Fall monitoring model training and fall monitoring method, apparatus, device, and medium

Abstract

Embodiments of the present application disclose a fall monitoring model training method, a fall monitoring method, a fall monitoring device, a fall monitoring equipment and a fall monitoring medium. The fall monitoring model training method comprises: obtaining fall monitoring data, wherein each fall monitoring data is provided with a corresponding fall action label; performing frequency domain filtering and time domain filtering on each piece of fall monitoring data to obtain frequency domain filtering data and time domain filtering data; taking the frequency domain filtering data, the time domain filtering data and the corresponding fall action label as a model training sample, training a preset fall monitoring model, and obtaining a target fall monitoring model. The technical solution of the present embodiment solves the problems of poor effect of the time domain filtering method alone and low efficiency of data conversion in the frequency domain filtering process in the prior art, realizes the fusion of time domain data and frequency domain data, makes the monitoring data fully filtered, improves the monitoring data processing efficiency, and improves the monitoring accuracy of the fall monitoring model.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a fall detection model training method, apparatus, device and medium for fall detection. Background Technology

[0002] Falls commonly cause direct or indirect harm to the elderly and are a frequent cause of accidental injury. In elderly care, timely detection and intervention of falls minimize potential harm. Among various fall detection methods, ultra-wideband radar signals offer advantages such as good privacy protection and ease of deployment in human posture recognition, and are being widely used in IoT technology and smart homes.

[0003] In fall detection using ultra-wideband radar, radar signal filtering is typically performed using methods such as time-domain filtering or frequency-domain filtering. However, some time-domain filtering methods may result in insufficient suppression of clutter and slow-moving targets, or blind speed or blind phase phenomena, which weaken the radar's ability to detect moving targets. While frequency-domain filtering can effectively filter clutter and environmental echoes, the filtering process involves inverse transformations of frequency and time domain data, increasing time and memory overhead. Summary of the Invention

[0004] This invention provides a fall monitoring model training method, apparatus, equipment, and medium for fall monitoring, which enables sufficient filtering of monitoring data, reduces the memory required for computation, and improves the efficiency of monitoring data processing and the monitoring accuracy of the fall monitoring model.

[0005] In a first aspect, embodiments of the present invention provide a method for training a fall detection model, the method comprising:

[0006] Acquire fall monitoring data, wherein each of the fall monitoring data is assigned a corresponding fall action label;

[0007] For each of the aforementioned fall monitoring data, frequency domain filtering and time domain filtering are performed respectively to obtain frequency domain filtered data and time domain filtered data;

[0008] The frequency domain filtered data, the time domain filtered data, and the corresponding fall action labels are used as model training samples to train the preset fall detection model and obtain the target fall detection model.

[0009] Optionally, frequency domain filtering is performed on the fall monitoring data to obtain frequency domain filtered data, including:

[0010] Perform a Fast Fourier Transform on the fall monitoring data to obtain frequency domain fall monitoring data;

[0011] The frequency domain filtered data is obtained by filtering out data whose frequency is lower than a preset lower frequency limit from the frequency domain fall monitoring data.

[0012] Optionally, the fall monitoring data is subjected to time-domain filtering to obtain time-domain filtered data, including:

[0013] Singular value decomposition was performed on the fall monitoring data to obtain data feature values;

[0014] The fall monitoring data is reconstructed by selecting feature values ​​within a preset range to obtain the time-domain filtered data.

[0015] Optionally, feature values ​​within a preset range corresponding to fall monitoring data are selected for data reconstruction to obtain the time-domain filtered data, including:

[0016] Sort the feature values ​​of the data obtained from singular value decomposition;

[0017] Fall monitoring data corresponding to the data feature values ​​within a preset sorting range are selected from the sorted data feature values ​​as the data in the time-domain filtered data, and the fall monitoring data corresponding to the feature values ​​outside the preset sorting range are replaced with zero values ​​to obtain the time-domain filtered data.

[0018] Optionally, before training the preset fall detection model, the method further includes:

[0019] The frequency domain filtered data and the time domain filtered data are compressed.

[0020] The compressed frequency domain filtered data and the compressed time domain filtered data, along with the corresponding fall action labels, are used as training samples for the model.

[0021] Optionally, the data compression of the frequency domain filtered data and the time domain filtered data includes:

[0022] Calculate the first sum of squares of the data corresponding to each row in each column of the frequency domain filtered data, and select data columns in the frequency domain filtered data whose first sum of squares is greater than a first preset threshold for a preset number of columns;

[0023] Calculate the second sum of squares of the data in each column corresponding to each row of the data in the preset number of columns, and select data rows in the data in the preset number of columns whose second sum of squares is greater than a second preset threshold to obtain frequency domain filtered data after data compression;

[0024] Based on the row and column coordinates of the frequency domain filtered data after data compression in the preprocessed fall monitoring data, the corresponding data is selected from the time domain filtered data to obtain the time domain filtered data after data compression.

[0025] Optionally, the preset fall detection model includes dual input channels, and the preset fall detection model is any one of the deep learning networks such as a sum-product network, a convolutional neural network, or a recurrent network.

[0026] Optionally, when the preset fall detection model is a convolutional neural network, the preset fall detection model includes three convolutional layers, three downsampling layers, two fully connected layers and one output layer, wherein the output layer is a classifier with a preset number of categories.

[0027] Secondly, embodiments of the present invention provide a fall detection method, the method comprising:

[0028] Obtain fall monitoring data;

[0029] The fall monitoring data are subjected to frequency domain filtering and time domain filtering respectively to obtain the corresponding frequency domain filtered data and time domain filtered data;

[0030] The frequency domain filtered data and time domain filtered data are input into the target fall monitoring model trained by the fall monitoring model training method described in Example 1 to determine the fall action corresponding to the fall monitoring data.

