Abnormal fall detection methods, electronic devices and software products

Abstract

This application provides an abnormal fall recognition method, electronic device, and program product, which can be used in the field of fall recognition technology. The method is applied to wearable devices and includes: acquiring a detection dataset; wherein the detection dataset includes WiFi signal data collected by a WiFi unit, audio data collected by an audio acquisition unit, and inertial measurement data collected by an inertial measurement unit; determining an indoor scene based on the WiFi signal data; inputting the indoor scene and audio data into a first fall recognition model to obtain a first fall probability; inputting the indoor scene and inertial measurement data into a second fall recognition model to obtain a second fall probability; and fusing the first and second fall probabilities based on the indoor scene to obtain a fall recognition result. This application simultaneously identifies fall behavior using audio data and inertial measurement data and introduces scene factors, thereby significantly improving the accuracy of abnormal fall recognition through multi-dimensional data.

Description

Technical Field

[0001] This application relates to the field of fall detection technology, and more particularly to an abnormal fall detection method, electronic device, and program product. Background Technology

[0002] With the increasing aging of society, the phenomenon of elderly people living alone is becoming more and more common. If an elderly person falls indoors, the lack of assistance makes it difficult to detect the fall, leading to delays in optimal treatment. Therefore, there is an urgent need for effective abnormal fall detection technology to improve this situation.

[0003] Existing fall detection technologies mainly rely on single-dimensional information (such as motion information or image information) to determine falls. However, due to the complexity of indoor environments, existing fall detection technologies are easily affected by environmental interference, leading to misjudgments and poor accuracy in identifying abnormal fall behaviors. Summary of the Invention

[0004] This application provides an abnormal fall recognition method, electronic device, and program product to solve the technical problem that existing abnormal fall recognition technologies have poor accuracy in recognizing abnormal fall behavior when applied in indoor environments.

[0005] According to a first aspect disclosed in this application, this application provides an abnormal fall detection method applied to a wearable device, the wearable device including a WiFi unit, an audio acquisition unit, and an inertial measurement unit, the method comprising: Obtain a detection dataset; wherein the detection dataset includes WiFi signal data collected by the WiFi unit, audio data collected by the audio acquisition unit, and inertial measurement data collected by the inertial measurement unit; Based on the WiFi signal data, the indoor scene is determined; The indoor scene and the audio data are input into the first fall recognition model to obtain the first fall probability output by the first fall recognition model; The indoor scene and the inertial measurement data are input into the second fall recognition model to obtain the second fall probability output by the second fall recognition model; The first fall probability and the second fall probability are fused based on the indoor scene to obtain the fall recognition result.

[0006] In one feasible implementation, the first fall recognition model includes a first input layer, a convolutional feature encoder, an acoustic attention weight generator, an acoustic attention module, and a first classifier. The indoor scene and the audio data are input into the first fall recognition model to obtain a first fall probability output by the first fall recognition model, including: Based on the first input layer, the audio data is preprocessed to obtain the Mel spectrogram corresponding to the audio data; The acoustic features are obtained by performing convolutional encoding on the Mel spectrogram based on the convolutional feature encoder. Based on the acoustic attention weight generator, attention weights are calculated based on the indoor scene to obtain acoustic attention weights; The acoustic attention module weights the acoustic features based on the acoustic attention weights to obtain acoustic weighted features; The acoustic weighted features are classified based on the first classifier to obtain the first fall probability.

[0007] In one feasible implementation, the first fall recognition model is obtained based on the following method: Obtain a first training set; wherein the first training set includes multiple audio training samples, and a first real fall label and a first scene real label corresponding to the audio training samples; Input the first training set into a pre-built first deep learning model to obtain the first predicted fall label output by the first deep learning model; Based on the first predicted fall label and the first actual fall label, a first loss function is constructed; wherein, the first loss function includes a focus loss function; The first deep learning model is iteratively trained based on the first loss function to obtain the first fall recognition model.

[0008] In one feasible implementation, the second fall recognition model includes a second input layer, a temporal feature encoder, a motion attention weight generator, a motion attention module, and a second classifier. The indoor scene and the inertial measurement data are input into the second fall recognition model to obtain the second fall probability output by the second fall recognition model, including: The inertial measurement data is preprocessed based on the second input layer to obtain the time series features corresponding to the inertial measurement data; Based on the temporal feature encoder, the time series features are temporally encoded to obtain motion features; Based on the motion attention weight generator, attention weights are calculated based on the indoor scene to obtain motion attention weights; The motion attention module weights the motion features based on the motion attention weights to obtain motion-weighted features; The motion-weighted features are classified based on the second classifier to obtain the second fall probability.

[0009] In one feasible implementation, the second fall recognition model is obtained based on the following method: Obtain a second training set; wherein the second training set includes multiple inertial measurement training samples, and a second real fall label and a second real scene label corresponding to the inertial measurement training samples; The second training set is input into the pre-built second deep learning model to obtain the second predicted fall label output by the second deep learning model; Based on the second predicted fall label and the second actual fall label, a second loss function is constructed; wherein, the second loss function includes a weighted binary cross-entropy loss function and a temporal consistency loss function; The second deep learning model is iteratively trained based on the second loss function to obtain the second fall recognition model.

[0010] In one feasible implementation, determining the indoor scene based on the WiFi signal data includes: The WiFi signal data is input into the scene recognition model to obtain the indoor scene output by the scene recognition model; wherein, the scene recognition model is trained based on a third deep learning model.

[0011] In one feasible implementation, the scene recognition model is obtained based on the following method: The indoor area is divided into multiple indoor scenes, and WiFi signal data at multiple locations within each indoor scene are measured to establish a WiFi dataset; wherein, the WiFi dataset includes multiple WiFi signal data and the scene real labels corresponding to the WiFi signal data; Input the WiFi dataset into the third deep learning model to obtain the scene prediction label output by the third deep learning model; Based on the scene's real labels and the scene's predicted labels, a third loss function is constructed for the third deep learning model; wherein, the third loss function is used to characterize the difference between the scene's real labels and the scene's predicted labels; The scene recognition model is obtained by iteratively training the third deep learning model based on the third loss function.

[0012] In one feasible implementation, the first fall probability and the second fall probability are fused based on the indoor scene to obtain a fall recognition result, including: The indoor scene is input into a weight mapping table to obtain the acoustic recognition weight and motion recognition weight corresponding to the indoor scene; wherein, the weight mapping table is used to indicate the acoustic recognition weight and motion recognition weight corresponding to different indoor scenes; The fall recognition result is obtained by weighting and fusing the acoustic recognition weights and the first fall probability, the motion recognition weights and the second fall probability.

