometer; inputting the collected data into an extended Kalman filter algorithm to perform nonlinear state estimation to obtain accurate Euler angle data as subsequent attitude measurement data;
[0009] Step (2) performs L2 regularization processing on the acquired acceleration, angular velocity and attitude measurement data, and then performs data segmentation to divide the processed nine-axis data set into a training set and a test set;
[0010] Step (3) constructing a fall detection model based on a wristband smart watch, the model including a spatial attention branch, a temporal attention branch and a recognition module; wherein the spatial attention branch is used to extract and integrate spatial features of the nine-axis sensor data, including a spatial attention feature extraction module and a spatial feature fusion module; the temporal attention branch is used to extract and integrate temporal features of the nine-axis sensor data, including a temporal attention feature extraction module and a temporal feature fusion module; the recognition module is used to integrate the output feature maps of the spatial attention branch and the temporal attention branch, and reduce the feature dimension to generate a recognition result;
[0011] Specifically, the nine-axis sensor data is input into the spatial attention feature extraction module to obtain a spatial attention feature map. The spatial attention feature extraction module is composed of four parallel two-dimensional convolutional layers and a maximum pooling layer, a two-dimensional convolutional layer, and a channel attention layer. The spatial attention feature map is input into the spatial feature fusion module to obtain a fused spatial feature map after feature integration. The spatial feature fusion module is composed of a two-dimensional convolutional layer, a batch normalization layer, and a maximum pooling layer.
[0012] The nine-axis sensor data is input into the temporal attention feature extraction module to obtain a temporal attention feature map. The temporal attention feature extraction module is composed of a bidirectional gated recurrent unit layer, a cross attention layer, a bidirectional gated recurrent unit layer, and a cross attention layer. The two cross attention layers share parameters. The temporal attention feature map is input into the temporal feature fusion module to obtain a fused temporal feature map after feature integration. The temporal feature fusion module is composed of two one-dimensional convolutional layers and a batch normalization layer.
[0013] The recognition module performs an element-wise feature fusion operation on the fused spatial feature map and the fused temporal feature map, and then passes them through three linear layers in sequence to achieve pattern recognition of fall and non-fall behaviors;
[0014] Step (4) training the fall detection model based on the wristband smartwatch built in step (3) to distinguish between fall and non-fall behaviors; adjusting hyperparameters to obtain optimal performance, repeatedly iterating the fall detection model so that the loss value converges to a minimum value and saving the optimal model;
[0015] Step (5) inputs the test set into the optimal performance fall detection model, and evaluates the fall detection model to verify the effectiveness and accuracy of its fall detection; after the verification is passed, the optimal performance fall detection model is integrated into the fall detection system based on the wristband smart watch to realize the real-time fall behavior detection function.
[0016] Furthermore, the channel attention layer in the spatial attention feature extraction module consists of an average pooling layer, two linear layers and a sigmoid activation layer in sequence, which assigns different weights to different channels of the feature map, making the model more focused on useful features and reducing the interference of irrelevant features.
[0017] Furthermore, in the cross-attention layer in the temporal attention feature extraction module, the input features are first linearly transformed to generate a query vector, a key vector, and a value vector; then, attention calculations are performed in the horizontal and vertical directions of the feature map: the correlation between different positions is measured by the dot product operation between the query vector and the key vector on the same row and column, and the dot product result is scaled and converted into an attention weight distribution through a normalized exponential function; then, the obtained attention weights are used to perform weighted summation on the value vectors at the corresponding positions to obtain the attention outputs in the horizontal and vertical directions respectively; finally, the horizontal and vertical attention outputs are fused and added to the input features at the corresponding positions of the elements to obtain a feature representation containing rich cross-spatial position dependencies.