[0031] Thirdly, embodiments of the present invention also provide a fall detection model training device, the device comprising:

[0032] The sample acquisition module is used to acquire fall monitoring data and preprocess the fall monitoring data, wherein each fall monitoring data is assigned a corresponding fall action label;

[0033] The sample processing module is used to perform frequency domain filtering and time domain filtering on the preprocessed fall monitoring data to obtain frequency domain filtered data and time domain filtered data.

[0034] The model training module is used to train the preset fall detection model by using the frequency domain filtered data, the time domain filtered data and the corresponding fall action labels as model training samples, so as to obtain the target fall detection model.

[0035] Optionally, the sample processing module is specifically used for:

[0036] Perform a Fast Fourier Transform on the fall monitoring data to obtain frequency domain fall monitoring data;

[0037] The frequency domain filtered data is obtained by filtering out data whose frequency is lower than a preset lower frequency limit from the frequency domain fall monitoring data.

[0038] Optionally, the sample processing module is further configured to:

[0039] Singular value decomposition was performed on the fall monitoring data to obtain data feature values;

[0040] The fall monitoring data is reconstructed by selecting feature values ​​within a preset range to obtain the time-domain filtered data.

[0041] Optionally, the sample processing module is further configured to:

[0042] Sort the feature values ​​of the data obtained from singular value decomposition;

[0043] Fall monitoring data corresponding to the data feature values ​​within a preset sorting range are selected from the sorted data feature values ​​as the data in the time-domain filtered data, and the fall monitoring data corresponding to the feature values ​​outside the preset sorting range are replaced with zero values ​​to obtain the time-domain filtered data.

[0044] Optionally, the fall detection model training device also includes a data compression module for:

[0045] Before training the preset fall detection model, the frequency domain filtered data and the time domain filtered data are compressed.

[0046] The compressed frequency domain filtered data and the compressed time domain filtered data, along with the corresponding fall action labels, are used as training samples for the model.

[0047] Optionally, the data compression module is specifically used for:

[0048] Calculate the first sum of squares of the data corresponding to each row in each column of the frequency domain filtered data, and select data columns in the frequency domain filtered data whose first sum of squares is greater than a first preset threshold for a preset number of columns;

[0049] Calculate the second sum of squares of the data in each column corresponding to each row of the data in the preset number of columns, and select data rows in the data in the preset number of columns whose second sum of squares is greater than a second preset threshold to obtain frequency domain filtered data after data compression;

[0050] Based on the row and column coordinates of the frequency domain filtered data after data compression in the preprocessed fall monitoring data, the corresponding data is selected from the time domain filtered data to obtain the time domain filtered data after data compression.

[0051] Optionally, the preset fall detection model includes dual input channels, and the preset fall detection model is any one of the deep learning networks such as a sum-product network, a convolutional neural network, or a recurrent network.

[0052] Optionally, when the preset fall detection model is a convolutional neural network, the preset fall detection model includes three convolutional layers, three downsampling layers, two fully connected layers and one output layer, wherein the output layer is a classifier with a preset number of categories.

[0053] Fourthly, embodiments of the present invention also provide a fall detection device, the device comprising:

[0054] The monitoring data acquisition module is used to acquire fall monitoring data;

[0055] The monitoring data processing module is used to perform frequency domain filtering and time domain filtering on the fall monitoring data to obtain corresponding frequency domain filtered data and time domain filtered data.

[0056] The monitoring result determination module is used to input the frequency domain filtered data and time domain filtered data into the target fall monitoring model trained by the fall monitoring model training method described in Example 1, and determine the fall action corresponding to the fall monitoring data.

[0057] Fifthly, embodiments of the present invention also provide a computer device, the computer device comprising:

[0058] One or more processors;

[0059] Memory, used to store one or more programs;

[0060] When the one or more programs are executed by the one or more processors, the one or more processors implement the fall detection model training method or fall detection method as provided in any embodiment of the present invention.

[0061] In a sixth aspect, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the fall detection model training method or fall detection method as provided in any embodiment of the present invention.

[0062] The embodiments of the above invention have the following advantages or beneficial effects:

[0063] In this embodiment of the invention, after acquiring fall monitoring data, frequency domain filtering and time domain filtering are performed on each piece of fall monitoring data to obtain frequency domain filtered data and time domain filtered data. Simultaneously, the frequency domain filtered data, time domain filtered data, and corresponding fall action labels are used as model training samples to train a preset fall monitoring model. This allows the detection model to learn from the data after both time and frequency domain filtering, resulting in a target fall monitoring model. The simultaneous parallel filtering in the time and frequency domains solves the problems of poor performance when using time domain filtering alone and low efficiency due to data conversion in the frequency domain filtering process in existing technologies. It achieves the fusion of time domain and frequency domain data, ensuring sufficient filtering of the monitoring data, improving monitoring data processing efficiency, and increasing the monitoring accuracy of the fall monitoring model. Attached Figure Description

[0064] Figure 1 This is a flowchart of a fall detection model training method provided in Embodiment 1 of the present invention;

[0065] Figure 2 This is a sample data diagram provided in Embodiment 1 of the present invention;

[0066] Figure 3 This is a schematic diagram of the frequency domain filtering effect of sample data provided in Embodiment 1 of the present invention;

[0067] Figure 4 This is a schematic diagram of the time-domain filtering effect of sample data provided in Embodiment 1 of the present invention;