[0013] According to a second aspect disclosed in this application, this application provides an abnormal fall detection device applied to a wearable device, the wearable device including a WiFi unit, an audio acquisition unit, and an inertial measurement unit, the device comprising: A data acquisition module is used to acquire a detection dataset; wherein the detection dataset includes WiFi signal data collected by the WiFi unit, audio data collected by the audio acquisition unit, and inertial measurement data collected by the inertial measurement unit; A scene recognition module is used to determine the indoor scene based on the WiFi signal data; The first recognition module is used to input the indoor scene and the audio data into the first fall recognition model to obtain the first fall probability output by the first fall recognition model; The second recognition module is used to input the indoor scene and the inertial measurement data into the second fall recognition model to obtain the second fall probability output by the second fall recognition model. The recognition fusion module is used to fuse the first fall probability and the second fall probability based on the indoor scene to obtain a fall recognition result.

[0014] According to a third aspect disclosed in this application, this application provides an electronic device, including a processor and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method described in any one of the first aspects.

[0015] According to the fourth aspect disclosed in this application, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the method described in any one of the first aspects.

[0016] According to the fifth aspect disclosed in this application, this application provides a computer program product, including a computer program, which, when executed, is used to implement the method described in any one of the first aspects.

[0017] Compared with the prior art, this application has the following advantages: This application provides an abnormal fall recognition method, electronic device, and program product. By acquiring kinematic parameters such as acceleration and angular velocity using inertial measurement unit (IMU) data, it can accurately capture key dynamic features during a fall, such as sudden changes in posture and loss of balance. Simultaneously, audio data can record environmental acoustic signals such as collision sounds and cries for help during the fall. Therefore, combining audio data and inertial measurement data for abnormal fall recognition enables multimodal information complementarity and collaborative verification. Compared to a single data source, this combination significantly reduces the false alarm rate caused by sensor drift and daily motion interference, while also improving the detection rate and recognition accuracy of fall events.

[0018] Meanwhile, by identifying the current indoor scene and using scene information as a factor affecting the fall recognition results of the model, multimodal and scene-based accurate intelligent discrimination can be achieved. Compared with recognition schemes based on single data or without scene distinction, it can further optimize the scene adaptability of the recognition model, effectively distinguish the difference between fall events in various scenes and similar daily actions, and improve the accuracy and robustness of fall recognition in different scenes. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] Figure 1 This is a schematic diagram of the structure of a wearable device provided in an embodiment of this application; Figure 2 A schematic diagram of the architecture of a wearable device and a data processor provided for an embodiment of this application; Figure 3 A flowchart illustrating an abnormal fall detection method provided in an embodiment of this application; Figure 4 A schematic diagram of the architecture of a first fall recognition model provided in an embodiment of this application; Figure 5 A schematic diagram illustrating a fall detection process based on a first fall detection model, provided as an embodiment of this application; Figure 6 A flowchart illustrating a first fall recognition model training method provided in an embodiment of this application; Figure 7 A schematic diagram of the architecture of a second fall recognition model provided in an embodiment of this application; Figure 8 A schematic diagram illustrating a fall detection process based on a second fall detection model, provided as an embodiment of this application; Figure 9 A flowchart illustrating a second fall recognition model training method provided in an embodiment of this application; Figure 10 This is a schematic diagram of the structure of an abnormal fall detection device provided in an embodiment of this application; Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0021] Figure label: 100-Wearable device, 101-WiFi unit, 102-Audio acquisition unit, 103-Data processing unit, 104-Inertial measurement unit; 200 - Data Processing Server; 400 - First fall recognition model, 401 - First input layer, 402 - Convolutional feature encoder, 403 - Acoustic attention weight generator, 404 - Acoustic attention module, 405 - First classifier; 700 - Second fall recognition model, 701 - Second input layer, 702 - Temporal feature encoder, 703 - Motion attention weight generator, 704 - Motion attention module, 705 - Second classifier; 1000 - Abnormal fall recognition device; 1001 - Data acquisition module; 1002 - Scene recognition module; 1003 - First recognition module; 1004 - Second recognition module; 1005 - Recognition fusion module; 1100 - Electronic device, 1101 - Processor, 1102 - Memory.

[0022] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0023] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0024] With the accelerating aging of society and the rapid increase in the number of elderly people, falls have become one of the major health threats facing the elderly. Furthermore, against the backdrop of deepening societal aging, the phenomenon of elderly people living alone is becoming increasingly common. Once an elderly person falls indoors, they often struggle to get up on their own due to the lack of assistance, and their solitary status makes it difficult for their cries for help to be heard, making them difficult to be discovered after a fall. In such cases, the elderly may experience delayed treatment due to prolonged lack of assistance, leading to worsening injuries and even endangering their lives. Therefore, there is an urgent need for effective abnormal fall detection technology to improve this situation.

[0025] Existing fall detection technologies mainly fall into two categories: one type uses wearable devices with built-in sensors such as accelerometers and gyroscopes to monitor human movement in real time, and combines algorithms to analyze data such as acceleration and posture angles. When an instantaneous acceleration is detected and the posture angle changes beyond a set threshold, a fall is detected and an alarm is triggered. The other type uses cameras to collect image data and computer vision technology to extract changes in the position of human joints to determine a fall.

[0026] However, most existing fall recognition technologies rely on only single-dimensional information (such as motion information or image information) to make fall recognition judgments. Furthermore, due to the complexity of indoor environments, existing fall recognition technologies are also susceptible to environmental interference, leading to misjudgments and resulting in poor accuracy in recognizing abnormal fall behaviors.

[0027] To address the aforementioned technical issues, this application proposes an abnormal fall recognition method, electronic device, and program product. It identifies fall behavior simultaneously through audio data and inertial measurement data, and incorporates scene factors, thereby significantly improving the accuracy of abnormal fall recognition through multi-dimensional data.

[0028] The technical solutions of the abnormal fall recognition method, electronic device, and program product provided in this application will be described in detail below through specific embodiments. It should be noted that the following embodiments may exist independently or in combination with each other, and the same or similar content may not be described again in different embodiments.

[0029] It should be noted that the execution subject of the abnormal fall recognition method provided in this application embodiment is a wearable device, and correspondingly, the abnormal fall recognition device is also set in the wearable device.

[0030] See Figure 1 The wearable device 100 includes a WiFi unit 101, an audio acquisition unit 102, an inertial measurement unit 104, and a data processing unit 103.

[0031] Specifically, the WiFi unit 101 is a commonly built-in functional module in smart devices, capable of receiving and transmitting WiFi signals. Therefore, WiFi signal data of the wearable device 100 can be collected through the WiFi unit 101. This WiFi signal data includes signal strength (RSSI), channel state information (CSI), and time of arrival (ToA / TDoA), among other things.

[0032] Specifically, the audio acquisition unit 102 is used to acquire audio data based on a preset time window. The audio acquisition unit 102 is a microphone.

[0033] Specifically, the inertial measurement unit 104 is used to collect inertial measurement data based on a preset time window, including sensors such as accelerometers and gyroscopes. The inertial measurement data includes acceleration, angular velocity, and attitude angle. The fusion processing of multiple inertial measurement data can fully reflect the user's motion status and eliminate the error of single data, thereby improving the accuracy and reliability of subsequent fall detection.