[0018] A fall detection system based on a wristband smartwatch, based on the above detection method, includes a sensor module, a main control module and an alarm module, wherein:
[0019] The sensor module detects the user's acceleration, angular velocity and magnetic field data in the x, y and z directions, and calculates the posture measurement data by fusing the Kalman filter;
[0020] The main control module performs L2 regularization and data segmentation on the acceleration, angular velocity, and attitude measurement data, and inputs the processed nine-axis motion data into a fall detection model based on a wristband smart watch to determine whether the user has fallen. When the judgment result is a fall, the alarm module is triggered;
[0021] The alarm module, when it is determined that the person has fallen, will sound an alarm and inform the guardian of the alarm information through wireless transmission.
[0022] The present invention has the following beneficial effects:
[0023] (1) The present invention is aimed at wristband-type smart watches, which monitor users in real time by measuring three motion signals: acceleration, angular velocity, and posture. It has a high coverage rate and is not restricted by usage scenarios.
[0024] (2) The present invention adopts a method of preprocessing data to make the connection between data stronger and maintain the structural correlation between features.
[0025] (3) The present invention proposes a fall detection model with parallel spatial attention branches and temporal attention branches, which solves the problems of insufficient feature extraction and large interpersonal differences in wrist motion data, and realizes accurate fall detection.
[0026] (4) The present invention is targeted at wristband-type smart watches and can be easily integrated into existing smart watches. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 is a schematic diagram of a fall detection system of the present invention;
[0028] Figure 2 is a flow chart of the fall detection method of the present invention;
[0029] Figure 3 This is a structural diagram of a fall detection model based on a wristband-type smart watch according to the present invention;
[0030] Figure 4 : This is a diagram of the channel attention layer structure of the spatial attention branch of the present invention;
[0031] Figure 5 It is a structural diagram of the cross-attention layer of the temporal attention branch of the present invention. DETAILED DESCRIPTION
[0032] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.
[0033] A fall detection system based on wristwatch motion data, such as Figure 1 As shown, it includes a sensor module, a main control module and an alarm module;
[0034] The sensor module includes a three-axis accelerometer, a three-axis gyroscope, a three-axis magnetometer, and a fusion Kalman filter. The three-axis accelerometer measures the user's acceleration data in the x, y, and z directions. The gyroscope measures the user's angular velocity data in the x, y, and z directions. The three-axis magnetometer measures the user's magnetic field data in the x, y, and z directions. The fusion Kalman filter calculates attitude measurement data using acceleration, angular velocity, and magnetic field data.
[0035] The main control module connects the sensor module and the alarm module; the main control module processes and analyzes the acceleration, angular velocity, and posture measurement data input by the sensor module, and performs pattern recognition to determine whether a fall has occurred; if a fall is detected, an alarm command is sent to the alarm module;
[0036] The alarm module is connected to the main control module and is used to send out an alarm signal; when a user falls, the alarm module will receive the alarm command from the main control module, send out an alarm, and inform the guardian of the alarm information through wireless transmission.
[0037] Based on the above fall detection system, the present invention also provides a fall detection method based on a wristband smart watch, such as Figure 2 The specific implementation process is as follows:
[0038] Step 1: Create a dataset;
[0039] Develop a wristband smartwatch-based data acquisition application to record acceleration, angular velocity, and magnetic data during human activities with a sampling frequency of 200Hz; use a fusion Kalman filter algorithm for nonlinear state estimation to obtain accurate Euler angle data as subsequent attitude measurement data; and combine the acceleration, angular velocity, and attitude measurement data into nine-axis sensor data.
[0040] Step 2: Preprocess the data and divide it into training set and test set;
[0041] For the nine-axis sensor data, the mean and standard deviation are calculated to obtain the L2 normalized data. The data is then segmented using a sliding window with a window size of 200 and an overlap rate of 50%.
[0042] An individual-based five-fold cross-validation strategy was used, in which 80% of the individual data were used to train the model and 20% of the individual data were used to test the model in each fold.