[0068] Figure 5 This is a flowchart of a fall detection model training method provided in Embodiment 2 of the present invention;

[0069] Figure 6 This is a schematic diagram of a sample data compression result provided in Embodiment 2 of the present invention;

[0070] Figure 7 This is a schematic diagram of the structure of a preset fall monitoring model provided in Embodiment 2 of the present invention;

[0071] Figure 8 This is a flowchart of a fall detection method provided in Embodiment 3 of the present invention;

[0072] Figure 9 This is a schematic diagram of the structure of a fall detection model training device provided in Embodiment 4 of the present invention;

[0073] Figure 10 This is a schematic diagram of the structure of a fall monitoring device provided in Embodiment 5 of the present invention;

[0074] Figure 11 This is a schematic diagram of the structure of a computer device provided in Embodiment Six of the present invention. Detailed Implementation

[0075] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0076] Example 1

[0077] Figure 1 This is a flowchart illustrating a fall detection model training method according to Embodiment 1 of the present invention. This embodiment is applicable to situations where a fall detection model is trained based on fall detection data. The method can be executed by a fall detection model training device, which can be implemented in software and / or hardware and integrated into an electronic device with application development capabilities.

[0078] like Figure 1 As shown, the training method for the fall detection model includes the following steps:

[0079] S110. Obtain fall monitoring data, wherein each of the fall monitoring data is assigned a corresponding fall action label.

[0080] In fall detection, the most commonly used non-contact sensors include radar sensors, ultrasonic sensors, and infrared sensors. These sensors offer high cost-effectiveness and good privacy protection, but require sophisticated data processing. In this embodiment, the fall detection data processing procedure can be applied to both radar sensor monitoring data and ultrasonic sensor detection data.

[0081] For example, in this embodiment, radar data is used as an example for explanation. First, fall detection data is collected using radar equipment. The data acquisition frequency can be set to 100Hz, and the number of parallel sampling channels is set to 512. Different channels collect radar signals reflected and scattered from different angles. Based on the characteristics of a fall—a sudden, brief descent—the sampling time for one radar data point can be set to 4 seconds. When abnormal radar data is generated, the data from the moment the abnormal data occurs for 4 seconds is acquired to form a fall detection radar data point. Each data point can be plotted as a 400×512 array to form a radar data spatiotemporal echo map, as shown below. Figure 2As shown, radar monitoring data can be obtained by having test personnel fall in different postures or states. The collected data can then be tagged with corresponding fall action labels, such as falling while standing, falling while bending over, or falling while squatting to tie shoelaces. This allows for detailed classification of fall actions, avoiding misjudgments. Through this data collection process, a certain number of radar data samples can be collected according to sample size requirements. Each radar spatiotemporal echo image corresponds to one fall action; the image height can be considered as 400 pixels, and the width as 512 pixels.

[0082] It should be noted that parameters such as sampling frequency, number of sampling channels, and sampling duration for each data point can be set according to actual sampling needs. It is understood that the method in this embodiment is not limited to the collection of fall data, but can also be applied to data collection in other scenarios. The data collected by the radar device can be normalized data, which can be achieved by setting the parameters of the radar device.

[0083] S120. For each of the aforementioned fall monitoring data, frequency domain filtering and time domain filtering are performed respectively to obtain frequency domain filtered data and time domain filtered data.

[0084] In this embodiment, each piece of fall monitoring data is subjected to frequency domain filtering and time domain filtering respectively. Data processing is performed simultaneously in the two data domains, and then the frequency domain filtered data and the time domain filtered data are fused to form a new signal group.

[0085] When performing frequency domain filtering on fall monitoring data, firstly, a Fast Fourier Transform (FFT) is performed on the fall monitoring data to obtain frequency domain fall monitoring data. Then, data with frequencies lower than a preset lower frequency limit are filtered out to obtain frequency domain filtered data. For example, the preset lower frequency limit can be set to 1Hz or other frequency values. Low-frequency and low-Doppler-shift signals within 1Hz usually originate from echo signals from a stationary environment. The presence of these echoes will seriously interfere with the monitoring results and therefore must be removed. The radar echo map of the frequency domain filtered data after removal can be referenced. Figure 3 The image shown. In particular, in this embodiment, it is not necessary to transform the frequency domain filtered data back to time domain data.

[0086] When performing time-domain filtering on fall monitoring data, the first step is not to use time-domain filtering methods such as single-order cancellers, but rather to perform Singular Value Decomposition (SVD) on the fall monitoring data, and then reconstruct the data based on the decomposition results. Assuming the fall monitoring data is matrix A, then the SVD of matrix A is A = U∑VT In this embodiment, A is a 400*512 matrix, U is a 400*400 matrix, Σ is a 400*512 matrix, and V is a 512*512 matrix. Each element on the main diagonal of Σ is called a singular value. It can be understood that 512 singular values, or data feature values, can be obtained through singular value decomposition. The magnitude of the data feature value indicates the amount of information contained in the corresponding data; the larger the value, the greater the information content and the higher the importance of the data. A preset range can be set based on the data characteristics of the monitoring data, and the monitoring data within this preset range can be selected as the result after time-domain data filtering. For example, the data feature values ​​obtained from singular value decomposition can be sorted; then, the fall monitoring data corresponding to the data feature values ​​within the preset sorting range after sorting can be selected as the data in the time-domain filtered data, and the fall monitoring data corresponding to feature values ​​outside the preset sorting range can be replaced with zero, thus obtaining the time-domain filtered data. In one specific implementation, the filtered fall monitoring data corresponding to the 5th to 80th data feature values ​​after sorting can be selected, and the fall monitoring data corresponding to the 1st to 4th data feature values ​​and the 81st to 512th data feature values ​​can be replaced with the value 0 as the reconstructed time-domain graph (i.e., the time-domain filtered graph). Figure 4 As shown. The reason why the monitoring data after singular value decomposition and data filtering based on singular values ​​is used as time-domain filtered data is that this data processing is equivalent to a bandpass filter, filtering out low-frequency environmental echoes and high-frequency noise. The fall monitoring data corresponding to the 1st to 4th data feature values ​​usually correspond to data with frequencies less than 1Hz in the frequency domain. In addition, the singular value decomposition and data reconstruction process can also compress the image and reduce memory usage.