[0034] Among them, acceleration reflects the linear acceleration of an object in three-dimensional space along the X, Y, and Z axes; angular velocity reflects the angular velocity of an object rotating around the X, Y, and Z axes; attitude angles reflect the attitude of an object in three-dimensional space, including pitch, roll, and yaw.

[0035] For example, the time window can be set to 2-5 seconds so that the audio data can include acoustic events before and after the fall, and the inertial measurement data can include motion events before and after the fall.

[0036] Specifically, the data processing unit 103 is used to process relevant data in order to identify and detect abnormal fall behavior.

[0037] See Figure 2 In addition to directly using the data processing unit 103 of the wearable device 100 to identify abnormal falls, the wearable device 100 can also be connected to the data processing server 200. After the wearable device 100 obtains the data, it transmits the data to the data processing server 200 for processing. After the data server 200 completes the processing, it returns the abnormal fall identification result to the wearable device.

[0038] Specifically, wearable devices 100 can be smartwatches, smart bracelets, smart belts, or other electronic devices such as smartphones that users carry for extended periods.

[0039] Specifically, the data processing server 200 can be a server cluster or a cloud server.

[0040] Figure 3This is a flowchart illustrating an abnormal fall detection method provided in an embodiment of this application. (See attached diagram.) Figure 3 In some embodiments, this abnormal fall detection method is applied to a wearable device, which includes a WiFi unit, an audio acquisition unit, and an inertial measurement unit. The specific process includes the following steps: S301, acquire the detection dataset; wherein, the detection dataset includes WiFi signal data collected by the WiFi unit, audio data collected by the audio acquisition unit, and inertial measurement data collected by the inertial measurement unit.

[0041] WiFi signal data is used to identify indoor scenes, while audio data and inertial measurement data are used to identify fall behavior.

[0042] Specifically, WiFi signal strength (RSSI) attenuates with increasing distance and is significantly affected by obstacles (such as walls and furniture). Furthermore, multipath environments at different locations (such as wall and furniture layouts) lead to variations in the channel state information (CSI) characteristics (such as amplitude distribution and phase pattern) of WiFi signals. Therefore, by measuring the RSSI and CSI values ​​in the WiFi signal data received by the wearable device, and inferring the indoor scene in which the wearable device is located based on the characteristics of the WiFi signal data, this study aims to determine the appropriate method for determining the location of the wearable device.

[0043] Specifically, when a person falls, the impact of the body with the ground produces a unique impact sound (such as the sound of a heavy object falling or the sound of limbs rubbing against a hard object). Its frequency, amplitude, and temporal characteristics (such as energy changes and duration) differ significantly from those of daily activities (such as walking or sitting). Furthermore, it may be accompanied by cries of surprise, groans, or calls for help. Therefore, audio data can be used to identify falls; by extracting relevant features from the audio data, it can be determined whether a fall has occurred.

[0044] Specifically, falls are usually accompanied by rapid, uncontrolled movements, contrasting with a slow sitting or squatting. For example, during a fall, the body's vertical acceleration suddenly increases and then approaches zero (when the body is at rest), while acceleration changes gradually during normal activity. Therefore, inertial measurement data can be used to identify falls by detecting changes in body acceleration, angular velocity, and posture angles.

[0045] S302 determines the indoor scene based on WiFi signal data.

[0046] When WiFi signals propagate indoors, they exhibit unique reflection, scattering, or attenuation patterns due to different spatial layouts, furniture distribution, and wall structures. Therefore, by utilizing the signal characteristics of WiFi signals in different indoor scenarios, it is possible to identify indoor scenes and effectively ensure the accuracy of scene identification.

[0047] S303: Input the indoor scene and audio data into the first fall recognition model to obtain the first fall probability output by the first fall recognition model.

[0048] In particular, the collected audio data will have different audio features for different indoor scenes. Therefore, by introducing indoor scene information into the first fall recognition model, the first fall recognition model can take into account the influence of scene factors on the fall recognition results, thereby improving the recognition accuracy of the first fall recognition model.

[0049] S304. Input the indoor scene and inertial measurement data into the second fall recognition model to obtain the second fall probability output by the second fall recognition model.

[0050] In particular, human movement characteristics will change for different indoor scenes. Therefore, by introducing indoor scene information into the second fall recognition model, the first fall recognition model can take into account the influence of scene factors on the fall recognition results, thereby improving the recognition accuracy of the second fall recognition model.

[0051] S305, based on the indoor scene, fuses the first fall probability and the second fall probability to obtain the fall recognition result.

[0052] The final fall recognition result is obtained by considering the fusion effect of different indoor scenarios on the probability of the first fall and the probability of the second fall. Specifically, the first and second fall probabilities are fused to obtain a fused probability, which is then compared with a pre-set probability threshold. If the fused probability is greater than the probability threshold, the fall is identified as a fall; if the fused probability is not greater than the probability threshold, the fall is identified as no fall.

[0053] In this embodiment, kinematic parameters such as acceleration and angular velocity collected by inertial measurement unit (IMU) can accurately capture key dynamic features during a fall, such as sudden changes in human posture and loss of balance. Meanwhile, audio data can simultaneously record environmental acoustic signals such as collision sounds and cries for help during the fall. Therefore, combining audio data and inertial measurement data for abnormal fall identification enables multimodal information complementarity and collaborative verification. Compared to a single data source, the combination not only significantly reduces the false alarm rate caused by sensor drift and interference from daily actions, but also improves the detection rate and recognition accuracy of fall events.

[0054] By identifying the current indoor scene (such as bathroom, hallway, living room) and using scene information as a factor influencing the model's fall recognition results, multimodal and scene-specific accurate intelligent discrimination can be achieved. Compared with recognition schemes based on single data or without scene distinction, it can further optimize the scene adaptability of the recognition model, effectively distinguish between fall events in various scenes and similar daily actions (such as bending over to wash in the bathroom, squatting down to pick up items in the kitchen, and sitting down to rest in the living room), and improve the accuracy and robustness of fall recognition in different scenes.

[0055] Optionally, after a fall is identified based on the fall detection results, the wearable device can call for help, such as automatically dialing medical emergency numbers or relatives' numbers, sending a text message containing the faller's location information to emergency contacts, or emitting an audible alarm accompanied by strong vibration through a built-in speaker, in order to provide assistance to the faller as soon as possible.

[0056] In some embodiments, determining an indoor scene based on WiFi signal data includes: inputting WiFi signal data into a scene recognition model to obtain the indoor scene output by the scene recognition model; wherein the scene recognition model is trained based on a third deep learning model.

[0057] In this embodiment, the deep learning model can learn the essential differences in WiFi signal data between different indoor scenes through data training, thereby achieving high accuracy and fast indoor scene recognition.