[0043] Step 3: Build a fall detection model based on a wristband smartwatch, such as Figure 2 As shown;
[0044] Includes spatial attention branch, temporal attention branch and recognition module;
[0045] The spatial attention branch includes a spatial attention feature extraction module and a spatial feature fusion module; the temporal attention branch includes a temporal attention feature extraction module and a temporal feature fusion module; the recognition module consists of three linear layers; Figure 3 As shown;
[0046] The spatial attention feature extraction module consists of four parallel two-dimensional convolutional layers and a maximum pooling layer, a two-dimensional convolutional layer and a channel attention layer. The input dimension of the four parallel two-dimensional convolutional layers and the maximum pooling layer is 9, the number of output channels is 8, 16, 32, 32, and the kernel size is 3, 5, 7, and 3, respectively. The number of input channels of a two-dimensional convolutional layer is 88, the number of output channels is 64, and the kernel size is 3; the structure of the channel attention layer is as follows: Figure 4 As shown in the figure, the input features pass through an average pooling layer, two linear layers and a Sigmoid activation layer, which can assign different weights to different channels of the feature map, allowing the model to focus more on useful features and reduce interference with irrelevant features; the number of input channels of the two linear layers is 64 and 4, respectively, and the number of output channels is 4 and 64, respectively.
[0047] The spatial feature fusion module consists of a two-dimensional convolutional layer, a batch normalization layer, and a maximum pooling layer. The two-dimensional convolutional layer has an input dimension of 64, an output dimension of 64, and a kernel size of 3.
[0048] The temporal attention feature extraction module consists of two bidirectional gated recurrent unit layers and a cross attention layer, where the input dimensions of the bidirectional gated recurrent unit layer are 9 and 128, the output dimensions are 128 and 64, and the number of hidden layers is 1. The structure of the cross attention layer is as follows: Figure 5As shown in the figure, the input features are first linearly transformed to generate a query vector, a key vector, and a value vector. Subsequently, attention calculations are performed in the horizontal and vertical directions of the feature map: the correlation between different positions is measured by performing a dot product operation between the query vector and the key vector in the same row and column. The dot product result is scaled and converted into an attention weight distribution using a normalized exponential function. The obtained attention weights are then used to perform a weighted summation of the value vectors at the corresponding positions, respectively, to obtain the attention outputs in the horizontal and vertical directions. Finally, the horizontal and vertical attention outputs are fused and added to the input features at the corresponding position of the element, forming a feature output containing rich cross-spatial position dependencies.
[0049] The temporal feature fusion module consists of two one-dimensional convolutional layers and a batch normalization layer. The input dimensions of the one-dimensional convolutional layers are 128 and 64, respectively, and the output dimensions are 64 and 64, respectively.
[0050] The recognition module flattens the feature maps obtained by the spatial attention branch and the temporal attention branch, then adds them element-wise, and passes the added feature vector into a multi-layer perceptron consisting of three linear layers, where the input dimensions of the linear layers are 64, 128, and 64, respectively, and the output dimensions are 128, 64, and 2, respectively.
[0051] Step 4: Set hyperparameters and train the model;
[0052] During model training, the epoch number was set to 100, the batch size was set to 64, and the cosine learning rate decay strategy was used to train the model, with the initial learning rate set to 0.0001 and the minimum learning rate set to 0.000001. In addition, the Youden index was used to determine the optimal classification threshold.
[0053] Step 5: Test the model and integrate it into a fall detection system based on wristwatch motion data.
[0054] The trained model was used to test the individual data in the test set, and the AUC value, sensitivity, specificity, F1 value and accuracy were calculated. The optimal model was integrated into the main control module of the fall detection system based on wristband watch motion data to determine whether a fall has occurred.