[0087] S130. The frequency domain filtered data, the time domain filtered data, and the corresponding fall action labels are used as model training samples to train the preset fall detection model and obtain the target fall detection model.

[0088] In this step, the frequency-domain filtered data, time-domain filtered data, and corresponding fall action label for each monitoring data point are used as training samples for the model. Simultaneously, the frequency-domain filtered data and time-domain filtered data are input into a preset fall monitoring model, which outputs a numerical value for the fall action type corresponding to the input data. If the output result is inconsistent with the label and the loss function has not converged, the preset fall monitoring model continues to learn from the input data until the loss function converges, satisfying the preset convergence condition, and the output result matches the label.

[0089] The preset fall detection model includes dual input channels and can be any deep learning network among sum-product networks, convolutional neural networks, or recurrent networks. In this embodiment, after acquiring fall detection data, frequency domain filtering and time domain filtering are performed on each piece of fall detection data to obtain frequency domain filtered data and time domain filtered data. Simultaneously, the frequency domain filtered data and time domain filtered data, along with the corresponding fall action labels, are used as model training samples to train the preset fall detection model. This allows the detection model to learn from the data after both time and frequency domain filtering, resulting in the target fall detection model. The simultaneous parallel filtering in the time and frequency domains solves the problems of poor performance when using time domain filtering alone and low efficiency due to data conversion in the frequency domain filtering process in existing technologies. It achieves the fusion of time domain and frequency domain data, ensuring sufficient filtering of the monitoring data, improving monitoring data processing efficiency, and increasing the monitoring accuracy of the fall detection model.

[0090] Example 2

[0091] Figure 5 This is a flowchart of a fall detection model training method provided in Embodiment 2 of the present invention. This embodiment further optimizes the above embodiment by compressing the sample data to improve the efficiency of model training. It belongs to the same inventive concept as the fall detection model training method in the above embodiment. This method can be executed by a fall detection model training device, which can be implemented by software and / or hardware and integrated into a server device with application development capabilities.

[0092] like Figure 5 As shown, the training method for the fall detection model includes the following steps:

[0093] S210. Obtain fall monitoring data, wherein each of the fall monitoring data is assigned a corresponding fall action label.

[0094] The acquisition of fall monitoring data in this step can refer to the fall monitoring data acquisition process described in step S110. Fall monitoring data can be ultra-wideband radar monitoring data or ultrasonic monitoring data.

[0095] In particular, when the monitoring data is ultra-wideband radar monitoring data, the monitoring data processing method in this embodiment is also applicable to monitoring the posture of people and animals, monitoring whether a car passes by, and can also be used to monitor actions in wall-penetrating scenarios, which can be used in the field of counter-terrorism.

[0096] S220. For each of the aforementioned fall monitoring data, frequency domain filtering and time domain filtering are performed respectively to obtain frequency domain filtered data and time domain filtered data.

[0097] Specifically, the filtering process of fall monitoring data in the time and frequency domains can be found in step S120.

[0098] S230. Compress the frequency domain filtered data and the time domain filtered data.

[0099] The frequency-domain filtered data and the time-domain filtered data reconstructed by singular value decomposition are still arrays with the same dimensions as the original fall monitoring data. Taking 512 sampling channels, a sampling frequency of 100Hz, and four seconds of fall monitoring data as an example, the frequency-domain filtered data and the time-domain filtered data are both 400*512 arrays. To reduce the amount of data computation and accelerate model convergence during subsequent model training, the frequency-domain filtered data and the time-domain filtered data can be compressed and simplified. Data reduction can be performed according to preset rules, while retaining data containing important information. For example, the 400*512 data can be reduced to an array of 200*200 rows or columns, or other arrays.

[0100] In a specific example, data compression can be achieved by segmenting data based on its energy. First, the sum of the squares of the data in each row of each column of the frequency-domain filtered data is calculated. Then, a predetermined number of columns in the frequency-domain filtered data are selected whose sum of the squares is greater than a predetermined threshold. That is, for each sampling channel, the sum of the squares of the sampled data in each channel is calculated, resulting in 512 sums of squares. Next, these 512 values ​​are iterated through, and a predetermined number of columns are selected. Data is then reduced within these columns; the predetermined number of columns can be 200 or other values ​​less than 512. The predetermined threshold can be a value set based on experience or a value calculated according to certain rules, such as the mean of the 512 sums of squares, or three-quarters of the mean of the 512 sums of squares. During the process of comparing the sum of squares of the traversed columns with the first preset threshold, the traversal process will stop when the number of columns with a sum of squares greater than the first preset threshold is greater than the preset number of columns. If the number of columns selected is less than the preset number of columns after the traversal process is completed, zeros will be added to the left and right sides of the selected data columns to ensure that the selected data meets the preset number of columns.