[0058] Optionally, the scene recognition model is obtained based on the following method: Step 1: Divide the indoor area into multiple indoor scenes and measure WiFi signal data at multiple locations within each indoor scene to establish a WiFi dataset; the WiFi dataset includes multiple WiFi signal data and the scene real labels corresponding to the WiFi signal data.

[0059] Since WiFi signal data differs in different indoor areas such as bathrooms and living rooms, multiple WiFi signal data points are collected in advance at different locations in these areas. Each WiFi signal data point is then labeled with a real-world scene tag representing the corresponding indoor scene.

[0060] Optionally, since the collected WiFi signal data is limited, data augmentation can be used to expand the dataset to improve the model's generalization ability and robustness.

[0061] Step 2: Input the WiFi dataset into the third deep learning model to obtain the scene prediction labels output by the third deep learning model.

[0062] Specifically, the WiFi dataset is input into the third deep learning model to obtain the scene prediction labels output by the third deep learning model, so as to use the WiFi dataset as the training set to iteratively train the third deep learning model.

[0063] Specifically, the third deep learning model can employ Convolutional Neural Network (CNN), Long Short-Term Memory Network (LSTM), or a hybrid CNN-LSTM model.

[0064] Step 3: Based on the real scene label and the predicted scene label, construct the third loss function of the third deep learning model; whereby the third loss function is used to characterize the difference between the real scene label and the predicted scene label.

[0065] The loss function quantifies the difference between the predicted results and the true labels, providing a clear direction for optimizing model parameters. For example, a larger loss value indicates that the model prediction is less accurate, and the optimization algorithm will adjust the model parameters (such as weights and biases) accordingly to gradually reduce the loss; conversely, a smaller loss value indicates that the model prediction is closer to the reality.

[0066] Specifically, the third loss function can be the cross-entropy loss function, the log loss function, etc.

[0067] Step 4: Iteratively train the third deep learning model based on the third loss function to obtain the scene recognition model.

[0068] Specifically, the third deep learning model is iteratively trained using a third loss function to optimize and adjust its parameters until the third loss function converges or the set number of training iterations is reached.

[0069] In this embodiment, a scene recognition model capable of quickly recognizing indoor scenes is obtained by iteratively training a deep learning model.

[0070] In some embodiments, the first fall probability and the second fall probability are fused based on the indoor scene to obtain a fall recognition result, including: Step 1: Input the indoor scene into the weight mapping table to obtain the acoustic recognition weight and motion recognition weight corresponding to the indoor scene; wherein, the weight mapping table is used to indicate the acoustic recognition weight and motion recognition weight corresponding to different indoor scenes.

[0071] Specifically, a weight mapping table indicates the acoustic recognition weight and motion recognition weight corresponding to different indoor scenes, thereby dynamically adjusting the recognition weight according to different indoor scenes, so that the fall recognition result can adapt to the current indoor scene and improve the fall recognition accuracy.

[0072] Specifically, the reliability of the first and second fall recognition models differs in different indoor scenarios. For example, in a bathroom environment, where there is often the sound of water, the second fall recognition model is more reliable, and thus, the second fall recognition model is given a higher weight.

[0073] Specifically, the weight mapping table can be obtained using the following method: 1. Assume the acoustic recognition weights are... The motion recognition weight is First, for each indoor scene, define a weighted search space, for example, The value ranges from 0.0 to 1.0, with a step size of 0.1. Because... ,so .

[0074] 2. For each candidate weight Using each sample in the validation set corresponding to this indoor scene, the first fall probability and the second fall probability are obtained respectively, and then the fused probability is calculated based on the subsequent weights:

[0075] in, This indicates the probability of fall detection. This indicates the probability of the first fall. This indicates the probability of falling down second.

[0076] according to The prediction result is obtained by using a pre-set probability threshold (e.g., 0.8). Among them, if >0.8, then =1 indicates that a fall was detected; if <0.8, then =0 indicates that the fall behavior could not be detected.

[0077] 3. Calculate the performance metric for the entire validation set under the current weights, typically the F1 score. The F1 score evaluates the classification model's performance by combining the model's precision and recall.

[0078] 4. After traversing all candidate weights, a weight-performance list is obtained. Select the weight that maximizes the F1 score. and As the optimal weight for the current indoor scene.

[0079] 5. Repeat the above process for all other indoor scenes to find the optimal weight for each indoor scene and generate a weight mapping table.

[0080] Step 2: Perform weighted fusion based on acoustic recognition weights, the first fall probability, motion recognition weights, and the second fall probability to obtain the fall recognition result.

[0081] Specifically, after determining the weights corresponding to the first and second fall probabilities through the indoor scene, the first fall probability and the second fall recognition are weighted and fused by acoustic recognition weight and motion recognition weight to obtain the fall recognition result.

[0082] Specifically, the fall detection probability satisfies the following formula:

[0083] In this embodiment, by determining the weights corresponding to the first and second fall probabilities based on the indoor scene, a greater weight is given to the more reliable fall recognition model in different indoor scenes, thereby improving the accuracy of the fall recognition results.

[0084] exist Figure 3 The abnormal fall detection method shown requires obtaining the first fall probability based on the first fall detection model. The following section combines... Figure 4 and Figure 5 The technical solution of the above-mentioned abnormal fall recognition method further introduces the content of obtaining the first fall probability based on the first fall recognition model.

[0085] See Figure 4 Optionally, the first fall recognition model 400 includes a first input layer 401, a convolutional feature encoder 402, an acoustic attention weight generator 403, an acoustic attention module 404, and a first classifier 405.

[0086] The first input layer 401 is used to input data and preprocess the data.

[0087] The convolutional feature encoder 402 is used to perform convolution processing on the input data to extract deep acoustic features of the data.

[0088] Among them, the acoustic attention weight generator 403 dynamically adjusts the attention weight by combining scene information.

[0089] The acoustic attention module 404 weights acoustic features based on acoustic attention weights. Taking a bathroom scene as an example, the attention module teaches the model to pay more attention to impact sounds, groans, etc., while ignoring the sound of flowing water.

[0090] The first classifier 405 is used to output the classification and recognition result based on the weighted features.

[0091] Figure 5A schematic diagram illustrating a fall detection process based on a first fall detection model, provided in this application embodiment, is shown below. Figure 5 In some embodiments, the fall detection process based on the first fall detection model includes the following steps: S501, based on the first input layer, preprocess the audio data to obtain the Mel spectrogram corresponding to the audio data.

[0092] Mel spectrograms can convert audio signals into two-dimensional image representations to visually display the energy distribution of sound at different times and frequencies, thus facilitating subsequent convolutional feature encoders to learn the relevant acoustic features of audio data from the Mel spectrograms.

[0093] S502 uses a convolutional feature encoder to perform convolutional encoding on the Mel spectrogram to obtain acoustic features.