Claims
1. A fall detection method based on a wristband smart watch, characterized in that: The steps include: Step (1) collecting sensor data from a wristband smartwatch, the sensor including a three-axis accelerometer, a gyroscope, and a magnetometer; inputting the collected data into an extended Kalman filter algorithm to perform nonlinear state estimation to obtain accurate Euler angle data as subsequent attitude measurement data; Step (2) performs L2 regularization processing on the acquired acceleration, angular velocity and attitude measurement data, and then performs data segmentation to divide the processed nine-axis data set into a training set and a test set; Step (3) constructing a fall detection model based on a wristband smart watch, the model including a spatial attention branch, a temporal attention branch and a recognition module; wherein the spatial attention branch is used to extract and integrate spatial features of the nine-axis sensor data, including a spatial attention feature extraction module and a spatial feature fusion module; the temporal attention branch is used to extract and integrate temporal features of the nine-axis sensor data, including a temporal attention feature extraction module and a temporal feature fusion module; the recognition module is used to integrate the output feature maps of the spatial attention branch and the temporal attention branch, and reduce the feature dimension to generate a recognition result; Specifically, the nine-axis sensor data is input into the spatial attention feature extraction module to obtain a spatial attention feature map. The spatial attention feature extraction module is composed of four parallel two-dimensional convolutional layers and a maximum pooling layer, a two-dimensional convolutional layer, and a channel attention layer. The spatial attention feature map is input into the spatial feature fusion module to obtain a fused spatial feature map after feature integration. The spatial feature fusion module is composed of a two-dimensional convolutional layer, a batch normalization layer, and a maximum pooling layer. The nine-axis sensor data is input into the temporal attention feature extraction module to obtain a temporal attention feature map. The temporal attention feature extraction module is composed of a bidirectional gated recurrent unit layer, a cross attention layer, a bidirectional gated recurrent unit layer, and a cross attention layer. The two cross attention layers share parameters. The temporal attention feature map is input into the temporal feature fusion module to obtain a fused temporal feature map after feature integration. The temporal feature fusion module is composed of two one-dimensional convolutional layers and a batch normalization layer. The recognition module performs an element-wise feature fusion operation on the fused spatial feature map and the fused temporal feature map, and then passes them through three linear layers in sequence to achieve pattern recognition of fall and non-fall behaviors; Step (4) training the fall detection model based on the wristband smartwatch built in step (3) to distinguish between fall and non-fall behaviors; adjusting hyperparameters to obtain optimal performance, repeatedly iterating the fall detection model so that the loss value converges to a minimum value and saving the optimal model; Step (5) inputs the test set into the optimal performance fall detection model, and evaluates the fall detection model to verify the effectiveness and accuracy of its fall detection; after the verification is passed, the optimal performance fall detection model is integrated into the fall detection system based on the wristband smart watch to realize the real-time fall behavior detection function.
2. A fall detection method based on a wristband smart watch according to claim 1, characterized in that: The channel attention layer in the spatial attention feature extraction module consists of an average pooling layer, two linear layers and a sigmoid activation layer in sequence to give different weights to different channels of the feature map.
3. The fall detection method based on a wristband smart watch according to claim 1, characterized in that: In the cross-attention layer in the temporal attention feature extraction module, the input features are first linearly transformed to generate query vectors, key vectors, and value vectors; then, attention calculations are performed in the horizontal and vertical directions of the feature map: the correlation between different positions is measured by the dot product operation between the query vector and the key vectors on the same row and column, and the dot product result is scaled and converted into an attention weight distribution through a normalized exponential function; then, the obtained attention weights are used to perform weighted summation on the value vectors at the corresponding positions to obtain the attention outputs in the horizontal and vertical directions respectively; finally, the horizontal and vertical attention outputs are fused and added to the input features at the corresponding positions of the elements to obtain a feature representation containing rich cross-spatial position dependencies.
4. A fall detection system based on a wristband smart watch, based on the detection method according to any one of claims 1 to 3, characterized in that: It includes sensor module, main control module and alarm module, among which: The sensor module detects the user's acceleration, angular velocity and magnetic field data in the x, y and z directions, and calculates the posture measurement data by fusing the Kalman filter; The main control module performs L2 regularization and data segmentation on the acceleration, angular velocity, and attitude measurement data, and inputs the processed nine-axis motion data into a fall detection model based on a wristband smart watch to determine whether the user has fallen. When the judgment result is a fall, the alarm module is triggered; The alarm module, when it is determined that the person has fallen, will sound an alarm and inform the guardian of the alarm information through wireless transmission.