[0101] Furthermore, within the pre-selected data with a pre-defined number of columns, a pre-defined number of rows are also filtered. Specifically, the second sum of the squares of the data in each column corresponding to each row of the pre-defined number of columns is calculated, and rows in the pre-defined number of columns whose second sum of the squares is greater than a second pre-defined threshold are selected. This yields the frequency-domain filtered data after data compression. The process for filtering data with a pre-defined number of rows can be referenced from the process for filtering data with a pre-defined number of columns.

[0102] Finally, based on the row and column coordinates of the frequency-domain filtered data after data compression in the fall monitoring data, corresponding data are selected from the time-domain filtered data to obtain the time-domain filtered data after data compression. The data compression result is as follows: Figure 6 As shown. Among them, Figure 6 The left image shows the result of frequency domain filtering data compression, and the right image shows the result of time domain filtering data compression.

[0103] S240. The compressed frequency domain filtered data, the compressed time domain filtered data, and the corresponding fall action labels are used as model training samples to train the preset fall detection model and obtain the target fall detection model.

[0104] Using compressed data as samples for model training can reduce the amount of data computation and improve model training efficiency.

[0105] In this embodiment, the preset fall detection module can use a dual-channel input convolutional neural network to learn the data features of radar signals, construct a model framework, and realize fall detection. In a specific example, the preset fall detection model may include three convolutional layers, three downsampling layers, two fully connected layers, and one output layer, wherein the output layer is a classifier with a preset number of categories. The structure of the preset fall detection model can be referred to... Figure 7The diagram shows the model structure. The input layer directly receives compressed 200*200×2 temporal and frequency-domain filtered data. The C1 convolutional layer outputs 64 feature maps, meaning it needs to learn 64 convolutional kernels during model training. Padding is 2, and the stride is 1, extracting 64 different features from each region of the input image. The kernel size is 5×5. Weight sharing is used in the convolutional layers because it significantly reduces the number of parameters the convolutional neural network needs to learn. Each kernel uses the same weights to convolve the temporal and frequency-domain filtered images, while different kernels use different weights. The M2 layer is a downsampling layer, outputting the same number of feature maps as the C1 convolutional layer: 64. The M2 downsampling layer has a 3×3 sampling region with no overlap. The C3 convolutional layer outputs 256 feature maps, therefore, the number of convolutional kernels trained for C3 is 256, meaning it extracts 256 different features for each region of the input image. The kernel size is 3×3, the stride is 1, and the padding is 1. The M2 layer feature map resolution is 3×3, the convolution operation stride is 2, and the output feature map after the C3 convolution is 256. Each convolutional kernel in the C3 layer convolves on all feature maps in the M2 layer, i.e., a fully connected approach. The M4 downsampling layer operates on the C3 convolutional layer output feature maps in the same way as M2, and M6 operates on the C5 convolutional layer output feature maps in the same way. The F7 fully connected layer has 256 neurons, each of which is connected to all neurons in the M6 ​​layer. The parameters that the F7 fully connected layer needs to learn are determined by the number and resolution of the feature maps that the network ultimately outputs to the M6 ​​downsampling layer. F8, also a fully connected layer, requires learning parameters determined by F7. The output layer is actually a multi-class Softmax classifier, with the number of output nodes determined by the classification task. Figure 7 The value is set to 3 (i.e., assuming there are three types of fall action labels).

[0106] Test sample data of known ultra-wideband radar fall signals were obtained. The 200×200 standardized radar echo images, after signal filtering and data compression, were input into a trained convolutional neural network model. The accuracy was calculated by comparing the final classification result with the actual category. The prediction accuracy for different fall movement types exceeded 95%, which is higher than the classification results of existing threshold-based algorithms.

[0107] The technical solution of this embodiment, after acquiring fall monitoring data, performs frequency domain filtering and time domain filtering on each piece of fall monitoring data to obtain frequency domain filtered data and time domain filtered data. Then, the frequency domain filtered data and time domain filtered data are compressed. Simultaneously, the compressed frequency domain filtered data and time domain filtered data, along with the corresponding fall action labels, are used as model training samples and input into a dual-input channel convolutional neural network to train a preset fall monitoring model. This allows the detection model to learn from the data filtered in both the time and frequency domains simultaneously, resulting in a target fall monitoring model. Simultaneous parallel filtering in the time and frequency domains solves the problems of poor performance when using time domain filtering alone and low efficiency due to data conversion in the frequency domain filtering process in existing technologies. It achieves the fusion of time domain and frequency domain data, ensuring sufficient filtering of the monitoring data, improving the efficiency of monitoring data processing and model training, and ultimately increasing the monitoring accuracy of the fall monitoring model.

[0108] Example 3

[0109] Figure 8 This is a flowchart of a fall detection method provided in Embodiment 3 of the present invention. This embodiment is applicable to scenarios where a fall condition is determined based on monitoring signal data. The method can be executed by a fall detection device, which can be implemented in software and / or hardware and integrated into an electronic device with application development capabilities.

[0110] like Figure 8 As shown, the fall detection training method includes the following steps:

[0111] S310, Obtain fall monitoring data.

[0112] Within the preset monitoring range, monitoring data can be acquired whenever abnormal data occurs. The monitoring data can be ultra-wideband radar data, ultrasonic data, or monitoring data collected by other sensors.

[0113] S320. Perform frequency domain filtering and time domain filtering on the fall monitoring data to obtain the corresponding frequency domain filtered data and time domain filtered data.

[0114] The acquired fall monitoring data is subjected to data filtering in both the time and frequency domains. For frequency domain filtering, firstly, a Fast Fourier Transform is performed on the fall monitoring data to obtain the frequency domain fall monitoring data; then, data with frequencies lower than a preset lower frequency limit are filtered out to obtain the frequency domain filtered data. For time domain filtering, Singular Value Decomposition (SVD) is used to reconstruct the data based on the decomposition results. The specific process can be found in Example 1.