[0094] Among them, by leveraging the local perception and hierarchical feature extraction capabilities of the Convolutional Feature Encoder (CNN), the time-frequency information in the Mel spectrogram is transformed into a more discriminative acoustic feature representation, thereby improving the performance of tasks such as speech recognition and audio classification.

[0095] S503 uses an acoustic attention weight generator to calculate attention weights based on indoor scenes, thus obtaining acoustic attention weights.

[0096] The acoustic attention weight generator generates corresponding acoustic attention weights based on the indoor scene. These weights are used to adjust the distribution of acoustic features output by the convolutional feature encoder, making them more suitable for the indoor scene context. For example, in a bathroom scene, features similar to water sound frequencies are suppressed, while features related to impact sounds are enhanced.

[0097] Specifically, the acoustic attention weight generator generates corresponding attention weights based on the indoor scene and adopts a hard-coded weight rule calculation method for attention weights. That is, different acoustic attention weight templates are set for different indoor scenes and can be switched directly according to the scene type.

[0098] Specifically, the attention weight is calculated by assigning a weight value between 0 and 1 to each feature by calculating the correlation between different features and the task objective. The higher the weight, the more important the feature is to the current task.

[0099] S504, based on the acoustic attention module, weights the acoustic features according to the acoustic attention weights to obtain acoustic weighted features.

[0100] In this model, acoustic features are weighted using acoustic attention weights. Important features are given higher weights to enhance their influence, while secondary features are weakened or even ignored. This mechanism allows the model to adaptively focus on key features, dynamically targeting crucial information, suppressing redundant interference, and improving model performance. For example, in a bathroom scenario, the model finds that the time steps of a "thump" sound and a groan are highly correlated with fall detection, and therefore assigns them high weights. Meanwhile, the sound of water is irrelevant to fall detection and receives lower weights.

[0101] Meanwhile, during model training, the weighting process can effectively filter noise and irrelevant information, reduce the interference of redundant features on model training, and enhance the robustness and generalization ability of the model.

[0102] S505, based on the first classifier, classifies the acoustic weighted features to obtain the first fall probability.

[0103] Among them, the acoustic weighted features after scene modulation and focusing are fed into the first classifier to obtain the first fall probability output by the first classifier.

[0104] In this embodiment, the first fall recognition model significantly improves the robustness, generalization ability and interpretability of the model by setting scene-specific acoustic attention weights for different scenarios.

[0105] In some embodiments, before obtaining the first fall probability based on the first fall recognition model, it is necessary to iteratively train the first deep learning model to obtain the first fall recognition model.

[0106] Figure 6 A flowchart illustrating a first fall recognition model training method provided in this application embodiment is shown below. Figure 6 In some embodiments, the training method for the first fall recognition model includes the following steps: S601, Obtain the first training set; wherein, the first training set includes multiple audio training samples, as well as the first real fall label and the first scene real label corresponding to the audio training samples.

[0107] The audio training samples include fall-related sounds (impact sounds, shouts, sounds of a body falling to the ground) and background ambient sounds of different indoor scenes, or simply background ambient sounds. These audio training samples are labeled with corresponding real fall labels (fall or non-fall) and scene labels (e.g., bathroom, hallway, living room, bedroom, etc.).

[0108] S602, input the first training set into the pre-built first deep learning model to obtain the first predicted fall label output by the first deep learning model.

[0109] When the first training set is input into the first deep learning model, the first deep learning model will output the first predicted fall label, which is the recognition and prediction result of the first deep learning model for the fall behavior.

[0110] S603, based on the first predicted fall label and the first actual fall label, construct a first loss function; wherein, the first loss function includes a focus loss function.

[0111] In fall detection, the number of non-fall samples far exceeds the number of fall samples. This leads the model to tend to predict all samples as "non-fall" to reduce overall loss, as this works well with a large number of negative samples. The focus loss function, on the other hand, reduces the weight of easily classified samples (non-fall samples), forcing the model to focus on harder-to-classify samples (fall samples or non-fall samples resembling falls), thereby improving the model's accuracy in fall detection.

[0112] Specifically, the first loss function satisfies the following formula:

[0113] in, This represents the model's predicted probability that a sample belongs to a certain class; its definition varies depending on the sample's true class. When the sample belongs to the positive class (That is, the probability of the model predicting a fall); When the sample belongs to the negative class (That is, the probability that the model predicts that the person will not fall).

[0114] This represents the class weight parameter, used to balance the importance of positive and negative samples. Its definition is as follows: When the sample belongs to the positive class ( This is a hyperparameter, usually set to a value less than 1 to reduce the weight of positive samples, since the number of positive samples is typically small. When the sample belongs to the negative class (To increase the weight of negative samples, but usually due to the large number of negative samples, their actual weight may still be low).

[0115] This represents the modulation factor parameter, used to control the rate at which the importance of easily classified samples decays. Greater than or equal to 0, and The larger the value, the greater the impact of difficult-to-classify samples.

[0116] By introducing category weight parameters The focus loss function balances the importance of positive and negative samples, preventing the model from being overly biased towards the class with the larger number of samples during training. This is achieved by introducing a modulation factor parameter. The focus loss function can reduce the loss contribution of easily classified samples, allowing the model to focus its training on difficult-to-classify samples, thereby improving the model's generalization ability.

[0117] S604, the first deep learning model is iteratively trained based on the first loss function to obtain the first fall recognition model.

[0118] Specifically, the first deep learning model is iteratively trained using a first loss function to optimize and adjust the parameters of the first deep learning model until the first loss function converges or the preset number of training iterations is reached, thus obtaining the first fall recognition model.

[0119] In this embodiment, scene information is used as additional input information, enabling the first fall recognition model to learn fall patterns in different scenes and achieve accurate recognition of fall behavior in different scenes.

[0120] exist Figure 3 The abnormal fall detection method shown requires obtaining the second fall probability based on the second fall detection model. The following section combines... Figure 7 and Figure 8 The technical solution of the above-mentioned abnormal fall recognition method further introduces the content on obtaining the second fall probability based on the second fall recognition model.

[0121] See Figure 7 Optionally, the first fall recognition model 700 includes a second input layer 701, a temporal feature encoder 702, a motion attention weight generator 703, a motion attention module 704, and a second classifier 705.

[0122] The second input layer 701 is used to input data and preprocess the data.

[0123] The temporal feature encoder 702 is used to perform temporal feature encoding on the input data in order to extract deep motion features from the data.

[0124] Among them, the motion attention weight generator 703 dynamically adjusts the attention weight by combining scene information.

[0125] The motion attention module 704 weights motion features based on motion attention weights. Taking a bathroom scene as an example, the attention module helps the model learn to be more sensitive to speed changes and less sensitive to changes in angular velocity.

[0126] The second classifier 705 is used to output the classification and recognition results based on the weighted features.

[0127] Figure 8 A schematic diagram illustrating a fall detection process based on a second fall detection model, provided in this application embodiment, is shown below. Figure 8 In some embodiments, the fall detection process based on the second fall detection model includes the following steps: S801 preprocesses the inertial measurement data based on the second input layer to obtain the time series features corresponding to the inertial measurement data.