[0115] S330. Input the frequency domain filtered data and time domain filtered data into the target fall monitoring model trained by the fall monitoring model training method described in Example 1, and determine the fall action corresponding to the fall monitoring data.

[0116] By inputting the filtered data into the target fall detection model, the model can output the identification results of the monitoring data, indicating whether it is real fall data and the corresponding fall action type.

[0117] It should be noted that if the target fall detection model is determined based on the training method of the fall detection model in Example 2, the filtered data needs to be compressed before inputting the filtered data into the target fall detection model. For the specific data compression process, please refer to the data compression process described in Example 2.

[0118] The technical solution of this embodiment, after acquiring fall monitoring data, performs frequency domain filtering and time domain filtering on each piece of fall monitoring data to obtain frequency domain filtered data and time domain filtered data. Simultaneously, the frequency domain filtered data and time domain filtered data are used as input data and input into the target fall monitoring model to obtain the analysis results of the fall monitoring data and determine the fall action classification. Simultaneous parallel filtering in the time and frequency domains solves the problems of poor performance when using time domain filtering alone and low efficiency due to data conversion in the frequency domain filtering process in existing technologies. It achieves the fusion of time domain data and frequency domain data, ensuring sufficient filtering of the monitoring data, improving the monitoring data processing efficiency, and increasing the monitoring accuracy of the fall monitoring model.

[0119] The following are embodiments of the fall detection model training device and fall detection device provided in this invention. This device belongs to the same inventive concept as the fall detection model training method or fall detection method in the above embodiments, and can implement the fall detection model training method or fall detection method in the above embodiments. For details not described in detail in the embodiments of the fall detection model training device or fall detection device, please refer to the embodiments of the fall detection model training method or fall detection method described above.

[0120] Example 4

[0121] Figure 9 This is a schematic diagram of a fall detection model training device provided in Embodiment 4 of the present invention. This embodiment can be applied to situations where a fall detection model is trained based on fall monitoring data.

[0122] like Figure 9 As shown, the fall detection model training device includes a sample acquisition module 410, a sample processing module 420, and a model training module 430.

[0123] The sample acquisition module 410 is used to acquire fall monitoring data and preprocess the fall monitoring data, wherein each fall monitoring data is assigned a corresponding fall action label; the sample processing module 420 is used to perform frequency domain filtering and time domain filtering on the preprocessed fall monitoring data to obtain frequency domain filtered data and time domain filtered data; the model training module 430 is used to use the frequency domain filtered data, the time domain filtered data and the corresponding fall action labels as model training samples to train a preset fall monitoring model to obtain a target fall monitoring model.

[0124] The technical solution of this embodiment, after acquiring fall monitoring data, performs frequency domain filtering and time domain filtering on each piece of fall monitoring data to obtain frequency domain filtered data and time domain filtered data. Simultaneously, the frequency domain filtered data, time domain filtered data, and corresponding fall action labels are used as model training samples to train a preset fall monitoring model. This allows the detection model to learn from the data after both time and frequency domain filtering, resulting in a target fall monitoring model. Simultaneous parallel filtering in the time and frequency domains solves the problems of poor performance when using time domain filtering alone and low efficiency due to data conversion in the frequency domain filtering process in existing technologies. It achieves the fusion of time domain and frequency domain data, ensuring sufficient filtering of the monitoring data, improving monitoring data processing efficiency, and increasing the monitoring accuracy of the fall monitoring model.

[0125] Optionally, the sample processing module 420 is specifically used for:

[0126] Perform a Fast Fourier Transform on the fall monitoring data to obtain frequency domain fall monitoring data;

[0127] The frequency domain filtered data is obtained by filtering out data whose frequency is lower than a preset lower frequency limit from the frequency domain fall monitoring data.

[0128] Optionally, the sample processing module 420 is further configured to:

[0129] Singular value decomposition was performed on the fall monitoring data to obtain data feature values;

[0130] The fall monitoring data is reconstructed by selecting feature values ​​within a preset range to obtain the time-domain filtered data.

[0131] Optionally, the sample processing module 420 is further configured to:

[0132] Sort the feature values ​​of the data obtained from singular value decomposition;

[0133] Fall monitoring data corresponding to the data feature values ​​within a preset sorting range are selected from the sorted data feature values ​​as the data in the time-domain filtered data, and the fall monitoring data corresponding to the feature values ​​outside the preset sorting range are replaced with zero values ​​to obtain the time-domain filtered data.

[0134] Optionally, the fall detection model training device also includes a data compression module for:

[0135] Before training the preset fall detection model, the frequency domain filtered data and the time domain filtered data are compressed.

[0136] The compressed frequency domain filtered data and the compressed time domain filtered data, along with the corresponding fall action labels, are used as training samples for the model.

[0137] Optionally, the data compression module is specifically used for:

[0138] Calculate the first sum of squares of the data corresponding to each row in each column of the frequency domain filtered data, and select data columns in the frequency domain filtered data whose first sum of squares is greater than a first preset threshold for a preset number of columns;

[0139] Calculate the second sum of squares of the data in each column corresponding to each row of the data in the preset number of columns, and select data rows in the data in the preset number of columns whose second sum of squares is greater than a second preset threshold to obtain frequency domain filtered data after data compression;

[0140] Based on the row and column coordinates of the frequency domain filtered data after data compression in the preprocessed fall monitoring data, the corresponding data is selected from the time domain filtered data to obtain the time domain filtered data after data compression.