[0128] Specifically, time series features include dynamic features such as the rate of change of acceleration, the rate of change of angular velocity, and the rate of change of attitude angle.

[0129] S802 uses a time-series feature encoder to perform time-series encoding on time-series features to obtain motion features.

[0130] In this process, a temporal feature encoder (LSTM) is used to encode time-series features into motion features that reflect human movement. First, this reduces data dimensionality. Original time-series feature data is often large and contains a lot of redundant information; encoding it into motion features allows for the extraction of the most critical motion information, reducing the burden of data storage and processing, and improving computational efficiency. Second, it enhances the robustness of the features. Motion features are more resistant to noise and interference. When faced with fluctuations in time-series features caused by sensor errors, environmental interference, etc., motion features can remain relatively stable, more accurately reflecting the true motion state of objects. Finally, motion features provide higher-quality and more representative input for subsequent fall recognition tasks, helping to improve the accuracy and reliability of these fall recognition methods. S803 uses a motion attention weight generator to calculate attention weights based on indoor scenes to obtain motion attention weights.

[0131] The motion attention weight generator generates corresponding motion attention weights based on the indoor scene. These weights are used to adjust the distribution of motion features output by the temporal feature encoder, making them more suitable for the indoor scene context. For example, in a bathroom scene, where people are more likely to slip, features related to angular velocity are suppressed, while features related to horizontal and vertical acceleration are enhanced.

[0132] Specifically, the motion attention weight generator generates corresponding attention weights based on the indoor scene and adopts a hard-coded weight rule calculation method for attention weights. That is, different motion attention weight templates are set for different indoor scenes and can be switched directly according to the scene type.

[0133] S804, based on the motion attention module, weights motion features according to motion attention weights to obtain motion-weighted features.

[0134] This mechanism utilizes motion attention weights to weight motion features, assigning higher weights to important features to enhance their influence, while weakening or even ignoring secondary features. This allows the model to adaptively focus on key features, dynamically prioritizing crucial information, suppressing redundant interference, and improving model performance. For example, in a bathroom scenario, the model found that changes in acceleration were highly correlated with fall detection, thus assigning them high weights. Changes in angular velocity, on the other hand, had a lower correlation with fall detection and therefore received lower weights.

[0135] Meanwhile, during model training, the weighting process can effectively filter noise and irrelevant information, reduce the interference of redundant features on model training, and enhance the robustness and generalization ability of the model.

[0136] S805 classifies motion-weighted features based on a second classifier to obtain a second fall probability.

[0137] Among them, the motion-weighted features after scene modulation and focusing are fed into the second classifier to obtain the second fall probability output by the second classifier.

[0138] In this embodiment, the second fall recognition model significantly improves the robustness, generalization ability and interpretability of the model by setting scene-specific motion attention weights for different scenarios.

[0139] In some embodiments, before obtaining the first fall probability based on the first fall recognition model, it is necessary to iteratively train the first deep learning model to obtain the first fall recognition model.

[0140] Figure 9 A flowchart illustrating a first fall recognition model training method provided in this application embodiment is shown below. Figure 9 In some embodiments, the training method for the first fall recognition model includes the following steps: S901, Obtain the second training set; wherein, the second training set includes multiple inertial measurement training samples, and the second real fall label and the second scene real label corresponding to the inertial measurement training samples.

[0141] The inertial measurement training samples include inertial measurement data (acceleration, angular velocity, and attitude angle) for different types of falls (forward, sideways, and backward) and non-falls. These inertial measurement training samples are labeled with corresponding real fall labels (fall or non-fall) and scene labels (e.g., bathroom, hallway, living room, bedroom, etc.).

[0142] S902, input the second training set into the pre-built second deep learning model to obtain the second predicted fall label output by the second deep learning model.

[0143] When the second training set is input into the second deep learning model, the second deep learning model will output the second predicted fall label, which is the recognition and prediction result of the fall behavior by the second deep learning model.

[0144] S903, based on the second predicted fall label and the second actual fall label, constructs a second loss function; wherein, the second loss function includes a weighted binary cross-entropy loss function and a temporal consistency loss function.

[0145] This approach employs a combined loss function of weighted binary cross-entropy loss and temporal consistency loss to balance classification accuracy in imbalanced scenarios with the dynamic constraints of temporal data. The weighted binary cross-entropy loss, by assigning differentiated weights to fallen and non-fallen samples, effectively mitigates model bias caused by the scarcity of fall samples in practical applications, improves the model's sensitivity to small-scale fall events, and avoids missed detections due to majority-class dominance in training. The temporal consistency loss, targeting the time-series attributes of inertial measurement data, forces the model to learn the dynamic correlation patterns between adjacent time points, constraining the smoothness and rationality of prediction results in the temporal dimension, reducing the risk of false detections caused by sensor noise or instantaneous abnormal data, and ensuring stable and reliable recognition results in continuous monitoring scenarios. Therefore, the combined effect of these two loss functions ensures both accurate discrimination of positive and negative samples and enhances the mining of inherent patterns in temporal data, ultimately achieving a simultaneous improvement in accuracy, recall, and stability in fall recognition tasks.

[0146] Specifically, the second loss function satisfies the following formula:

[0147] in, This represents the second loss function. This represents the weighted binary cross-entropy loss function. Represents the time-series consistency loss function. and They represent the weights, .

[0148] Weighted binary cross-entropy loss means the same thing as standard BCE, but different weights are assigned to samples of different classes. The formula is usually as follows:

[0149] Where N represents the total number of samples, and i represents the i-th sample. This represents the true label of the i-th sample. This represents the probability that the model predicts the i-th sample as positive (label 1), with a value range of [0,1]. The weight of the positive sample is... This is used to increase the contribution of positive samples in loss calculation, especially when the number of positive samples is small. It should be greater than 1. This represents the weight of negative samples, used to increase the contribution of negative samples in the loss calculation. When the number of negative samples is small... It should be greater than 1.

[0150] The core advantage of the weighted binary cross-entropy (WBCE) loss function is that it specifically addresses the sample imbalance problem in binary classification tasks. By assigning differentiated weights to positive and negative samples, it amplifies the contribution of minority class samples in the loss calculation and avoids the model from tending to predict the majority class due to the dominance of majority class samples.

[0151] It is known that for a real fall event, the movement pattern before and after the fall should be continuous, and the model's prediction for the continuous window should also be smooth. Therefore, the temporal consistency loss utilizes the continuity of time series data and satisfies the following formula:

[0152] in, It represents the mean square error between the predicted probabilities p of two consecutive time windows t and t+1.

[0153] Temporal consistency loss characterizes the stability of the model's predictions. Without this loss, the model might output wildly fluctuating predictions (e.g., "fall-not fall-fall-not fall") for continuous, subtly varying segments of motion data, which is clearly unreasonable. This loss function penalizes such flickering predictions, encouraging the model to make consistent judgments for continuous motion segments.

[0154] S904, the second deep learning model is iteratively trained based on the second loss function to obtain the second fall recognition model.

[0155] Specifically, the first deep learning model is iteratively trained using a first loss function to optimize and adjust the parameters of the first deep learning model until the first loss function converges or the preset number of training iterations is reached, thus obtaining the first fall recognition model.

[0156] In this embodiment, scene information is used as additional input information, enabling the second fall recognition model to learn fall patterns in different scenes and achieve accurate recognition of fall behavior in different scenes.

[0157] Figure 10 This is a schematic diagram of the structure of an abnormal fall detection device provided in an embodiment of this application. (See attached diagram.) Figure 10 The abnormal fall detection device includes various functional modules for implementing the aforementioned abnormal fall detection method, and any functional module can be implemented by software and / or hardware.

[0158] In some embodiments, the abnormal fall detection device 1000 is applied to a wearable device, which includes a WiFi unit, an audio acquisition unit, and an inertial measurement unit. The abnormal fall detection device 1000 includes a data acquisition module 1001, a scene recognition module 1002, a first recognition module 1003, a second recognition module 1004, and a recognition fusion module 1005. Wherein: The data acquisition module 1001 is used to acquire the detection dataset; wherein, the detection dataset includes WiFi signal data collected by the WiFi unit, audio data collected by the audio acquisition unit, and inertial measurement data collected by the inertial measurement unit; The scene recognition module 1002 is used to determine the indoor scene based on WiFi signal data; The first recognition module 1003 is used to input indoor scene and audio data into the first fall recognition model to obtain the first fall probability output by the first fall recognition model; The second recognition module 1004 is used to input indoor scene and inertial measurement data into the second fall recognition model to obtain the second fall probability output by the second fall recognition model. The recognition fusion module 1005 is used to fuse the first fall probability and the second fall probability based on the indoor scene to obtain the fall recognition result.

[0159] In some embodiments, the first fall recognition model includes a first input layer, a convolutional feature encoder, an acoustic attention weight generator, an acoustic attention module, and a first classifier. Indoor scene and audio data are input into the first fall recognition model. The first recognition module 1003 is specifically used for: The audio data is preprocessed based on the first input layer to obtain the Mel spectrogram corresponding to the audio data; Acoustic features are obtained by performing convolutional encoding on the Mel spectrogram based on a convolutional feature encoder. Acoustic attention weights are obtained by calculating attention weights based on indoor scenes using an acoustic attention weight generator. The acoustic attention module uses acoustic attention weights to weight acoustic features, thus obtaining acoustic weighted features. The acoustic weighted features are classified based on the first classifier to obtain the first probability of falling.

[0160] In some embodiments, the first identification module 1003 is further configured to: Obtain the first training set; wherein, the first training set includes multiple audio training samples, as well as the first real fall label and the first real scene label corresponding to the audio training samples; Input the first training set into the pre-built first deep learning model to obtain the first predicted fall label output by the first deep learning model; Based on the first predicted fall label and the first actual fall label, a first loss function is constructed; wherein, the first loss function includes the focus loss function; The first deep learning model is iteratively trained based on the first loss function to obtain the first fall recognition model.

[0161] In some embodiments, the second fall recognition model includes a second input layer, a temporal feature encoder, a motion attention weight generator, a motion attention module, and a second classifier. The second recognition module 1004 is specifically used for: The inertial measurement data is preprocessed based on the second input layer to obtain the time series features corresponding to the inertial measurement data; Motion features are obtained by temporally encoding time-series features using a temporal feature encoder. Based on a motion attention weight generator, attention weights are calculated for indoor scenes to obtain motion attention weights; Motion features are obtained by weighting motion features based on motion attention weights using a motion attention module. The second fall probability is obtained by classifying the motion-weighted features based on the second classifier.

[0162] In some embodiments, the second identification module 1004 is further configured to: Obtain the second training set; wherein the second training set includes multiple inertial measurement training samples, and the second real fall label and the second real scene label corresponding to the inertial measurement training samples; Input the second training set into the pre-built second deep learning model to obtain the second predicted fall label output by the second deep learning model; A second loss function is constructed based on the second predicted fall label and the second actual fall label; wherein, the second loss function includes a weighted binary cross-entropy loss function and a temporal consistency loss function; The second deep learning model is iteratively trained based on the second loss function to obtain the second fall recognition model.

[0163] In some embodiments, the scene recognition module 1002 is specifically used for: WiFi signal data is input into the scene recognition model to obtain the indoor scene output by the scene recognition model; the scene recognition model is trained based on a third deep learning model.

[0164] In some embodiments, the scene recognition module 1002 is further configured to: The indoor area is divided into multiple indoor scenes, and WiFi signal data at multiple locations within each indoor scene are measured to establish a WiFi dataset. The WiFi dataset includes multiple WiFi signal data and the scene real labels corresponding to the WiFi signal data. Input the WiFi dataset into the third deep learning model to obtain the scene prediction labels output by the third deep learning model; Based on the real scene labels and the predicted scene labels, a third loss function is constructed for the third deep learning model; the third loss function is used to characterize the difference between the real scene labels and the predicted scene labels. The third deep learning model is iteratively trained based on the third loss function to obtain the scene recognition model.

[0165] In some embodiments, the identification fusion module 1005 is specifically used for: The indoor scene is input into the weight mapping table to obtain the acoustic recognition weight and motion recognition weight corresponding to the indoor scene; the weight mapping table is used to indicate the acoustic recognition weight and motion recognition weight corresponding to different indoor scenes; The fall recognition result is obtained by weighting and fusing the acoustic recognition weight, the first fall probability, the motion recognition weight, and the second fall probability.

[0166] The abnormal fall recognition device 1000 provided in this application embodiment is used to execute the technical solution provided in the aforementioned abnormal fall recognition method embodiment. Its implementation principle and technical effect are similar to those in the aforementioned method embodiment, and will not be repeated here.

[0167] It should be noted that the division of the various modules in the above device is merely a logical functional division. In actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. Furthermore, these modules can be implemented entirely in software via processing element calls, entirely in hardware, or partially in software calls via processing element calls, with some modules implemented in hardware. For example, the identification fusion module 1005 can be a separate processing element, or it can be integrated into a chip within the above device. Alternatively, it can be stored as program code in the device's memory, and its functions can be called and executed by a processing element within the device. The implementation of other modules is similar. Moreover, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0168] Figure 11This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. (See attached diagram.) Figure 11 The electronic device 1100 includes a processor 1101 and a memory 1102 communicatively connected to the processor 1101; Memory 1102 stores computer-executed instructions; The processor 1101 executes computer execution instructions stored in the memory 1102 to implement the aforementioned technical solution of the abnormal fall recognition method.