[0141] Optionally, the preset fall detection model includes dual input channels, and the preset fall detection model is any one of the deep learning networks such as a sum-product network, a convolutional neural network, or a recurrent network.

[0142] Optionally, when the preset fall detection model is a convolutional neural network, the preset fall detection model includes three convolutional layers, three downsampling layers, two fully connected layers and one output layer, wherein the output layer is a classifier with a preset number of categories.

[0143] The fall monitoring model training device provided in this embodiment of the invention can execute the fall monitoring model training method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0144] Example 5

[0145] Figure 10This is a schematic diagram of a fall monitoring device provided in Embodiment 5 of the present invention. This embodiment can be applied to scenarios where the fall condition is determined based on monitoring signal data.

[0146] like Figure 10 As shown, the fall monitoring device includes a monitoring data acquisition module 510, a monitoring data processing module 520, and a monitoring result determination module 530.

[0147] The monitoring data acquisition module 510 is used to acquire fall monitoring data; the monitoring data processing module 520 is used to perform frequency domain filtering and time domain filtering on the fall monitoring data to obtain corresponding frequency domain filtered data and time domain filtered data; the monitoring result determination module 530 is used to input the frequency domain filtered data and time domain filtered data into the target fall monitoring model trained by the fall monitoring model training method described in Example 1 to determine the fall action corresponding to the fall monitoring data.

[0148] The fall monitoring device provided in this embodiment of the invention can execute the fall monitoring method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method.

[0149] The technical solution of this embodiment, after acquiring fall monitoring data, performs frequency domain filtering and time domain filtering on each piece of fall monitoring data to obtain frequency domain filtered data and time domain filtered data. Simultaneously, the frequency domain filtered data and time domain filtered data are used as input data and input into the target fall monitoring model to obtain the analysis results of the fall monitoring data and determine the fall action classification. This solves the problems of poor performance when using time domain filtering alone and low efficiency due to data conversion in the frequency domain filtering process in the prior art. It achieves the fusion of time domain data and frequency domain data, enabling the monitoring data to be fully filtered, improving the monitoring data processing efficiency, and increasing the monitoring accuracy of the fall monitoring model.

[0150] Example 6

[0151] Figure 11 This is a schematic diagram of the structure of a computer device provided in Embodiment Six of the present invention. Figure 11 A block diagram of an exemplary computer device 12 suitable for implementing embodiments of the present invention is shown. Figure 11 The computer device 12 shown is merely an example and should not be construed as limiting the functionality or scope of the embodiments of the present invention. The computer device 12 can be any terminal device with computing capabilities, such as intelligent controllers and servers, mobile phones, and other terminal devices.

[0152] like Figure 11As shown, the computer device 12 is represented in the form of a general-purpose computing device. The components of the computer device 12 may include, but are not limited to: one or more processors or processing units 16, system memory 28, and a bus 18 connecting different system components (including system memory 28 and processing unit 16).

[0153] Bus 18 represents one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0154] Computer device 12 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by computer device 12, including volatile and non-volatile media, removable and non-removable media.

[0155] System memory 28 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32. Computer device 12 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 34 may be used to read and write non-removable, non-volatile magnetic media (…). Figure 11 Not shown; usually referred to as a "hard drive"). Although Figure 11 Not shown, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 18 via one or more data media interfaces. System memory 28 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0156] A program / utility 40 having a set (at least one) of program modules 42 may be stored, for example, in system memory 28. Such program modules 42 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 42 typically perform the functions and / or methods described in the embodiments of the present invention.

[0157] Computer device 12 can also communicate with one or more external devices 14 (e.g., keyboard, pointing device, display 24, etc.), and with one or more devices that enable a user to interact with the computer device 12, and / or with any device that enables the computer device 12 to communicate with one or more other computing devices (e.g., network card, modem, etc.). This communication can be performed via input / output (I / O) interface 22. Furthermore, computer device 12 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 20. As shown, network adapter 20 communicates with other modules of computer device 12 via bus 18. It should be understood that, although... Figure 11 As not shown, it can be used in conjunction with computer device 12 with other hardware and / or software modules, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0158] The processing unit 16 executes various functional applications and data processing by running programs stored in the system memory 28, such as implementing a fall detection model training method or a fall detection method provided in this embodiment.

[0159] The training methods for fall detection models include:

[0160] Acquire fall monitoring data, wherein each of the fall monitoring data is assigned a corresponding fall action label;

[0161] For each of the aforementioned fall monitoring data, frequency domain filtering and time domain filtering are performed respectively to obtain frequency domain filtered data and time domain filtered data;

[0162] The frequency domain filtered data, the time domain filtered data, and the corresponding fall action labels are used as model training samples to train the preset fall detection model and obtain the target fall detection model.

[0163] Fall detection methods include:

[0164] Obtain fall monitoring data;

[0165] The fall monitoring data are subjected to frequency domain filtering and time domain filtering respectively to obtain the corresponding frequency domain filtered data and time domain filtered data;

[0166] The frequency domain filtered data and time domain filtered data are input into the target fall monitoring model trained by any of the fall monitoring model training methods described in claims 1-8 to determine the fall action corresponding to the fall monitoring data.

[0167] Example 7

[0168] This embodiment seven provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the fall detection model training method or fall detection method as provided in any embodiment of the present invention.

[0169] The training methods for fall detection models include:

[0170] Acquire fall monitoring data, wherein each of the fall monitoring data is assigned a corresponding fall action label;

[0171] For each of the aforementioned fall monitoring data, frequency domain filtering and time domain filtering are performed respectively to obtain frequency domain filtered data and time domain filtered data;

[0172] The frequency domain filtered data, the time domain filtered data, and the corresponding fall action labels are used as model training samples to train the preset fall detection model and obtain the target fall detection model.