[0169] In the aforementioned electronic device 1100, the memory 1102 and the processor 1101 are electrically connected directly or indirectly to achieve data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses or signal lines, such as bus connections. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be classified as address buses, data buses, control buses, etc., but this does not mean that there is only one bus or one type of bus. The memory 1102 stores computer execution instructions that implement the aforementioned abnormal fall detection method, including at least one software functional module that can be stored in the memory 1102 in the form of software or firmware. The processor 1101 executes various functional applications and data processing by running the software programs and modules stored in the memory 1102.

[0170] The memory 1102 includes at least one type of readable storage medium, not limited to Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory 1102 stores programs, and the processor 1101 executes the programs after receiving execution instructions. Furthermore, the software programs and modules within the memory 1102 may also include an operating system, which may include various software components and / or drivers for managing system tasks (e.g., memory management, storage device control, power management, etc.) and can communicate with various hardware or software components to provide an operating environment for other software components.

[0171] Processor 1101 can be an integrated circuit chip with signal processing capabilities. The aforementioned processor 1101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), etc. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor, or processor 1101 can be any conventional processor.

[0172] The electronic device 1100 is used to execute the technical solution provided in the aforementioned abnormal fall recognition method embodiment. Its implementation principle and technical effect are similar to those in the aforementioned method embodiment, and will not be repeated here.

[0173] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the technical solution of the aforementioned abnormal fall recognition method.

[0174] The aforementioned computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The computer-readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0175] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Alternatively, the readable storage medium can be an integral part of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components within the control unit of an abnormal fall detection device.

[0176] This application also provides a computer program product, including a computer program, which, when executed, is used to implement the technical solution of the aforementioned abnormal fall recognition method.

[0177] In the above embodiments, those skilled in the art will understand that the above method embodiments can be implemented entirely or partially by software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented entirely or partially in the form of a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless network, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0178] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0179] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the appended claims.

[0180] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for identifying abnormal falls, characterized in that, Applied to wearable devices, the wearable devices including a WiFi unit, an audio acquisition unit, and an inertial measurement unit, the method includes: Obtain a detection dataset; wherein the detection dataset includes WiFi signal data collected by the WiFi unit, audio data collected by the audio acquisition unit, and inertial measurement data collected by the inertial measurement unit; Based on the WiFi signal data, the indoor scene is determined; The indoor scene and the audio data are input into the first fall recognition model to obtain the first fall probability output by the first fall recognition model; The indoor scene and the inertial measurement data are input into the second fall recognition model to obtain the second fall probability output by the second fall recognition model; The first fall probability and the second fall probability are fused based on the indoor scene to obtain the fall recognition result.

2. The method according to claim 1, characterized in that, The first fall detection model includes a first input layer, a convolutional feature encoder, an acoustic attention weight generator, an acoustic attention module, and a first classifier. The indoor scene and the audio data are input into the first fall detection model to obtain the first fall probability output by the first fall detection model, including: Based on the first input layer, the audio data is preprocessed to obtain the Mel spectrogram corresponding to the audio data; The acoustic features are obtained by performing convolutional encoding on the Mel spectrogram based on the convolutional feature encoder. Based on the acoustic attention weight generator, attention weights are calculated based on the indoor scene to obtain acoustic attention weights; The acoustic attention module weights the acoustic features based on the acoustic attention weights to obtain acoustic weighted features; The acoustic weighted features are classified based on the first classifier to obtain the first fall probability.

3. The method according to claim 2, characterized in that, The first fall detection model was obtained based on the following method: Obtain a first training set; wherein the first training set includes multiple audio training samples, and a first real fall label and a first scene real label corresponding to the audio training samples; Input the first training set into a pre-built first deep learning model to obtain the first predicted fall label output by the first deep learning model; Based on the first predicted fall label and the first actual fall label, a first loss function is constructed; wherein, the first loss function includes a focus loss function; The first deep learning model is iteratively trained based on the first loss function to obtain the first fall recognition model.

4. The method according to claim 1, characterized in that, The second fall detection model includes a second input layer, a temporal feature encoder, a motion attention weight generator, a motion attention module, and a second classifier. The indoor scene and the inertial measurement data are input into the second fall detection model to obtain the second fall probability output by the second fall detection model, including: The inertial measurement data is preprocessed based on the second input layer to obtain the time series features corresponding to the inertial measurement data; Based on the temporal feature encoder, the time series features are temporally encoded to obtain motion features; Based on the motion attention weight generator, attention weights are calculated based on the indoor scene to obtain motion attention weights; The motion attention module weights the motion features based on the motion attention weights to obtain motion-weighted features; The motion-weighted features are classified based on the second classifier to obtain the second fall probability.

5. The method according to claim 4, characterized in that, The second fall recognition model is obtained based on the following method: Obtain a second training set; wherein the second training set includes multiple inertial measurement training samples, and a second real fall label and a second real scene label corresponding to the inertial measurement training samples; The second training set is input into the pre-built second deep learning model to obtain the second predicted fall label output by the second deep learning model; Based on the second predicted fall label and the second actual fall label, a second loss function is constructed; wherein, the second loss function includes a weighted binary cross-entropy loss function and a temporal consistency loss function; The second deep learning model is iteratively trained based on the second loss function to obtain the second fall recognition model.

6. The method according to any one of claims 1-5, characterized in that, Based on the WiFi signal data, the indoor scene is determined, including: The WiFi signal data is input into the scene recognition model to obtain the indoor scene output by the scene recognition model; wherein, the scene recognition model is trained based on a third deep learning model.

7. The method according to claim 6, characterized in that, The scene recognition model is obtained based on the following method: The indoor area is divided into multiple indoor scenes, and WiFi signal data at multiple locations within each indoor scene are measured to establish a WiFi dataset; wherein, the WiFi dataset includes multiple WiFi signal data and the scene real labels corresponding to the WiFi signal data; Input the WiFi dataset into the third deep learning model to obtain the scene prediction label output by the third deep learning model; Based on the scene's real labels and the scene's predicted labels, a third loss function is constructed for the third deep learning model; wherein, the third loss function is used to characterize the difference between the scene's real labels and the scene's predicted labels; The scene recognition model is obtained by iteratively training the third deep learning model based on the third loss function.

8. The method according to any one of claims 1-5, characterized in that, Based on the indoor scene, the first fall probability and the second fall probability are fused to obtain a fall recognition result, including: The indoor scene is input into a weight mapping table to obtain the acoustic recognition weight and motion recognition weight corresponding to the indoor scene; wherein, the weight mapping table is used to indicate the acoustic recognition weight and motion recognition weight corresponding to different indoor scenes; The fall recognition result is obtained by weighting and fusing the acoustic recognition weights and the first fall probability, the motion recognition weights and the second fall probability.

9. An electronic device, characterized in that, Includes a processor and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 8.

10. A computer program product, characterized in that, Includes a computer program, which, when executed, is used to implement the method described in any one of 1 to 8.

Images