[0173] Fall detection methods include:

[0174] Obtain fall monitoring data;

[0175] The fall monitoring data are subjected to frequency domain filtering and time domain filtering respectively to obtain the corresponding frequency domain filtered data and time domain filtered data;

[0176] The frequency domain filtered data and time domain filtered data are input into the target fall monitoring model trained by any of the fall monitoring model training methods described in claims 1-8 to determine the fall action corresponding to the fall monitoring data.

[0177] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. For example, a computer-readable storage medium can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0178] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0179] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0180] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0181] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computing device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0182] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A method for training a fall detection model, characterized in that, include: Acquire fall monitoring data, wherein each of the fall monitoring data is assigned a corresponding fall action label; For each piece of fall monitoring data, frequency domain filtering and time domain filtering are performed respectively to obtain frequency domain filtered data and time domain filtered data. Specifically, frequency domain filtering of the fall monitoring data includes performing a Fast Fourier Transform (FFT) on the fall monitoring data to obtain frequency domain fall monitoring data; filtering out data in the frequency domain fall monitoring data whose frequency is lower than a preset lower frequency limit to obtain the frequency domain filtered data. Time domain filtering of the fall monitoring data includes performing singular value decomposition (SVD) on the fall monitoring data to obtain data feature values; selecting feature values ​​within a preset range to reconstruct the fall monitoring data to obtain the time domain filtered data. The frequency domain filtered data, the time domain filtered data, and the corresponding fall action labels are used as model training samples to train the preset fall detection model and obtain the target fall detection model.

2. The method according to claim 1, characterized in that, Selecting feature values ​​within a preset range corresponding to fall monitoring data for data reconstruction yields the time-domain filtered data, including: Sort the feature values ​​of the data obtained from singular value decomposition; Fall monitoring data corresponding to the data feature values ​​within a preset sorting range are selected from the sorted data feature values ​​as the data in the time-domain filtered data, and the fall monitoring data corresponding to the feature values ​​outside the preset sorting range are replaced with zero values ​​to obtain the time-domain filtered data.

3. The method according to claim 1, characterized in that, Before training the preset fall detection model, the method further includes: The frequency domain filtered data and the time domain filtered data are compressed. The compressed frequency domain filtered data and the compressed time domain filtered data, along with the corresponding fall action labels, are used as training samples for the model.

4. The method according to claim 3, characterized in that, The data compression of the frequency domain filtered data and the time domain filtered data includes: Calculate the first sum of squares of the data corresponding to each row in each column of the frequency domain filtered data, and select data columns in the frequency domain filtered data whose first sum of squares is greater than a first preset threshold for a preset number of columns; Calculate the second sum of squares of the data in each column corresponding to each row of the data in the preset number of columns, and select data rows in the data in the preset number of columns whose second sum of squares is greater than a second preset threshold to obtain frequency domain filtered data after data compression; Based on the row and column coordinates of the frequency domain filtered data after data compression in the fall monitoring data, the corresponding data is selected from the time domain filtered data to obtain the time domain filtered data after data compression.

5. The method according to claim 1, characterized in that, The preset fall detection model includes dual input channels, and the preset fall detection model is any one of the deep learning networks such as sum-product network, convolutional neural network or recurrent network.

6. The method according to claim 5, characterized in that, When the preset fall detection model is a convolutional neural network, the preset fall detection model includes three convolutional layers, three downsampling layers, two fully connected layers and one output layer, wherein the output layer is a classifier with a preset number of categories.

7. A fall detection method, characterized in that, include: Obtain fall monitoring data; The fall monitoring data are subjected to frequency domain filtering and time domain filtering respectively to obtain the corresponding frequency domain filtered data and time domain filtered data; The frequency domain filtered data and time domain filtered data are input into the target fall monitoring model trained by any of the fall monitoring model training methods described in claims 1-6 to determine the fall action corresponding to the fall monitoring data.

8. A fall detection model training device, characterized in that, include: The sample acquisition module is used to acquire fall monitoring data and preprocess the fall monitoring data, wherein each fall monitoring data is assigned a corresponding fall action label; The sample processing module is used to perform frequency domain filtering and time domain filtering on the preprocessed fall monitoring data to obtain frequency domain filtered data and time domain filtered data. Specifically, the sample processing module is used to perform Fast Fourier Transform on the fall monitoring data to obtain frequency domain fall monitoring data; filter out data in the frequency domain fall monitoring data whose frequency is lower than a preset lower frequency limit to obtain the frequency domain filtered data; perform singular value decomposition on the fall monitoring data to obtain data feature values; select feature values ​​within a preset range to reconstruct the fall monitoring data to obtain the time domain filtered data. The model training module is used to train the preset fall detection model by using the frequency domain filtered data, the time domain filtered data and the corresponding fall action labels as model training samples, so as to obtain the target fall detection model.

9. A fall detection device, characterized in that, include: The monitoring data acquisition module is used to acquire fall monitoring data; The monitoring data processing module is used to perform frequency domain filtering and time domain filtering on the fall monitoring data to obtain corresponding frequency domain filtered data and time domain filtered data. The monitoring result determination module is used to input the frequency domain filtered data and time domain filtered data into the target fall monitoring model trained by any of the fall monitoring model training methods described in claims 1-6, and determine the fall action corresponding to the fall monitoring data.

10. A computer device, characterized in that, The computer device includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the fall detection model training method or fall detection method as described in any one of claims 1-7.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the fall detection model training method or fall detection method as described in any one of claims 1-7.

Images